Bagaimana Penyelidik Mengukur, Mendeteksi dan Benchmark Manipulasi AI

Penulis : Enes Altuncu, ea483@kent.ac.uk (University of Kent, UK) Virginia N. L. Franqueira, V.Franqueira@kent.ac.uk (University of Kent, UK) Shujun Li, S.J.Li@kent.ac.uk (University of Kent, UK) Penulis : Enes Altuncu, ea483@kent.ac.uk (Universiti Kent, UK) Virginia N. L. Franqueira, V.Franqueira@kent.ac.uk (Universiti Kent, UK) Shujun Li, S.J.Li@kent.ac.uk (Universiti Kent, UK) Abstract abstrak Kemajuan baru-baru ini dalam AI, terutamanya pembelajaran mendalam, telah menyumbang kepada peningkatan yang signifikan dalam penciptaan media sintetik baru (video, imej, dan audio) yang kelihatan realistis dan manipulasi media sedia ada, yang telah membawa kepada penciptaan istilah baru "deepfake". Berdasarkan kedua-dua literatur penyelidikan dan sumber dalam bahasa Inggeris dan Cina, kertas ini juga memberikan gambaran menyeluruh tentang deepfake, meliputi pelbagai aspek penting konsep yang muncul, termasuk 1) definisi yang berbeza, 2) metrik prestasi dan piawaian yang biasa digunakan, dan 3) set data yang berkaitan dengan deepfake, cabaran, pertandingan dan penilaian. : Deepfake, Survey, Definition, Datasets, Benchmarks, Challenges, Competitions, Standards, Performance Metrics. Kata Kunci 1 Pengenalan Kemajuan baru-baru ini dalam AI dan pembelajaran mesin telah meningkatkan keupayaan untuk menghasilkan media yang lebih realistik, contohnya, video, imej, dan audio. Terutamanya, kaedah pembelajaran mendalam moden membolehkan penciptaan "deepfakes", media yang dimanipulasi atau sintetik yang realiti tidak mudah dikenali oleh mata manusia. Walaupun deepfake merupakan fenomena yang agak baru (yang pertama kali muncul pada akhir 2017), pertumbuhannya telah luar biasa. Bilangan video deepfake di internet berbahasa Inggeris meningkat daripada 7,964 (Desember 2018) kepada 14,678 (Juli 2019) kepada 85,047 (Desember 2020), mewakili peningkatan 968% dari 2018 hingga 2020. 2 Dalam kerja ini, kami meninjau ekosistem penyelidikan yang berkaitan dengan deepfake yang sedia ada dalam pelbagai aspek, termasuk metrik dan piawaian prestasi, set data, cabaran, pertandingan, dan benchmarks.Selain itu, kami menyediakan kajian meta daripada 12 kertas kajian yang dipilih yang berkaitan dengan deepfake yang merangkumi beberapa aspek tambahan selain daripada yang dinyatakan secara sistematik, seperti perbandingan prestasi, cabaran utama, dan cadangan. Walaupun ia merupakan istilah yang sangat popular, terdapat kurangnya persetujuan mengenai definisi "deepfake" dan sempadan antara deepfakes dan non-deepfakes tidak jelas. Untuk kajian ini, kami mengambil pendekatan yang lebih inklusif untuk meliputi semua bentuk media dimanipulasi atau sintetik yang dianggap deepfakes dalam makna yang lebih luas. Kami juga meliputi topik yang berkaitan erat termasuk biometrik dan forensik multimedia, kerana deepfakes sering digunakan untuk melancarkan serangan persembahan terhadap sistem pengesahan berasaskan biometrik dan pengesanan deepfakes boleh dianggap sebahagian daripada forensik multimedia. Perbincangan yang lebih terperinci mengenai definisi yang berbeza "deepfake" diberikan seterusnya. 1.1 Pengenalan kepada istilah Deepfake Seperti namanya, istilah "deepfake" berasal daripada gabungan "deep" (mengacu kepada (DL)) dan "palsu". Ia biasanya digunakan untuk merujuk kepada manipulasi media yang sedia ada (gambar, video dan / atau audio) atau pengeluaran media baru (sintetik) menggunakan pendekatan berasaskan DL. Data deepfake yang paling biasa dibincangkan adalah imej muka palsu, penipuan ucapan palsu, dan video palsu yang menggabungkan kedua-dua imej palsu dan penipuan ucapan palsu. Walaupun mempunyai "palsu" dalam perkataan menunjukkan media yang dimanipulasi atau disintesis, terdapat banyak aplikasi yang baik daripada teknologi deepfake, contohnya, untuk hiburan dan seni kreatif. Dalam hal ini, istilah lain "sintesis dalam" telah dicadangkan sebagai alternatif yang lebih netral [ ].Tetapi istilah baru ini belum diterima secara meluas. Pembelajaran mendalam 60 Selain daripada kurangnya definisi universal, seperti yang telah disebutkan, sempadan antara deepfakes dan non-deepfakes sebenarnya bukan potongan yang jelas. Pertama, pengesanan deepfakes sering mengikuti pendekatan yang sangat serupa dengan pengesanan deepfakes tradisional yang dihasilkan tanpa menggunakan teknik DL. Kaedah pengesanan lanjutan juga telah mula memanfaatkan DL untuk meningkatkan prestasi mereka, tetapi mereka tidak perlu tahu bagaimana media sasaran dicipta (deep atau tidak). Pada tahap tertentu, seseorang boleh berpendapat bahawa pengesanan deepfakes tidak melibatkan membangunkan kaedah deepfakes tertentu (walaupun sesetengah penyelidik memilih untuk melakukannya), tetapi detektor yang lebih kukuh dan universal yang boleh menangani mana-mana (deep atau tidak) media palsu. Ini boleh dilihat untuk dua topik yang berkaitan erat: biometrik dan forensik multimedia. Untuk biometrik, terdapat kecenderungan untuk menggunakan teknik pembelajaran mendalam untuk menghasilkan isyarat biometrik palsu (contohnya, imej muka dan video) untuk serangan biometr Kedua, seseorang boleh berpendapat bahawa perkataan "deep" dalam "deepfake" tidak semestinya merujuk kepada penggunaan "pelajaran mendalam", tetapi apa-apa teknologi "mendalam" (iaitu canggih) yang mewujudkan media palsu yang sangat boleh dipercayai. ] dianggap deepfake sebagai manipulasi audio-visual menggunakan "spektrum kecerdasan teknikal ... dan teknik". dan , merujuk kepada “pengendalian tahap rendah media audio-visual yang dicipta dengan perisian yang mudah diakses [atau tanpa perisian] untuk mempercepat, memperlambat, mengembalikan atau mengkontekstualkan semula kandungan”. pemahaman yang lebih luas tentang “deepfake” juga telah diadopsi oleh pembuat undang-undang untuk undang-undang baru yang memerangi deepfakes jahat. 9 ShallowPerkosa Palsu 2018 Akta Larangan Palsu Berbahaya 1: Perkara 4 (b) (1) : ” Istilah "dosa mendalam" bermaksud rekod audio-visual yang dicipta atau diubahsuai sedemikian rupa sehingga rekod itu secara palsu akan kelihatan kepada penonton yang munasabah sebagai rekod sebenar perkataan atau tingkah laku sebenar seseorang. 2019 DEEP FAKES Akta Tanggungjawab2: § 1041 (n).(3) : “ Istilah "palsu mendalam" bermaksud mana-mana rekod video, filem filem, rekod bunyi, imej elektronik, atau fotografi, atau mana-mana representasi teknologi ucapan atau tingkah laku yang secara substansial berasal daripada mereka. (a) yang kelihatan mewakili perkataan atau tingkah laku mana-mana orang yang tidak benar-benar terlibat dalam perkataan atau tingkah laku itu; dan ” (B) yang pengeluarannya bergantung secara substansial kepada cara-cara teknikal, dan bukan keupayaan orang lain untuk secara fizikal atau lisan menggambarkan orang tersebut. Seperti yang kita boleh lihat daripada definisi undang-undang di atas "deepfake", penggunaan DL sebagai teknologi tidak disebutkan sama sekali. 1.2 Saiz dan sumbangan Berdasarkan perbincangan di atas mengenai definisi deepfake, kita boleh melihat bahawa tidak selalu mudah atau bermakna untuk membezakan deepfakes daripada non-deep fakes. Selain itu, untuk tumpuan kami pada penilaian dan perbandingan prestasi, sempadan antara deepfakes dan non-deep fakes lebih kabur. Ini kerana DL hanyalah bentuk khas (deeper) pembelajaran mesin (ML), dan sebagai hasilnya, kaedah DL dan non-deep ML berkongsi banyak konsep, metrik dan prosedur yang biasa. Walaupun fakta bahawa deepfake boleh dipahami dalam makna yang lebih luas, dalam kerja ini, kita mempunyai tumpuan yang cukup sempit untuk mengelakkan merangkumi terlalu banyak topik. Untuk metrik dan piawaian, kami memilih untuk memasukkan semua yang biasa digunakan untuk menilai kaedah ML umum dan yang secara khusus ditakrifkan untuk menilai kaedah penciptaan atau pengesanan deepfake. For datasets, challenges, competitions and benchmarks, we considered those related to fake media covered in the deepfake-related survey papers and those with an explicit mention of the term “deepfake” or a comparable term. Untuk meta-review, kami hanya mempertimbangkan makalah kajian yang penulis secara eksplisit merujuk kepada istilah "deepfakes" dalam meta data (judul, abstrak dan kata kunci). 2 Metodologi Maklumat penyelidikan yang dibincangkan dalam kajian ini (iaitu maklumat kajian yang berkaitan dengan penipuan mendalam) telah diidentifikasi melalui carian sistematik dalam pangkalan data saintifik, Scopus dan China Online Journals (COJ) Pertanyaan carian berikut digunakan untuk melakukan carian pada Scopus dan COJ, masing-masing: 3 (deepfake* OR deep-fake* OR “deep fake*”) DAN (review OR survey OR overview OR systemati* OR SoK) (dipfake OR 深度偽造) DAN (综述 OR 进展) Pencarian mengembalikan 41 makalah kajian dalam bahasa Inggeris dan 15 makalah kajian dalam bahasa Cina.Daripada makalah ini, lapan diterbitkan dalam bahasa Inggeris dan empat diterbitkan dalam bahasa Cina telah dipilih untuk pertimbangan. Cabaran yang berkaitan dengan Deepfake, pertandingan dan benchmarks telah diidentifikasi melalui pelbagai sumber: kertas kajian yang dipilih, kertas penyelidikan daripada koleksi peribadi rakan-rakan, carian Google Web, dan pemeriksaan manual laman web persidangan berkaitan AI utama yang diadakan pada tahun 2020 dan 2021 (di mana cabaran dan persidangan seperti itu dianjurkan secara rutin). persidangan yang disiasat termasuk yang disenaraikan dalam ACL (Association for Computational Linguistics) Anthology , ICCV, CVPR, AAAI, ICML, ICLR, KDD, SIGIR, WWW, dan banyak lagi. Selain itu, senarai komprehensif set data telah dikumpulkan berdasarkan kertas kajian yang dipilih dan cabaran, persaingan, dan benchmarks yang diidentifikasi. piawaian yang berkaitan telah diidentifikasi terutamanya melalui kertas penyelidikan yang disenaraikan dalam kajian ini, pengetahuan peribadi rakan-rakan penulis, dan carian Google Web. Untuk metrik prestasi, kami merangkumi yang biasa digunakan berdasarkan piawaian yang berkaitan, kertas kajian, dan cabaran, persaingan, dan benchmarks yang diidentifikasi. 4 3 Metrik dan piawaian prestasi yang berkaitan dengan Deepfake Dalam kajian ini, kami memberi tumpuan kepada penilaian prestasi dan perbandingan kaedah pengeluaran dan pengesanan deepfake. Metrik yang digunakan untuk penilaian prestasi sedemikian adalah di tengah-tengah perbincangan kami. Dalam seksyen ini, kami meninjau metrik prestasi yang biasa digunakan untuk menilai algoritma pengeluaran dan pengesanan deepfake. Sila ambil perhatian bahawa semua metrik yang dibahas dalam seksyen ini juga biasa digunakan untuk menilai prestasi sistem serupa yang tidak untuk menghasilkan atau mendeteksi deepfakes. Oleh itu, seksyen ini boleh dilihat sebagai tutorial yang sangat singkat mengenai metrik prestasi umum. Dengan "standar formal", kita merujuk kepada standard yang ditakrifkan mengikut prosedur formal, sering oleh satu atau lebih badan standardisasi yang ditubuhkan seperti Pertubuhan Antarabangsa untuk Standardisasi (ISO) Jawatankuasa Elektroteknik Antarabangsa (IEC) Perhatikan bahawa kami mempertimbangkan pelbagai dokumen yang ditakrifkan sebagai standard oleh badan-badan standardisasi, contohnya, Kesatuan Telekomunikasi Antarabangsa (ITU). Rencana dan laporan teknikal ISO (TRs) 5 6 7 4.1 Matriks kebingungan Deepfake Detection ialah masalah klasifikasi binari.Binary classifier mengambil input yang atau dan mengeluarkan nilai binari yang menandakan bahawa ia atau Contohnya, sistem pengesanan deepfake akan mengambil imej yang disyaki sebagai input yang mungkin atau dan output atau . Sebenarnya positif Sebenarnya negatif Memperkirakan positif Memprediksi negatif Sebenarnya palsu actually real Pencapaian palsu predicted real Alat penting yang digunakan dalam menilai klasifikator binari ialah yang merangkumi kejayaan dan kegagalan model klasifikasi. values and on the other axis are the two Nilai-nilai ini ialah klasifikasi (sebenarnya positif dan benar-benar negatif) apabila nilai sebenar dan nilai yang diramalkan sepadan. (False Positive dan False Negative) apabila nilai sebenar dan dijangka tidak sepadan. menunjukkan matriks kebingungan untuk pengklasifikasian deepfake binari (detektor). Dua sel dalam hijau, TP (jumlah ) dan TN ( bilangan ), menunjukkan hasil ramalan yang betul, dan dua sel dalam merah, FN (jumlah ) dan FP (jumlah ), menunjukkan dua jenis kesilapan yang berbeza apabila membuat hasil ramalan yang salah. confusion matrix semasa Memprediksi Sukses / betul / benar Kegagalan / salah / salah 1 true positives true negatives false negatives false positives Table 1: Confusion matrix for a binary classifier for detecting deepfake. fake (predicted) real (predicted) fake (actual) TP FN real (actual) FP TN Palsu (Kehadiran yang sebenar) TP yang PBB sebenar (sesuatu yang sebenar) fp Tn 3.2 Ketepatan dan pengingat Berdasarkan empat nilai asas yang diperkenalkan dalam Seksyen , i.e., TP, TN, FP and FN, we define two important performance metrics for a binary classifier – dan . 3.1 precision recall Ketepatan sebuah klasifikator binari ditakrifkan sebagai sebahagian daripada sampel di antara semua Perbezaan di bawah ini ialah perbezaan di bawah ini ialah perbezaan di bawah ini ialah perbezaan di bawah ini ialah perbezaan di bawah ini ialah perbezaan di bawah ini ialah perbezaan di bawah ini ialah perbezaan di bawah ini ialah perbezaan di bawah ini ialah perbezaan di bawah ini ialah perbezaan di bawah ini ialah perbezaan ( ) daripada Sebenarnya positif Memperkirakan positif 1 Apabila rasio "asli" antara sampel positif dan negatif secara signifikan berbeza daripada set ujian, ia sering berguna untuk menyesuaikan berat positif palsu, yang membawa kepada (Bahasa Inggeris) yang ditakrifkan dalam Eq. ) di mana 0 ialah berat yang ditentukan oleh perbandingan antara sampel negatif dan positif. weighted precision 2 α » Mengingatkan klasifikator binari ialah sebahagian daripada sampel di antara contoh, seperti yang dinyatakan dalam Eq. ( Dalam matriks kebingungan, ia adalah sebahagian daripada sampel sebenar dalam baris pertama. Memperkirakan positif Sebenarnya positif 3 Mari kita pertimbangkan contoh klasifikator binari yang meramalkan sama ada imej dari pangkalan data yang mengandungi kedua-dua imej deepfake dan sebenar (asli) adalah palsu atau tidak. ketepatan klasifikator ialah bahagian imej yang diklasifikasikan dengan betul di antara semua imej yang diklasifikasikan sebagai deepfake. 3.3 True and False Positive Rates Berfokus pada sampel positif yang diramalkan, kita juga boleh mendefinisikan dua metrik: (TPR), also called (CDR), sebagai sebahagian daripada sampel positif yang diramalkan antara sampel positif sebenar dan (FPR), juga dikenali sebagai (FAR), as the fraction of the predicted positive samples among the actually negative samples, as shown in Eqs. ( ) dan true positive rate correct detection rate false positive rate false alarm rate 4 ( Dalam matriks kebingungan, TPR adalah sebahagian daripada sampel positif yang diramalkan dalam baris pertama dan FPR adalah sebahagian daripada sampel positif yang diramalkan dalam baris kedua. (Bahasa Melayu) )). 5 recall 3 3.4 Kadar negatif sebenar dan palsu Sama seperti kadar positif benar dan palsu, kita boleh menentukan dua kadar lain yang memberi tumpuan kepada hasil negatif yang diramalkan: (TNR) indicating the fraction of the predicted negative samples among the actually negative samples, and (FNR) menunjukkan sebahagian daripada sampel negatif dijangka di antara sampel positif sebenarnya, seperti yang ditunjukkan dalam Eqs. ( ) dan ( ). true negative rate false negative rate 6 7 3.5 Sensitiviti dan spesifikasi Dalam sesetengah aplikasi klasifikator binari, terutamanya dalam biologi dan perubatan, TPR dan TNR digunakan lebih kerap, dan mereka sering dipanggil (TPR) dan (TNR) Tumpuan kedua-dua istilah ini adalah pada kedua-dua jenis ketepatan hasil yang diramalkan.Ia kurang digunakan dalam penyelidikan yang berkaitan dengan deepfake, oleh itu, kita tidak akan merujuk kepada mereka dalam sisa kertas ini. sensitivity specificity 3.6 Kadar kesilapan yang sama Berfokus pada kadar kesilapan bermakna kita perlu mempertimbangkan FPR dan FNR. Kedua-dua kadar ini biasanya berkonflik dengan satu sama lain sehingga mengurangkan satu kadar biasanya membawa kepada peningkatan pada yang lain. Oleh itu, alih-alih cuba mengurangkan kedua-dua kadar kesilapan pada masa yang sama, yang biasanya tidak mungkin, tugas yang lebih realistis dalam aplikasi praktikal adalah untuk mencari keseimbangan yang betul supaya kedua-duanya berada di bawah ambang batas yang boleh diterima. Dalam sesetengah aplikasi, seperti biometrik, orang-orang terutamanya berminat dalam menubuhkan apa yang dipanggil (Dalam hal ini) atau EER/CER tidak semestinya merupakan metrik yang baik untuk sesetengah aplikasi, terutamanya apabila kedua-dua jenis kesilapan mempunyai tahap penting yang berbeza, contohnya, untuk mengesan deepfakes kritikal (contohnya, berita palsu yang boleh mempengaruhi cara orang mengundi mereka) kita sering boleh mentolerir lebih banyak positif palsu (alarm palsu) daripada negatif palsu (alarm terlewat). equal error rate crossover error rate 3.7 Accuracy and F-Score Selain EER/CER, terdapat juga metrik lain yang cuba mencerminkan kedua-dua jenis kesilapan, untuk memberikan indikasi yang lebih seimbang mengenai prestasi keseluruhan klasifikator binari. dan (juga dikenali sebagai Kedua-dua metrik boleh ditakrifkan berdasarkan empat nilai asas (TP, TN, FP, dan FN). accuracy F-score F-measure Ketepatan sebuah klasifikator binari ditakrifkan sebagai sebahagian daripada Perbezaan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilangan bilion ( ) daripada Diprediksi dengan betul 8 F-score klasifikator binari sebenarnya merupakan keluarga metrik. bentuk umumnya boleh digambarkan berdasarkan parameter as defined in Eq. ( ) daripada β 9 The most widely used edition of all F-scores is the so-called , which is effectively the F-score with = 1. lebih tepatnya, ia ditakrifkan seperti yang ditunjukkan dalam Eq. ( ). F1-score β 10 3.8 Receiver Operating Characteristic Curve dan kawasan di bawah kurva Curves (ROC) biasanya digunakan untuk mengukur prestasi klasifikator binari yang mengeluarkan skor (atau kebarangkalian) prediksi. Receiver operating characteristic Pertimbangkan perkara berikut. biarkan menjadi set semua sampel ujian dan biarkan skor output ( ) (for all ∈ ) lie in the interval [ ] on the real line. Let menjadi ambang prediksi untuk model, dan menganggap bahawa klasifikator berfungsi seperti berikut untuk semua ∈ : S f s s S A , B t A , B s S Ia mudah untuk dilihat, kerana = daripada , semua sampel akan diklasifikasikan sebagai positif, yang membawa kepada FN = TN = 0 jadi TPR = FPR = 1; manakala untuk = , semua sampel akan diklasifikasikan sebagai negatif, yang membawa kepada FP = TP = 0 jadi TPR = FPR = 0. dan , nilai TPR dan FPR biasanya antara 0 dan 1. from to continuously, we can normally get a continuous curve that describes how the TPR and FPR values change from (0,0) to (1,1) on the 2D plane. This curve is the ROC curve of the binary classifier. t a t b a b t a b For a random classifier, assuming that (Daripada ) distributes uniformly on [ ] for the test set, we can mathematically derive its ROC curve being the TPR = FPR line, whose area under the ROC curve (AUC) is 0.5. For a binary classifier that performs better than a random predictor, we can also mathematically prove that its AUC is always higher than 0.5, with 1 being the best possible value. Note that no binary classifier can have an AUC below 0.5, since one can simply flip the prediction result to get a better predictor with an AUC of 1 − AUC. The relationship between the ROC and the AUC is graphically illustrated in Figure . f s A , B 1 3.9 Log Loss Another widely used performance metric for binary classifiers that can return a probability score for the predicted label is Untuk klasifikasi binari dengan label sebenar ∈ {0*,* 1} and an estimated probability = Pr( Perbezaan antara pengambilan sampel dan pengambilan sampel ialah pengambilan sampel ( ). log loss y p y 12 Given a testing set with samples, the log loss score of a binary classifier can be calculated using Eq. ( ), where is 1 if the -th sample is true and 0 if false, and ˆ is the predicted probability of = 1. n 13 yi i y i yi 3.10 Extension to Multi-class Classifiers Semua metrik yang ditakrifkan berdasarkan empat nilai asas TP, TN, FP dan FN boleh dengan mudah diperluaskan kepada Contohnya, jika sistem mengklasifikasikan haiwan (kucing, anjing, kuda, singa, harimau, dan lain-lain), maka ramalan positif yang benar mengenai imej kucing akan secara bersamaan menjadi ramalan negatif yang benar bagi kelas-kelas yang tersisa (anjing, kuda, singa, harimau, dan lain-lain). multi-class classification Penilaian Kualiti Perceptual (PQA) Metrik By definition, the main goal of deepfakes is to make it hard or impossible for human consumers (listeners or viewers) to distinguish fake media from real media. Therefore, when evaluating the quality of deepfake media, the quality perceived by human consumers of the media is key. This calls for subjective assessment of the perceptual quality of the deepfake media as the “gold standard”. The most widely used subjective perceptual quality assessment (PQA) metric for audio-visual signals is (MOS), which has been widely used by the signal processing and multimedia communication communities, including digital TV and other multimedia-related consumer applications. As its name implies, MOS is calculated by averaging the subjective scores given by a number of human judges, normally following a numerical scale between 1 and 5 or between 0 and 100. MOS has been used in some deepfake-related challenges (see Section ) dan juga untuk menilai dan membandingkan kualiti (realitas / alam semula jadi) set data deepfake (lihat Seksyen ). mean opinion score 5.2 4.6 Sebagai metrik PQA subjektif umum, MOS telah disetorkan oleh ITU . There are also ITU standards defining more specific subjective Video Quality Assessment (VQA) metrics and the standard procedures one should follow to conduct VQA user studies, e.g., ITU-T Recommendation P.910 “Subjective video quality assessment methods for multimedia applications” . Note that the ITU standards focus more on traditional perceptual quality, i.e., how good a signal looks or sounds, even if it looks or sounds not real (e.g., too smooth). On the other hand, for deepfakes, the focus is rather different because what matters is the realness and naturalness of the created media, i.e., how real and natural it looks or sounds, even if it is of low quality. To some extent, we can also consider realness and naturalness as a special aspect of perceptual quality. 8 9 One major problem of subjective PQA metrics like MOS is the need to recruit human judges and to have a well-controlled physical testing environment and protocol, which are not easy for many applications. To help reduce the efforts and costs of conducting PQA-related user studies, various objective PQA metrics have been proposed, where the term “objective” refers to the fact that such metrics are human-free, i.e., automatically calculated following a computational algorithm or process. Depending on whether a reference exists, such objective PQA metrics can be largely split into three categories: full-reference (FR) metrics (when the original “perfect-quality” signal is available as the reference), reduced-reference (RR) metrics (when some features of the original “perfect-quality” signal are available as the reference), and no-reference (NR) metrics (when the original signal is unavailable or such an original signal does not exist). For deepfakes, normally NR or RR metrics are more meaningful because the “fake” part of the word means that part of the whole data does not exist in the real world, hence a full reference cannot be obtained. RR metrics are still relevant because deepfakes are often produced for a target’s specific attributes (e.g., face and voice), where the reduced reference will be such attributes. NR metrics will be useful to estimate the realness and naturalness of a deepfake, simulating how a human judge would rate it in a controlled subjective PQA user study. PQA is a very active research area and many PQA metrics have been proposed, some of which have been widely used in real-world products and services, e.g., (MSE), (PSNR) and (SSIM) untuk FR PQA gambar digital dan video yang didefinisikan sebagai dalam Eqs. (14), (15), dan (16), masing-masing, di mana X = {xi} n i adalah rujukan (sinyal asal), Y = {yi} n i ialah isyarat yang kualiti visual dinilai, n ialah bilangan piksel dalam X dan Y , L adalah nilai piksel maksimum yang mungkin X dan Y (contohnya, 255 untuk 8-bit gambar kelabu skala), c1 = (k1L) 2 dan c2 = (k2L) 2 ) adalah dua parameter stabilisasi (k1 = 0.01 dan k2 = 0.03 secara lalai). mean squared error peak signal-to-noise ratio structural similarity index measure 3.12 More about Standards Many of the basic performance metrics described in this section have been widely used by deepfake researchers as de facto standards, e.g., EER, log loss and MOS have been widely used in deepfake-related challenges (see Section ). Also, the combination of precision, recall and F1-score has been widely used to assess performance of binary classifiers. While there have been a number of ITU standards on PQA to date, there does not seem to be many standardisation efforts on the performance metrics for evaluation of binary classifiers. This was the case until at least 2017, when ISO and IEC jointly set up the ISO/IEC JTC 1/SC 42 , a standardisation subcommittee (SC) focusing on AI under ISO/IEC JTC 1 Jawatankuasa Teknikal Gabungan untuk Standardisasi "Teknologi Maklumat". 5 10 11 One recent effort that ISO/IEC JTC 1/SC 42 made is to produce the ISO/IEC TR 24029-1:2021 “Artificial Intelligence (AI) – Assessment of the robustness of neural networks – Part 1: Overview” , a technical report (TR) that systematically covers many commonly used performance assessment concepts, methods and metrics. Although the technical report has “neural networks” in its title, most performance assessment concepts, methods and metrics included are common ones for all supervised machine learning models. 12 In terms of performance metrics, two other ongoing work items of the ISO/IEC JTC 1/SC 42 that deserve attention are as follows: ISO/IEC DTS (Draft Technical Specification) 4213 “Information technology – Artificial Intelligence – Assessment of machine learning classification performance” 13 ISO/IEC AWI (Approved Work Item) TS (Technical Specifications) 5471 “Artificial intelligence – Quality evaluation guidelines for AI systems” 14 While the ISO/IEC JTC 1/SC 42 was created very recently, another standardisation subcommittee under ISO/IEC JTC1 has a much longer history of nearly 20 years: the ISO/IEC JTC 1/SC 37 yang memberi tumpuan kepada teknologi yang berkaitan dengan biometrik. Sub-komite standardisasi ini sangat berkaitan dengan deepfake kerana muka deepfake boleh digunakan untuk menipu sistem pengesahan pengguna berasaskan biometrik. 15 : Standard ini merangkumi metrik umum mengenai penilaian sistem biometrik. (FAR) and (FRR), which refer to the standard FPR and FNR, respectively. This standard also deprecates the use of single-number metrics including the EER and AUC (which were widely used in biometrics-related research in the past). ISO/IEC 19795-1:2021 “Information technology – Biometric performance testing and reporting – Part 1: Principles and framework” 16 false accept rate false reject rate : This standard defines a general framework about (PAD) mechanisms, where the term “ ” refers to the “ Ia memberi tumpuan kepada sistem pengiktirafan biometrik, di mana mekanisme PAD adalah klasifikator binari yang cuba untuk meramalkan serangan persembahan (juga dipanggil persembahan serangan, contohnya, muka palsu) sebagai positif dan persembahan bona fide (benar) sebagai negatif. ISO/IEC 30107-1:2016 “Information technology – Biometric presentation attack detec-tion – Part 1: Framework” 17 presentation attack detection presentation attack penyampaian artefak atau ciri-ciri manusia kepada subsistem penangkapan biometrik dengan cara yang bertujuan untuk mengganggu dasar sistem : This standard defines a number of special performance metrics for evaluating PAD mechanisms standardised in the ISO/IEC 30107-1:2016. Three such metrics look at error rates: (APCER) yang merujuk kepada standard FPR, (NPCER/BPCER) yang merujuk kepada standard FNR, dan (ACER) that is defined as the average of the APCER and the NPCER/BPCER. Such metrics have been used in biometrics-related challenges such as Face Anti-spoofing (Presentation Attack Detection) Challenges . When deepfake images or videos are used to spoof a biometric system, such standardised metrics will become relevant. ISO/IEC 30107-3:2017 “Information technology – Biometric presentation attack detection – Part 3: Testing and reporting” 18 attack presentation classification error rate normal/bona fide presentation classification error rate average classification error rate 19 3.13 Perbincangan: Metrik dan piawaian prestasi This section provided a comprehensive summary of performance metrics used for evaluating and bench-marking binary classifiers. It is rare that all such metrics are used for a specific application. Instead, one or several are chosen based on specific needs. For a deepfake detection system as a binary classifier, many researchers have chosen to use overall metrics such as accuracy, AUC, EER and log loss, but the combination of precision, recall and F1-score is also common. Some deepfake-related challenges and competitions have introduced their own specific metrics, some of which will be described in Section . The use of different performance metrics can make comparison of different reported results more difficult, so we hope the expected new ISO/IEC standard particularly ISO/IEC 4213 will help. 5 It is worth mentioning that, in addition to evaluating performance of deepfake detectors, the introduced performance metrics for evaluating binary classifiers can also be used to evaluate performance of deepfake generation methods by considering how deepfake detectors fail. For instance, organisers of the Voice Conversion Challenge 2018 and 2020 used this approach to benchmark how well voice conversion (VC) systems can generate high-quality fake speech samples. Another point we would like to mention is that for deepfake videos there are two levels of performance metrics: those at the frame level (metrics of each frame), and those at the video level (metrics for the whole video). Generally speaking, the latter can be obtained by averaging the former for all frames, potentially following an adaptive weighting scheme, so that more important (key) frames will be counted more. 4 Deepfake-Related Datasets In this section, we cover all deepfake-related datasets we identified from the meta-review of deepfake-related survey papers, deepfake-related challenges, competitions and benchmarks covered, one online collection of deepfake-related datasets on GitHub , and the co-authors’ personal collections. Table shows basic information about these datasets. We explain them in four categories: deepfake image datasets, deepfake video datasets, deepfake audio/speech datasets, and hybrid deepfake datasets (mainly mixed image and video datasets). 20 2 Note that many datasets of real (authentic) media were also used by deepfake researchers for two purposes. First, any detectors would need both fake and real media to demonstrate their performance. Second, real media have also been used to train deepfake generators as the training set. In this section, we include only datasets containing deepfake media, some of which contain both deepfake and real media. Some datasets, especially those created for deepfake-related challenges and competitions, have separate subsets for training and evaluation (testing) purposes. The split is necessary for such challenges and competitions, but not very useful for people who just want to use such datasets. Therefore, in this section when introducing such datasets we will ignore that level of details and focus on the total number of data including the number of real and fake samples. 4.1 Deepfake Image Datasets [ ]: This dataset contains 4,310 images, including 2,300 real images and 2,010 fake images created using FaceSwap and the SwapMe iOS app (now discontinued). SwapMe and FaceSwap dataset 78 21 [ ]: This dataset contains 131,500 face images, including 78,500 images extracted from 150 videos in the FaceForensics dataset and 53,000 images extracted from 150 fake videos collected from YouTube. Fake Faces in the Wild (FFW) dataset 32 : This is a number of commercial datasets provided by the Generated Media, Inc., with up to nearly 2.7 million synthetic face images generated by StyleGAN. A free edition with 10,000 128x128 synthetic images is made available for academic research. The website also provides an interactive face generator and an API . The generated.photos datasets have a good diversity: five age groups (infants, children, youth, adults, middle-aged), two genders (male and female), four ethnicities (white, black, Latino, Asian), four eye colours (brown, grey, blue, green), four hair colours (brown, black, blond, gray), three hair length (short, medium, long), facial expressions, three head poses (front facing, left facing, right facing), two emotions (joy and neutral), two face styles (natural, beautified). (According to a number of research papers we read, an earlier 100K-Faces dataset was released by generated.photos for academic research in 2018, which was used by many researchers. This dataset is not currently available any longer.) generated.photos datasets 22 23 24 [Bahasa Melayu ]: This dataset includes 19,457 face images, including 7,948 deepfake images generated from on 175 forged videos collected online and 11,509 real face images collected from various online sources. (Table 2 of the paper shows the dataset size is 19,509, but the dataset downloaded from pCloud contains just 19,457 images.) MesoNet Deepfake Dataset 1 [Bahasa Melayu ]: This dataset includes 100,000 synthesised face, bedroom, car and cat images by a GAN generator trained based on real images in the FFHQ and LSUN datasets (three object types – bedrooms, cars and cats – for the latter). Note that the name “100K-Generated-Images” was not a proper one as the authors [ ] just used this to name a sub-folder of their Google Drive shared space, but it was used in one of the survey papers [ ]. 100K-Generated-Images 30 25 26 30 65 [ Set data ini mengandungi 420,053 imej selebriti, termasuk 156,930 imej sebenar yang dimuat turun menggunakan Google Image API dan 263,123 imej palsu yang ditukar muka yang dibuat menggunakan dua kaedah yang berbeza (Metode Nirkin dan Auto-Encoder-GAN) Ding et al.’s swapped face dataset 17 [ ]: This dataset includes 87,000 224x224 face images, generated by processing some StyleGAN-generated synthetic images using the GAN-fingerprint Removal approach (GANprintR) proposed by . It is the replaced version of the dataset, which contains 150,000 face images generated using an earlier version of GANprintR. iFakeFaceDB 48 Neves et al. FSRemovalDB [ ]: This dataset includes 40,000 images, half real and half deepfake. The images were collected from four sources: the CelebA-HQ dataset Kumpulan Data Flickr-Faces-HQ , the 100K-Faces dataset (not available any longer, see the description of generated.photos datasets), and . Faces-HQ 21 27 28 29 thisperson-doesnotexist.com [ Set data ini termasuk 625,537 imej muka disintesis daripada 10,177 selebriti, dengan 43 atribut kaya pada muka, pencahayaan, persekitaran dan jenis spoof. 43 atribut termasuk 40 untuk imej sebenar, yang meliputi semua komponen dan aksesori muka (contohnya, kulit, hidung, mata, mata, bibir, rambut, topi, kacamata), dan 3 untuk imej palsu, yang meliputi jenis spoof, persekitaran dan keadaan pencahayaan. CelebA-Spoof 75 30 [ Set data ini mengandungi 299,039 imej, termasuk 58,703 imej sebenar yang diambil sampel daripada tiga set data (FFHQ) , CelebA and FaceForensics++ ) and 240,336 fake ones in four main facial manipulation types (identity swap, expression swap, attribute manipulation, and entire synthesis). The images cover two genders (male and female), a wide age groups (the majority between 21 and 50 years old), and both low- and high-quality levels. Diverse Fake Face Dataset (DFFD) 11 31 32 33 4.2 Deepfake Video Datasets [Bahasa Melayu ]: This dataset contains 620 deepfake face videos, generated by face swapping without manipulation of audio, covering 32 subjects and two quality levels (high and low). DeepfakeTIMIT 35 (FF) [ ]: This dataset contains 1,004 face videos with over 500,000 frames, covering various quality levels and two types of facial manipulation. This dataset is now replaced by the larger FaceForensics++ dataset (see below). FaceForensics 55 (FF++) [ ]: This dataset contains 5,000 face videos with over 1.8 million manipulated frames, including 1,000 real videos (with 509,914 frames) downloaded from YouTube, and 4,000 fake videos created using four face manipulation methods (Deepfakes, Face2Face, FaceSwap and NeuralTextures). The videos cover two genders (male and female), and three quality levels (VGA/480p, HD/720p, and FHD/1080p). FaceForensics++ 56 [ ]: This dataset contains 98 face videos, half (49) are real ones downloaded from Youtube, and the other half are fake ones generated using the FakeApp mobile application (which is now discontinued). The video dataset was created to used to demonstrate a deepfake video detection method based on detection of eye blinking behaviours, so all videos contain at least one eye-blinking event. All fake videos were created by swapping the original face in each of the real videos with the face of the actor Nicolas Cage , thus, only one subject is represented. UADFV dataset 39 34 [ ]: This dataset contains 142 “in the wild” deepfake portrait videos, collected from a range of online sources including news articles, online forums, mobile apps, and research presentations. The videos are diverse, covering the source generative model, resolution, compression, illumination, aspect-ratio, frame rate, motion, pose, cosmetics, occlusion, content, and context. Deep Fakes Dataset 10 [ ]: This dataset contains 5,244 face videos of 66 subjects with both face and voice manipulation. It was released as a preview of the full dataset of the 2020 Deepfake Detection Challenge (DFDC, see below). DFDC (Deepfake Detection Challenge) preview dataset 18 : This dataset contains 1,203 face videos of celebrities, including 408 real videos collected from YouTube with subjects of different ages, ethic groups and genders, and 795 deepfake videos synthesised from these real videos. Celeb-DF v1 35 [Bahasa Melayu ]: This dataset contains 6,229 face videos of celebrities, including 590 real videos collected from YouTube with subjects of different ages, ethic groups and genders, and 5,639 deepfake videos synthesised from these real videos. Celeb-DF v2 40 [ Set data ini mengandungi 3,363 video muka, meliputi 28 subjek, jantina, dan warna kulit. ia telah dicipta sebagai usaha bersama antara dua unit Google, Inc.: Google AI and JigSaw . DeepFake Detection (DFD) Dataset 20 36 37 [ ]: This dataset contains 60,000 indoor face videos (with 17.6 million frames) generated by face swapping, covering 100 subjects, four skin tones (white, black, yellow, brown), two gen-ders (male and female), different age groups (20-45), 26 nationalities, 7 different angles, 8 face expressions, and different head poses. DeeperForensics-1.0 27 [ ]: This dataset contains 128,154 face videos of 960 subjects, including 23,654 real videos from 3,426 paid actors and 104,500 deepfake videos created using eight different methods (DF-128, DF-256, MM/NN face swap, NTH, FSGAN, StyleGAN, refinement, and audio swap). DFDC (Deepfake Detection Challenge) full dataset 18 10 [ ]: This dataset contains 10,000 high-quality forgery videos, with video- and face-level annotations. The dataset focuses on a more challenging case for forgery detection: each video involves one to 15 individuals, but only some (a minority of) faces are manipulated. FFIW K (Face Forensics in the Wild) dataset 79 FFIW [ ]: This dataset contains 37,942 videos of paid subjects (395 Koreans and 8 Southeastern Asians), including 62,166 real videos and 175,776 fake ones created using six methods – FaceSwap, DeepFaceLab, FSGAN, First Order Motion Model (FOMM), Audio-driven Talking Face HeadPose (ATFHP) and Wav2Lip. The videos cover a balanced gender ratio and a wide range of age groups. Korean DeepFake Detection Dataset (KoDF) 36 [ ]: This dataset contains 1,737 videos with 1,666,816 frames, including 1,339,843 real frames and 326,973 fake frames generated using the Deep Video Portraits (DVP) [ ] method. The original videos were obtained from three sources: the dataset used in [ ], the Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS) [ Kebanyakan video mempunyai resolusi 1280×720. VideoForensicsHQ 23 34 33 42 [ ]: This dataset contains 7,314 face sequences extracted from 707 deepfake videos that were collected completely from the Internet. It covers diverse scenes, multiple persons in each scene and rich facial expressions. Different from other deepfake video datasets, WildDeepfake contains only face sequences not the full videos. This makes the dataset more like between an image dataset and a video one. We decided to keep it in the video category since the selection process was still more video-focused. WildDeepfake 81 4.3 Deepfake Audio/Speech Datasets Voice conversion (VC) is a technology that can be used to modify an audio and speech sample so that it appears as if spoken by a different (target) person than the original (source) speaker. Obviously, it can be used to generate deepfake audio/speech samples. The biennial Voice Conversion Challenge that started in 2016 is a major challenge series on VC. Datasets released from this challenge series are very different from other deepfake datasets: the deepfake data is not included in the original dataset created by the organisers of each challenge, but in the participant submissions (which are retargeted/fake utterances produced by VC systems built by participants). The challenge datasets also include the evaluation (listening-based) results of all submissions. Some fake utterances may be produced by DL-based VC systems, so we consider all datasets from this challenge series relevant for our purpose of this survey. 38 [ ]: The original dataset created by the challenge organisers was derived from the DAPS (Device and Produced Speech) Dataset [ ]. It contains 216 utterances (162 for training and 54 for testing) per speaker from 10 speakers. Participating teams (17) developed their own VC systems for all 25 source-target speaker pairs, and then submitted generated utterances for evaluation. At least six participating teams used DL-related techniques (LSTM, DNN) in their VC systems (see Table 2 of the result analysis paper ), so the submitted utterances can certainly be considered deepfakes. Voice Conversion Challenge 2016 database 62 47 39 [ ]: The original dataset created by the challenge organisers was also based on the DAPS dataset. It contains 116 utterances (81 for training and 35 for testing) per speaker from 12 speakers in two different tasks (called Hub and Spoke). Participating teams (23 in total, all for Hub and 11 for Spoke) developed their own VC systems for all 16 source-target speaker pairs, and then submitted generated utterances for evaluation. Comparing with the 2016 challenge, more participating teams used DL-related techniques (e.g., WaveNet, LSTM, DNN, CycleGAN, DRM – deep relational models, and ARBM – adaptive restricted Boltzmann machines) in their VC systems. Voice Conversion Challenge 2018 database 44 [ ]: This dataset is based on the Effective Multilingual Interaction in Mobile Environments (EMIME) dataset , a bilingual (Finnish/English, German/English, and Mandarin/English) database. It contains 145 utterances (120 for training and 25 for testing) per speaker from 14 speakers for two different tasks (with 4 × 4 and 4 × 6 source-target speaker pairs, respectively). Participating teams (33 in total, out of which 31 for Task 1 and 28 for Task 2) developed their own VC systems for all source-target speaker pairs, and then submitted generated utterances for evaluation. Comparing with the 2018 challenge, DL-based VC systems were overwhelmingly used by almost all participating teams (WaveNet and WaveGAN among the most used DL-based building blocks). Voice Conversion Challenge 2020 database 70 40 A major set of deepfake speech datasets were created for the (Automatic Speaker Verification Spoofing and Countermeasures) Challenge (2015-2021, held biannually). The datasets for the 2019 and 2021 contain speech data that can be considered deepfakes. ASVspoof 41 [ ]: This dataset is based on the Voice Cloning Toolkit (VCTK) corpus , pangkalan data bahasa Inggeris berbilang penutur ditangkap daripada 107 penutur (46 lelaki dan 61 wanita). Dua skenario serangan dipertimbangkan: akses logik (LA) yang melibatkan ucapan yang disalahpahami (sintetik atau ditukar) dan akses fizikal (PA) yang melibatkan serangan pemulihan rekod bonafide yang disenaraikan sebelumnya). Untuk tujuan kami dalam kajian ini, skenario LA lebih relevan. Bahagian LA dalam set data termasuk 12,483 ucapan bonafide (sebenar) dan 108,978 ucapan disalahpahami. Sesetengah data ucapan yang disalahpahami untuk skenario LA dihasilkan menggunakan model generatif yang melibatkan teknik berbasis DL seperti ingatan jangka pendek (LSTM) , WaveNet [ ], WaveRNN [ ], WaveCycleGAN2 [ ]. Note that the challenge organisers did not use the term “deepfake” explicitly, despite the fact that the DL-generated spoofed speech data can be considered as deepfakes. ASVspoof 2019 Challenge database 67 42 43 50 28 58 [Bahasa Melayu ]: This dataset contains bona fide and spoofed speech data for the logical access (LA) task. The challenge is still ongoing and we did not find a detailed paper on the dataset, so cannot include more details other than its size (7.8 GB after compression). Although we did not see details of the generative algorithms used to produce spoofed speech data, we believe similar DL-based algorithms were used like for the 2019 challenge. ASVspoof 2021 Challenge – Logical Access Database 14 [ ]: In 2021, the challenge included an explicitly defined track on deepfake, but the task description suggests that the organisers of the challenge considered a broader definition of the term “deepfake” by looking at spoofing human listeners rather than ASV (Automatic Speaker Verification) systems. The size of the dataset is 34.5 GB after compression. ASVspoof 2021 Challenge – Speech Deepfake Database 15 Possibly because of the long history and wide participation of the community in the ASVspoof challenges for creating the dedicated datasets, there are very few other deepfake audio/speech datasets. One such dataset was created by a group of researcher from Baidu Research [ ]. This dataset was created to demonstrate a proposed voice cloning method. It is relatively small, and contains 134 utterances, including 10 real ones, 120 cloned ones, and 4 manipulated ones. Another dataset was created by Google AI and Google News Initiative , but it was made part of the ASVspoof 2019 dataset. This dataset contains thousands of phrases spoken by 68 synthetic “voices” covering a variety of regional accents. 5 44 4.4 Hybrid Deepfake Datasets : These datasets were created by the DARPA Media Forensics (MediFor) Program for the 2020 OpenMFC . There are two GAN-generated deepfake datasets, one with more than 1,000 deepfake images and the other with over 100 deepfake videos. The datasets were made available to registered participants of the competition only. NIST OpenMFC (Open Media Forensics Challenge) Datasets 45 46 47 [ ]: This dataset is named as “a versatile benchmark for comprehensive forgery analysis”. It contains 2,896,062 images and 221,247 videos, including 1,457,861 fake images and 121,617 fake videos. The videos and images cover seven image-level and eight video-level manipulation approaches, 36 different types of perturbations and more mixed perturbations, and a large number of annotation labels (6.3 million classification labels, 2.9 million manipulated area annotations and 221,247 temporal forgery segment labels). The dataset is being used for supporting the Face Forgery Analysis Challenge 2021 at the SenseHuman 2021 (3rd Workshop on Sensing, Understanding and Synthesizing Humans) , co-located at the ICCV 2021 conference . ForgeryNet 25 48 49 50 5.4 Pengeluar Dataset Deepfake [ ]: This is not actually a dataset per se, but a system for producing large datasets more automatically, including generating deepfake datasets. One may argue the automatically generated datasets are fake since they are not produced from real-world scenes. DatasetGAN 74 4.6 Subjective Quality of Deepfakes in Different Databases As mentioned in Section , subjective quality evaluation is necessary to evaluate the realness, realisticness, and naturalness of deepfake media. While there has been very limited work on this topic, in 2020, Jiang et al. [ ] conducted a user study on realness of deepfake videos. They recruited 100 professional participants (most of whom are computer vision researchers), who were asked to evaluate the realness of 30 randomly selected videos from 7 deepfake video datasets (DeeperForensics-1.0, UADFV, DeepFake-TIMIT, Celeb-DF, FaceForensics++, Deep Fake Detection, and DFDC). Participants were asked to respond to the statement “The video clip looks real.” and gave scores following a five-point Likert scale (1 – clearly disagree, 2 – weakly disagree, 3 – borderline, 4 – weakly agree, 5 – clearly agree). 4.7 27 Table shows the results. Interestingly, we can see a huge difference between the realness levels of different datasets. What is probably quite surprising is that FaceForensics++, one of the most widely used deepfake datasets, has a very low MOS score and less than 9% of participants considered the 30 selected videos as real. 3 Jadual 3: Kualiti subjektif yang dinilai oleh manusia (realitas) video deepfake dalam 7 set data. , but calculated by us based on the raw data shown in Table 3 of [ ]. Jiang dan al. 27 4.7 Discussion: Datasets Among all deepfake image and video datasets, a significant majority are about face images and videos. This is not surprising since face swapping, face attribution manipulation, and fully synthesised face images are among the hottest topics within deepfake research and real-world applications. We hope more non-face deepfake image and video datasets can be produced to support a broader range of research activities on deepfake. Hasil kualiti subjektif yang dipaparkan dalam Jadual indicate that it is important to check realness of deep-fake media to support any performance evaluation or comparison. To ensure that the quality evaluation of datasets is fair, transparent and reliable, standard procedures need defining and a common pool of qualified human experts should be used. 3 Many authors of deepfake-related datasets attempted to classify such datasets into different generations. Chronologically speaking, we could broadly split such datasets into two generations: before 2019 and since 2019. Typically, datasets created before 2019 are relatively less advanced and smaller, while those created after 2019 tend to be larger, more diverse (i.e., covering more attributes), and of higher quality (i.e., produced by more advanced generative models). This can also be seen from the data in Table , di mana dua set data teratas (DeeperForensics-1 dan Celeb-DF) termasuk dalam generasi baru (2020), manakala yang lain termasuk generasi lama. selain daripada dua generasi, generasi yang lebih baru juga telah muncul pada tahun 2021: beberapa set data yang sangat baru mula memberi tumpuan kepada deepfakes yang lebih realistik (iaitu, di alam liar) atau kawasan yang lebih tertentu deepfakes (contohnya, 10 focusing on multiple faces in the same video, and KoDF focusing on Korean faces). This trend shows that the deepfake research community has grown significantly in the past few years so that narrower topics have also started gaining attention and interest from some researchers. 3 FFIW K 5 Deepfake-Related Challenges, Competitions & Benchmarks This section reviews initiatives aiming to advance the state-of-the-art of detection and generation of synthetic or manipulated media (such as video, image and audio) via competitions or challenges open to the public, and ongoing benchmarks tackling specific problems. 5.1 Detection of Manipulated Media The Deepfake Detection Challenge (DFDC) was an initiative promoted by an AI and Media Steering Committee , including BBC, Facebook, Amazon, Microsoft and New York Times, and some universities around the world including the University of Oxford. The competition remained open from 5 September 2019 till 31 March 2020, and involved 3 stages. At first, the DFDC preview dataset was released. At a later stage, the DFDC full dataset was also made available to the 2,114 participants of the competition incorporating face and audio swap techniques for generation of deepfake content. At the final stage, the submitted models were evaluated using a test dataset (referred to as the “black box dataset”) of 10,000 videos which included prestasi terbaik pada set data kotak hitam mempunyai ketepatan 65.18%, menurut hasil yang dikeluarkan [ ]. Submissions were ranked according to the overall log loss score, as defined in Eq. ( ). All top five ranked models (the winner had the lowest overall log loss) are available on GitHub. Results indicate how challenging the detection of deepfake is since the best accuracy was low and “ ”, according to Dolhansky et al. [ ]. Figure shows a screenshot of the leaderboard with the five finalists. The first top ranked model used MTCNN (Multi-tasked Cascaded Convolutional Network), the second used WS-DAN (Weakly Supervised Data Augumentation Network), and the third used the EfficientNetB7 architecture. Meta compiling the common themes observed in the winning models, they were: clever augmentations, architectures, and absence of forensics methods. Moving forward, they called for “ ”. 51 52 in-the-wild 22 53 13 many submissions were simply random 19 2 solutions that go beyond analysing images and video. Considering context, provenance, and other signals may be the way to improve deepfake detection models The Automatic Speaker Verification Spoofing And Countermeasures Challenge Workshop (ASVspoof) has been running biennially since 2015. This competition is organised by an international consortium that includes Inria and EURECOM (France), University of Eastern Finland, National Institute of Informatics (Japan), and Institute for Infocomm Research (Singapore). This year the ASVspoof challenge includes, for the first time, a sub-challenge focused on where the envisioned use case is an adversary trying to fool a human listener. The metric used for evaluating performance of submitted solutions (i.e., classifiers) is EER. Four baseline solutions (also called “countermeasures”), each using a different technique, were made available to participants with their corresponding EER metric values. The ASVspoof 2021 Speech Deepfake Database containing audio recordings with original and spoofed utterances has also been made available. The competition involves three phases : a progress phase, an evaluation phase and a post-evaluation phase; it is unclear how teams move from one phase to the next. More information about the 2021 competition is available in the published evaluation plan [ ]. The organisers of the competition noted that they opted for the EER as the performance evaluation met-ric for countermeasures submitted to the speech deepfake task for legacy reasons. They acknowledged, however, that “ ” by the ISO/IEC 19795-1:2021 standard. Despite the fact that only the 2021 ASVspoof competition contained a track explicitly related to deepfake, some data in the ASVspoof 2019 dataset (Logical Access task) used for the 2019 competition was generated using DL-based algorithms as mentioned in Section . We expect that this also holds for the ASVspoof 2021 dataset (Logical Access task). The ASVspoof 2019 competition used the EER as secondary metric; the primary performance metric used was the tandem detection cost function (t-DCF) [ ]. According to its evaluation plan [ ], t-DCF menilai prestasi keseluruhan sistem tandem di mana " ”. It is calculated according to Eq. ( ), where cm ( ) and cm( ) are, respectively, “ ”. 54 Speech DeepFake 55 56 13 EER reporting is deprecated 57 4 63 69 a CM berfungsi sebagai 'pintu gerbang' untuk menentukan sama ada input ucapan yang diberikan berasal daripada pengguna bona fide (asli), sebelum melepasi pengesahan biometrik utama (sistem ASV) 17 P s P s the miss rate and the false alarm rate of the CM system at threshold s For further information about Eq. ( ), including constants 1 and 2, please refer to the ASVspoof 2019 evaluation plan [ ]. 17 C C 69 An implementation of the t-DCF metric has been made available by the ASVspoof 2019’s organisers in Python daripada MATLAB formats. 58 59 Tindakan Anti-Spoofing Face (Presentation Attack Detection) started in 2019. Its first two editions were held at the 2019 and 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2020), respectively. Its third edition was moved to be co-located with the 2021 IEEE/CVF International Conference on Computer Vision (ICCV 2021). This competition series was organised by a group of researchers from academia and industry in China, Mexico, Spain, Finland and the US. The 2021 competition was focused on 3D high-fidelity mask attacks, and followed a 2-phased process. The first phase is the “development phase”; it started in April 2021 when the CASIA-SURF HiFiMask dataset was released to participants. The second phase is the “final ranking phase” (June 2021), when the competition ended. The competition adopted the following performance metrics for evaluation daripada penyelesaian yang diajukan: kadar kesilapan penilaian persembahan serangan (APCER), kadar kesilapan penilaian persembahan normal/bona fide (NPCER/BPCER), dan kadar kesilapan penilaian purata (ACER), mengikut ISO/IEC 30107-3:2017 Kegunaan : Figure provides the leaderboard for the top three solutions. 60 61 62 63 64 3 Penilaian Benchmark is an ongoing automated benchmark for detection of face manipulation. The organisers of the benchmark made the FaceForensics++ dataset available for training. Manipulated videos (4,000 in total) were created using four techniques, i.e., two computer graphics-based approaches (Face2Face and FaceSwap) and two learning-based approaches (DeepFakes and Neural Textures). The deepfakes videos were generated using a slightly modified version of FaceSwap , and the Neural Textures videos were created using the approach proposed by Thies et al. [ ]. The benchmark test dataset is created from the collection of 1,000 images randomly selected from either the manipulation methods or the original videos [ ]. Participants have to submit results to the benchmark, rather then code like other competitions; this is illustrated in Figure . The outcome of a submission is illustrated in Figure , where the scores are a measure of accuracy (Eq. ( )). 65 66 61 56 4a 4b 8 The Open Media Forensics Challenge (OpenMFC, formerly DARPA MFC) is an annual image and video forensics evaluation aiming to facilitate development of multimedia manipulation detection systems. It has been organised annually starting from 2017 under the name of DARPA MFC. In 2020, the National Institute of Standards and Technology (NIST) initiated the as a new evaluation platform, based on their previous experiences with the DARPA MFC series, to make the participation more convenient for all researchers. In OpenMFC 2020, two deepfake-related tasks were included for the first time: Image GAN Manipulation Detection (IGMD) and Video GAN Manipulation Detection (VGMD). The organisers provided an image evaluation dataset for the IGMD task, containing 1,000 images from over 200 image journals , and a video evaluation dataset for the VGMD task, including over 100 test videos. Furthermore, they provided the datasets used in the previous MFC challenges as development datasets. The challenge is composed of two main phases for development and evaluation, respectively, and a pre-challenge phase for quality control testing. For evaluation of submissions, AUC-ROC is used as the primary metric. Furthermore, CDR@FAR, where CDR refers to correct detection rate or TPR (Eq. ( )) and FAR refers to false alarm rate or FPR (Eq. ( )), juga digunakan sebagai metrik [ ]. The DeeperForensics Challenge 2020 is a deepfake face detection challenge held at the 2020 ECCV 67 68 OpenMFC 69 70 4 5 49 71 SenseHuman Workshop . The challenge used the DeeperForensics1.0 dataset. 72 The organisers provided a hidden test dataset to better simulate real-world scenarios. The challenge involved two phases: the “development phase” that started in August 2020 allowing 100 successful sub-missions, and the “final test phase” that started in October 2020 allowing 2 successful submissions until the end of the month. The submissions were evaluated using the binary cross-entropy loss (BCELoss) metric, calculated according to Eq. ( ), where is the number of videos in the hidden test set, Adalah Label Kebenaran Tanah Video (fake:1, real:0), and ( ) is the predicted probability that video is fake. 18 N yi i p yi i Results of the competition were discussed by Jiang et al. [ ]. The top solution used three models, i.e., EfficientNet-B0, EfficientNet-B1 and EfficientNet-B2, for classification. The second top used EfficientNet-B5 for both an image-based model and a video-based model. The third ranked solution used a 3D convolutional neural network (3DCNN). 73 26 The Face Forgery Analysis Challenge 2021 is a competition hosted at the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2021). It is organised by researchers from a number of organisations in China including universities and SenseTime Research (the research arm of SenseTime , one of the major AI “unicorns” in China). The challenge aims to advance the state-of-the-art in detection of photo-realistic manipulation of images and videos. Participants are able to use a large annotated face dataset (i.e., the ForgeryNet dataset) that was obtained by applying a number of techniques for manipulation (15) and perturbation (36) to train their solutions. The phases comprise of Forgery Image Analysis, Forgery Video Analysis, Forgery Video Temporal Localization phases, and the final phase (i.e., “private test”) where participants’ models will be tested against an unseen dataset. The following metrics will be used [ ]: AUC, average precision (AP) at some “temporal Intersection over Union” (AP@tIoU) compared to a threshold ∈ [0*. ,* 0*.*95) dan peringatan purata (AR) pada (Bahasa Melayu ) where is the top labels returned for multi-class classifiers. 74 75 25 t 5 K K K K The 2020 CelebA-Spoof Face Anti-Spoofing Challenge was hosted at the 16 European Conference on Computer Vision (ECCV 2020). The challenge ran between August and October 2020, and aimed to advance the state-of-the-art in detecting “ ” [ ]. The organisers made the face CelebA-Spoof dataset available for the competition containing rich annotation across a range of attributes. The competition only had one phase where participants submitted their solutions to be evaluated against a test dataset; the spoof class was considered as “positive” and the live class as “negative”. Metric TPR@FPR was used and collected at three points where the TPR when FPR = 10 4 determined the final ranking. The top three finalists (see Figure ) used deep learning models ResNet, EfficientNet-B7, and a novel architecture combining Central Difference Convolutional Networks (CDCN) and Dual Attention Network (DAN). The two top ranked solutions used different strategies to boost their models’ performance: a heuristic voting scheme was used by the top-ranked solution, and a weight-after-sorting strategy was used by the second ranked solution. 76 th sama ada muka yang dipaparkan ialah hidup atau berbohong 76 − 5 Peluang 2019 CSIG is the second edition of a challenge organised by the China Society of Image and Graphics . The 2021 challenge has the Fake Media Forensic Challenge as its 6 track, co-organised by CSIG’s Digital Media Forensics and Security Technical Committee and Institute of Information Engineering, Chinese Academy of Sciences . This track has two tasks, one on deepfake video detection, and the other on deepfake audio/speech detection. For the deepfake video detection task, the dataset used contains a public training set with 10,000 sound-free face videos (including 4,000 fake videos), a public test set with 20,000 face videos (the percentage of deepfake videos is unknown to participants), and a private test set that will be determined and used at the final session for selecting the winners. All videos contain faces of Eastern Asian people, and cover a wide range of parameters such as multiple resolutions and encoding quality factors, the use of blurring or sharpening filters, and added noise. Deepfake videos were created using public tools including DeepFaceLab [ ], Faceswap Perbezaan : Perbezaan : Perbezaan : Perbezaan ] and ALAE (Adversarial Latent Autoencoders) [ ]. For the deepfake audio/speech detection task, the dataset used contains a public training set with 10,000 speech samples (including 6,000 fake ones), a public test set with 20,000 face videos (the percentage of deepfake videos is unknown to participants), and a private test set for the final session (the same as the deepfake video detection task). The tools used for generating the fake speech samples include TTS (text-to-speech) voice synthesis tools and VC (voice conversion) tools. The main TTS tools used include open-source tools such as DeepVoice, TensorFlowTTS and GAN-TTS [ ] and commercial software tools such as those from iFlytek and IBM. The main VC tools used include Adaptive-VC and CycleGAN-VC [ ]. For both deepfake detection tasks, the performance metric used is log loss. 77 78 79 th 80 81 53 82 6 54 83 8 84 29 2020 China Artificial Intelligence was the second edition of a Chinese AI competition open for the general public to participate, organised by the municipal government of the City of Xiamen in China. In 2020, it had two sub-competitions, Multimedia Information Recognition Technology Competition and Language and Knowledge Technology Competition . The Multimedia Information Recognition Technology Competition included two tasks on deepfakes: one on deepfake video detection dan satu pada pengesanan audio / suara deepfake . The deepfake video detection task used 3,000 videos, and log loss was used as the sole performance metric. The deepfake audio/speech detection task used 20,000 audio samples (mostly in Chinese, and the remaining in English), and EER was used as the sole performance metric. For both tasks, the ratio between real and deepfake samples was 1:1. We did not find where to download the datasets used for the tasks nor a more detailed technical description of the datasets. For the deepfake video detection tasks, the top two winning teams (with an A prize) were from Netease (Hangzhou) Network Co., Ltd. and Beijing RealAI Technology Co., Ltd., followed by three other teams winning a B prize: Xiamen Fuyun Information Technology Co., Ltd.; Institute of Computing Technology, Chinese Academy of Sciences; and Wuhan Daqian Information Technology Co., Ltd. For the deepfake audio/speech task, there was no team winning an A prize, but one team winning a B prize: SpeakIn Technologies Co., Ltd. The final results of some teams were published, but some teams were allowed to hide their results. We did not find a detailed technical report summarising the results and explaining the work of the winning teams. 85 86 87 88 89 One of the B-prize winning team is from Beijing RealAI Technology Co., Ltd., a Chinese company active in deepfake-related R&D. 5.2 Generation of Manipulated Media The Voice Conversion Challenge is a biennial competition that has been running since 2016. The challenge and the corresponding workshop, hosted at the INTERSPEECH conference , is supported by the SynSig (Speech Synthesis Special Interest Group) of the International Speech Communication Association (ISCA) Tujuannya adalah untuk mempromosikan kemajuan dalam teknologi penukaran suara (VC) yang boleh digunakan untuk pelbagai kes penggunaan positif dan negatif, seperti sistem biometrik suara yang merosakkan. Cabaran 2020 memberi tumpuan kepada penukaran penutur, sub-masalah VC, dan termasuk dua tugas. Untuk tugas pertama "penukaran suara semi-paralel intra-bahasa", peserta terpaksa membangunkan 16 sistem VC (kombinasi penutur-pasangan) termasuk penutur lelaki dan wanita dan ayat-ayat Bahasa Inggeris, menggunakan pangkalan data Voice Conversion Challenge 2020 v1.0 yang disediakan untuk latihan (lihat Seksyen ). For the second task “cross-lingual voice conversion”, participants had to develop 24 VC systems, also including male and female speakers, but uttering sentences in three languages (Finnish, German and Mandarin), based on the provided training dataset. Figure illustrates the process of training and generation of VC systems. 90 91 92 93 4 6 Submissions were evaluated for “ ” . As such, the organisers used [ ] and recruited both native and non-native English speakers (i.e., Japanese native speakers) via crowd-sourcing for the listening tests. Naturalness (answering the question “How natural does the converted voice sound? ”) was measured using the metric MOS (covered in Section 4.6), and similarity (answering the question “how similar the converted voice sound comparing source and target speakers? ”) was measured in terms of speaker recognition as “same” or “different”, as elaborated by Wester et al. [68]. Tests also focused on the effects of language differences on the performance of VC systems submitted to the competition. The most popular CNN/RNN/GANbased VC systems submitted used WaveNet, WaveRNN, and Parallel WaveGAN. Results indicated that, in terms of similarity, the best performing VC systems were as good as natural speech but none reached human-level naturalness for task 1; scores were lower for task 2 which was more complex [70]. The organisers of the 2020 competition also used objective evaluation [12]. The metrics used for evaluation of speaker similarity were: equal error rate (EER), false acceptance rate of target (P tar fa ), miss rate of source (P src miss), and cosine similarity of speaker embedding vectors (cos-sim) according to Eq. (19) where A is the speaker embedding vectors for the converter audio and B is the speaker embedding vectors for the original audio. The performance of the VC systems as a spoof countermeasure was also evaluated using EER, while to evaluate the quality of the subjective MOS obtained via listening tests, a DL-based model to predict MOS, called MOSNet [43], was used. Lastly, to evaluate intelligibility of the converted transcribed speech, in comparison with the original transcribed speech, the word error rate (WER) [4] was used. WER is calculated according to Eq. (20) where I refers to insertions, D refers to deletions, S refers to substitutions, and N refers to the total number of words in the original transcript. perceived naturalness and similarity through listening tests 94 subjective evaluation 70 The Deepfake Africa Challenge (2021) is a new initiative of the AI Africa Expo, in partnership with a film and media production company (Wesgro) and the African Data Science competition platform Zindi. Its aim is “ Para peserta yang layak diminta untuk menjadi warganegara dan penduduk benua Afrika. Submission, diterima sehingga akhir Julai 2021, boleh menjadi video atau audio. Penilaian submission ditakrifkan dalam hal kreativiti artistik, relevansi topik cabaran, dan inovasi dalam proses pengeluaran selagi peserta menggunakan alat dan paket yang tersedia secara awam. Tiga finalis teratas akan menerima hadiah, memaparkan kerja mereka di Expo, dan mesti memberikan hak cipta kepada Zindi. Tidak seperti persaingan lain yang ditinjau dalam seksyen ini, yang memberi tumpuan kepada memajukan state-of-the-art dalam pengesanan media sintetik atau dimanipulasi, persaingan ini memberi tumpuan kepada pengeluaran deepfake yang kelihatan lebih berpusat kepada kemanusiaan. Ini adalah trend yang diamati dalam seni [] ] and culture [ ].5.3 Generation and Detection of Manipulated Media 95 to create convincing deepfakes to highlight the power of this synthetic media, illustrating its creative potential for exploitation for both positive and negative outcomes and focusing debate about its ethical use / misuse in an African context 31 57 The DeepFake Game Competition (DFGC) adalah dalam edisi pertama, diselenggarakan pada 2021 International Joint Conference on Biometrics (IJCB 2021). Penyelenggaraannya terutamanya dari Institut Automation China Academy of Sciences (CASIA). Gagasan persaingan adalah untuk mempromosikan permainan lawan antara agensi mendorong kemajuan dalam kedua-dua penciptaan deepfake dan pengesanan. Untuk mencapai ini, protokol enam peringkat telah direka untuk meninggalkan tiga fasa penciptaan (C-fase) dan fasa pengesanan (D-fase), biasanya satu minggu terpisah; penghantaran ditutup pada April 2021. Kedua-dua fasa C dan D-fase dikaitkan dengan set data Celeb-DF (v2) [ ], containing 6,229 videos (590 real/original videos and 5,639 fake/manipulated videos), for training purposes. As such, submissions to a C-phase would consist of datasets extracted from Celeb-DF (v2) which included novel face-swap approaches to obtain evaluation results. Submissions to a D-phase would consist of detection models/codes to obtain evaluation results. The models submitted for a D-phase were evaluated against the datasets submitted for the previous C-phase [ Metrik yang digunakan untuk penilaian were: a detection score, used for evaluation of a D-phase, and a creation score, used for evaluation of a C-phase. The top three finalists for the detection phase employed CNN-based classifiers EfficientNet-B3, Efficientnet-B0 and EfficientNetV2. 96 40 52 97 The Detection Score ( ) metric captures the models’ ability to correctly classify fake images submitted to the previous C-phase against a set of real images in the CelebDF test dataset. It is calculated using Eq. ( ), where is the number of valid submissions of created synthesis test sets in the last C-phase. ds 21 NC yang The Creation Score ( ) metric used to evaluate creation models submitted to this challenge is calculated by Eq. ( ), where is the number of valid submissions of detection methods in the last D-phase, the noise score ( noise) penalises noisy images, the other three parts of the equation relate to the following : “ ”. cs 22 ND S 98 ID level similarity to the donor ID, image level similarity to the target frame, and the deception ability against detection models. ID level similarity is scored by a face recognition model using dot product of two ID features (fake face ID and donor ID). The image level similarity is scored by SSIM [Structural Similarity Index] to make sure the face-swapped image is similar to the corresponding target image in content and quality Peng et al. [ ] observed a commonality between the three winning teams for the creation task, i.e., the use of the FaceShifter [ ] framework for face swapping. They highlighted two overall reflections about the competition: (1) the limited diversity of the deepfake datasets submitted and the use of repetitive methods to generate them, and (2) the limited size of the Celeb-DF (v2) dataset itself flagging the need for a larger dataset for next year’s competition. The organisers of the competition also applied the top two detection models to unseen datasets (DFDC and FaceForensics++) and noticed that they do not generalise well. 52 37 6 A Meta-Review of Deepfake-Related Surveys This section presents a meta-review of 12 selected deepfake-related survey papers, including eight published in English [ , , , – daripada , ] and four published in Chinese [ , , , ]. It covers the following aspects in a systematic manner: definitions and scope, performance metrics, datasets, challenges/competitions/benchmarks, performance comparison, key challenges and recommendations. 16 45 46 64 66 71 73 7 38 41 59 The meta-review aims at drawing some high-level insights for monitoring future development of deepfake-related technologies and their applications. 6.1 Definitions and Scope As we discussed in Section , among researchers, practitioners and law makers there is no universally accepted definition of “deepfake” as a term. This is also reflected in how the authors of the 12 survey papers considered this aspect. Most authors talked about the history of deepfakes and pointed out that the term reflects the combination of “deep learning” and “fake”, but some used a broader definition, e.g., Lyu [ ] ditakrifkan deepfake sebagai “ ”. Some authors also referred to deepfake-related legislations, but none of them pointed out that the definitions in some such legislations are completely different from the more technical definitions involving the use of deep learning. No authors discussed the blurred boundary between deepfakes and non-deepfakes, although some surveys actually cover both, e.g., Tao et al. [ ] focused on speech forgery and did not explicitly highlight “deepfake”. 1.1 45 Video dan audio palsu berkualiti tinggi yang dihasilkan oleh algoritma AI 59 In terms of the scope, while some authors (correctly) considered all types of media that can be produced by deepfake-related techniques [ , , , [ ], sesetengah mempertimbangkan hanya skop sempit, contohnya, penulis [ , , , ] considered only videos, and only authors of [ , Fenomena lain yang kami amati adalah bahawa banyak penulis memberi tumpuan lebih kepada imej muka dan video, dan penulis tiga kaji selidik [...] , , ] even limited the definition of “deepfake” to such a narrow scope: 38 41 45 65 7 64 71 73 16 66 16 64 71 Deshmukh and Wankhade [ ] defined it as “ ”; 16 a technology which creates fake images or videos of targeted humans by swapping their faces [by] another character saying or doing things that are not absolutely done by them and humans start believing in such fake as it is not always recognisable with the everyday human eye Younus and Hasan [ ] considered deepfake as a technique allowing “ ”; and 71 any computer user to exchange the face of one person with another digitally in any video Tolosana et al. [ ] defined it as “ ”. 64 a deep learning based technique able to create fake videos by swapping the face of a person by the face of another person Such unnecessarily narrow definitions and scopes can lead to confusion and do not help exchanges between researchers and practitioners working on different types of deepfakes. We call on more researchers to accept a broader definition of “deepfake” so that highly realistic/natural media of any kind generated by a sophisticated automated method (often AI-based) is considered deepfake. Here, we provide two examples of such a broader definition: the image2image (or pixel2pixel) technique [ ] that allows the production of deepfake images and videos of any objects (e.g., the “horse2zebra” deepfake image shown in Figure ), and the the so-called “deepfake geography [ ]”, where AI-based techniques are used to generate realistic-looking satellite images. 80 7 77 Another important fact missed or not sufficiently discussed by authors of all the 12 surveys is that deepfake techniques can be used for positive applications, e.g., creative arts, entertainment and protecting online users’ privacy. We call for more researchers and practitioners to follow the proposal in the 2020 Tencent AI White Paper [ Oleh itu, kita boleh menggunakan kata-kata yang berbeza untuk jenis data yang berbeza yang dihasilkan menggunakan teknik "sintesis mendalam", contohnya, "art mendalam", "animasi mendalam", "musik mendalam", dan "deepfake". manakala penulis 12 makalah kaji selidik tidak mengenali aplikasi positif teknologi "deepfake", beberapa penyelidik lain melakukannya, contohnya, penyelenggara Voice Conversion Challenge 2020 yang berkata teknologi VC (untuk ucapan deepfake) “ ”. 60 99 is useful in many applications, such as customizing audio book and avatar voices, dubbing, movie industry, teleconferencing, singing voice modification, voice restoration after surgery, and cloning of voices of historical persons 6.2 Performance Metrics Surprisingly, none of the 12 surveys have covered performance metrics explicitly. Some directly used performance metrics to explain and compare performance of covered deepfake generation and detection methods. The most used performance metrics include accuracy, ERR, and AUC. This may be explained by the page constraints of such survey papers, which did not allow the authors to extend their coverage significantly to cover performance metrics systematically. The subjective quality of deepfakes is an area least covered by the surveys, which seems related to an unbalanced coverage on deepfake generation and deepfake detection in terms of performance evaluation and comparison (the former much less than the latter). 6.3 Pengumpulan data Many of the 12 survey papers list a number of deepfake-related datasets, but none of them have coverage as complete as ours shown in Section . For instance, none of the surveys have covered the Voice Conversion Challenge 2016/2018/2020 datasets and the ASVspoof 2019/2021 datasets are covered briefly only in two surveys [ , ]. In addition, more recent deepfake datasets especially those released in 2021 are also not covered by any of the surveys. We believe that our Section is the most comprehensive review of deepfake-related datasets so far. 4 38 59 4 Some survey papers include datasets that are likely deepfakes, e.g., Verdoliva [ ] covered many general fake image datasets where the manipulated images were not generated by deep learning or even AI-based methods, and some surveys (e.g., [ ]) mentioned ASVspoof 2015 datasets but we did not see the use of deep learning for generating data used in the dataset. 66 38 6.4 Challenges, Competitions and Benchmarks Many surveys cover deepfake-related challenges, competitions and benchmarks. The coverage is, however, mostly limited, and some challenges (e.g., the Voice Conversion Challenge 2016/2018/2020 and the two Chinese challenges we covered in Section ) are not covered by any of the surveys. The level of detail of challenges, competitions and benchmarks is also normally limited, compared with what we chose to include in Section . Similar to the datasets we covered in Section , we believe that our coverage of deepfake-related challenges, competitions and benchmarks in Section is also the most comprehensive so far. 5 5 4 5 6.5 Performance Comparison Most surveys have a good coverage of related methods for deepfake generation and detection, but only some explicitly covered performance comparison between different methods [ , , » » » 38 46 64 Among all the survey papers, Li et al. [ ] conducted the most comprehensive study on performance of different deepfake detection methods. In addition to showing the performance metrics of a number of deepfake detection methods in Table 3 of [ ], they also looked at general characteristics and issues of different types of deepfake detection methods, as shown in Table . Furthermore, they also looked at research on robustness of deepfake detection methods against adversarial samples, referring to some work that showed a lack of such robustness. 38 38 4 Due to quality issues of many deepfake-related datasets (discussed in Section ), kita perlu merawat apa-apa metrik prestasi dan perbandingan kaedah pengesanan yang berbeza dengan berhati-hati. Tanpa menguji semua kaedah pada set data deepfake yang cukup besar, pelbagai dan berkualiti tinggi, hasil perbandingan prestasi boleh menyesatkan. Ini menyoroti pentingnya mempunyai lebih banyak cabaran, persaingan dan benchmarks untuk menggalakkan perbandingan prestasi pada set data standard dan menggunakan metrik prestasi yang konsisten. 4.6 6.6 Challenges and Recommendations The authors of some surveys identified some key challenges and future research directions for the deepfake community. Tidak menghairankan, bagaimana membangunkan kaedah pengesanan deepfake yang lebih kukuh, boleh diperluas, boleh digeneralisasikan dan boleh dijelaskan adalah salah satu cabaran utama yang paling dibincangkan dan juga arah penyelidikan utama masa depan. , , , , , , , , ]. Considering the arms race between deepfake generation and detection, this research direction will likely remain the hottest topic in deepfake research. 7 16 38 41 45 59 65 66 71 A couple of surveys [ , ] mentioned fusion as a key future research direction, where “fusion” refers to combining different methods (e.g., combining multiple detectors of different types) and data sources (e.g., jointly considering audio-visual analysis) to achieve better performance for deepfake detection. Lyu [ ] mencadangkan bahawa, untuk pengesanan video deepfake, kita perlu mempertimbangkan pengesanan tahap video lebih, yang boleh dianggap penggabungan hasil pengesanan semua bingkai video. 38 66 45 The authors of three surveys, Lyu [ ] , Deshmukh and Wankhade [ ] and Younus and Hasan [ ], argued that better (higher-quality, more up-to-date, and more standard) deepfake datasets are needed to develop more effective deepfake detection methods. Lyu [ ] also suggested that we need to consider effects in training data and improve the evaluation of datasets. We agree with them on these points. 45 16 71 45 media sosial pencucian Tao et al. [ Ini merupakan cadangan yang sah kerana kaedah ringan akan membolehkan peranti pengkomputeran yang kurang kuat (contohnya, peranti IoT) untuk mendapat manfaat daripada teknologi tersebut. 59 Two Chinese surveys [ , ] juga menyebutkan keperluan untuk mempunyai undang-undang baru yang berkaitan dengan deepfake yang memerangi penggunaan jahat deepfakes dan keperluan untuk melatih pengguna akhir seperti wartawan. 38 41 Terdapat juga cadangan ad-hoc lain yang diberikan oleh penulis beberapa kajian. ] argued that deepfake detection should be considered a (more complicated) multi-class, multi-label and local detection problem. Tolosana et al. [ ] discussed specific research directions for different deep-fake generation methods (face synthesis, identity swap, attribute manipulation, and expression swap). Liang et al. [ [ ] dan Li et al. ] recommended more active defence mechanisms such as using digital watermarking and blockchain technologies to build trustworthy media frameworks against deepfakes. 45 64 41 38 7 Conclusion Pertumbuhan pesat dalam keupayaan untuk memanipulasi media atau mencipta media sintetik yang kelihatan realistis dan semulajadi membuka jalan bagi deepfakes. Pada mulanya, makalah ini mengambil pendekatan kritikal untuk melihat definisi yang berbeza daripada istilah "deepfake". Dalam hal ini, kami menunjukkan definisi yang berbeza yang bertentangan dan memanggil masyarakat yang lebih luas untuk mempertimbangkan bagaimana untuk mendefinisikan istilah baru yang mempunyai skop dan makna yang lebih konsisten. Sebagai contoh, menggantikan "deepfake" dengan "sintesis mendalam" boleh menjadi lebih inklusif dengan merangkumi aplikasi positif teknik deepfake, contohnya, dalam hiburan dan untuk tujuan simulasi. This paper provided a comprehensive overview of multiple aspects of the deepfake ecosystem drawing from the research literature and other online sources published in two languages: English and Chinese. It covers commonly used performance metrics and standards, related datasets, challenges, competitions and benchmarks. It also presents a meta-review of 12 selected deepfake-related survey papers published in 2020 and 2021, covering not only the above mentioned aspects, but also highlighting key challenges and recommendations. References [1] Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. 2018. MesoNet: A Compact Facial Video Forgery Detection Network. In . IEEE, 1–7. Proceedings of the 2018 IEEE International Workshop on Information Forensics and Security https://doi.org/10.1109/WIFS.2018.8630 761 [2] Henry Ajder, Giorgio Patrini, Francesco Cavalli, and Laurence Cullen. 2019. The State of Deepfakes: Landscape, Threats, and Impact. Deeptrace. , 27 pages. https://sensity.ai/reports/ [3] Zahid Akhtar and Tiago H. Falk. 2017. Audio-Visual Multimedia Quality Assessment: A Comprehensive Survey. 5 (2017), 21090–21117. IEEE Access https://doi.org/10.1109/ACCESS.2017.2750918 [4] Ahmed Ali dan Steve Renals. 2018. Perkiraan Kadar Kesilapan Kata untuk Pengiktirafan Ucapan: e-WER. . Association for Computational Linguistics, 20–24. Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics https://doi.org/10.18653/v1/P18-2004 [5] Sercan O¨ . Arık, Jitong Chen, Kainan Peng, Wei Ping, and Yanqi Zhou. 2018. Neural Voice Cloning with a Few Samples. In . Curran Associates Inc., 10040–10050. Proceedings of the 32nd International Conference on Neural Information Processing Systems https://papers.nips.cc/paper/201 8/hash/4559912e7a94a9c32b09d894f2bc3c82-Abstract.html [6] Aayush Bansal, Shugao Ma, Deva Ramanan, dan Yaser Sheikh. 2018. Recycle-GAN: Retargeting Video Tanpa Pengawasan. dalam . Springer, 17 pages. Proceedings of the 2018 European Conference on Computer Vision https://doi.org/10.1007/978-3-030-01228-1 8 [7] Yu-xuan Bao, Tian-liang Lu, and Yan-hui Du. 2020. Overview of Deepfake Video Detection Technology. 47, 9 (2020), 283–292. Computer Science https://doi.org/10.11896/jsjkx.200400130 [8] Mikol-aj Bin´kowski, Jeff Donahue, Sander Dieleman, Aidan Clark, Erich Elsen, Norman Casagrande, Luis C. Cobo, and Karen Simonyan. 2019. High Fidelity Speech Synthesis with Adversarial Networks. https://doi.org/10.48550/ARXIV.1909.11646 [9] ![](file:///C:/Users/user/AppData/Local/Temp/msohtmlclip1/01/clip_image050.gif)Madeline Brady. 2020. Deepfakes: A New Desinformation Threat? Report by the Democracy Reporting International. , 9 pages. https://democracy-reporting.org/dri publications/de epfakes-a-new-disinformation-threat/ [10] Umur Aybars Ciftci, Ilke Demir, dan Lijun Yin. 2020. FakeCatcher: Pengesanan video potret sintetik menggunakan isyarat biologi. (2020), 17 buah halaman IEEE Transactions on Pattern Analysis and Machine Intelligence https://doi.org/10.1109/TPAMI.2020.3009287 [11] Hao Dang, Feng Liu, Joel Stehouwer, Xiaoming Liu, and Anil K. Jain. 2020. On the Detection of Digital Face Manipulation. In . IEEE, 10 pages. Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition https://doi.org/10.1109/CVPR42600.2020.00582 [12] Rohan Kumar Das, Tomi Kinnunen, Wen-Chin Huang, Zhen-Hua Ling, Junichi Yamagishi, Zhao Yi, Xiaohai Tian, dan Tomoki Toda. . International Speech Communication Association, 99–120. Proceedings of the Joint Workshop for the Blizzard Challenge and Voice Conversion Challenge 2020 https://doi.org/10.21437/VCC BC.2020-15 [13] H´ector Delgado, Nicholas Evans, Tomi Kinnunen, Kong Aik Lee, Xuechen Liu, Andreas Nautsch, Jose Patino, Md Sahidullah, Massimiliano Todisco, Xin Wang, and Junichi Yamagishi. 2021. ASVspoof 2021: Automatic Speaker Verification Spoofing and Countermeasures Challenge Evaluation Plan. https://www.asvspoof.org/asvspoof2021/asvspoof2021 evaluation plan.pdf [14] H ́ector Delgado, Nicholas Evans, Tomi Kinnunen, Kong Aik Lee, Xuechen Liu, Andreas Nautsch, Jose Patino, Md Sahidullah, Massimiliano Todisco, Xin Wang, dan Junichi Yamagishi. ASVspoof 2021 Challenge - Logical Access Database. https://doi.org/10.5281/zenodo.4837263 [15] H´ector Delgado, Nicholas Evans, Tomi Kinnunen, Kong Aik Lee, Xuechen Liu, Andreas Nautsch, Jose Patino, Md Sahidullah, Massimiliano Todisco, Xin Wang, and Junichi Yamagishi. 2021. ASVspoof 2021 Challenge - Speech Deepfake Database. https://doi.org/10.5281/zenodo.4835108 [16] ![](file:///C:/Users/user/AppData/Local/Temp/msohtmlclip1/01/clip_image051.gif)Anushree Deshmukh and Sunil B. Wankhade. 2021. Deepfake Detection Approaches Using Deep Learning: A Systematic Review. In . Springer, 293–302. Intelligent Computing and Networking: Proceedings of IC-ICN 2020 (Lecture Notes in Networks and Systems, Vol. 146) https://doi.org/ 10.1007/978-981-15-7421-4 27 [17] Xinyi Ding, Zohreh Raziei, Eric C. Larson, Eli V. Olinick, Paul Krueger, and Michael Hahsler. 2020. Swapped Face Detection using Deep Learning and Subjective Assessment. Tahun 2020, 1 (2020), 1–12. EURASIP Journal on Information Security https://doi.org/10.1186/s13635-020-00109-8 [18] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. 2020. The DeepFake Detection Challenge (DFDC) Dataset. https://doi.org/10.48550/ARXIV.2006.07397 [19] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. 2020. The DeepFake Detection Challenge (DFDC) Dataset. arXiv:2006.07397. https://arxiv.org/abs/2006.07397 [20] Nick Dufour and Andrew Gully. 2019. Contributing Data to Deepfake Detection Research. Google AI Blog. https://ai.googleblog.com/2019/09/contributing-data-to-deepfake-detectio n.html [21] Ricard Durall, Margret Keuper, Franz-Josef Pfreundt, dan Janis Keuper. 2019. Mendedahkan Deep-Fakes dengan Ciri-ciri Sederhana. https://doi.org/10.48550/ARXIV.1911.00686 [22] Cristian Canton Ferrer, Brian Dolhansky, Ben Pflaum, Joanna Bitton, Jacqueline Pan, and Jikuo Lu. 2020. Deepfake Detection Challenge Results: An Open Initiative to Advance AI. Meta AI Blog. Melayu https://ai.facebook.com/blog/deepfake-detection-challenge-results-an-open-initia tive-to-advance-ai/ [23] Gereon Fox, Wentao Liu, Hyeongwoo Kim, Hans-Peter Seidel, Mohamed Elgharib, and Christian Theobalt. 2021. Videoforensicshq: Detecting High-Quality Manipulated Face Videos. In . IEEE, 1–6. Proceedings of the 2021 IEEE International Conference on Multimedia and Expo https://doi.or g/10.1109/ICME51207.2021.9428101 [24] Haiying Guan, Andrew Delgado, Yooyoung Lee, Amy N. Yates, Daniel Zhou, Timothee Kheyrkhah, and Jon Fiscus. 2021. User Guide for NIST Media Forensic Challenge (MFC) Datasets. https://doi.org/10.6028/NIST.IR.8377 [25] Yinan He, Bei Gan, Siyu Chen, Yichun Zhou, Guojun Yin, Luchuan Song, Lu Sheng, Jing Shao, and Ziwei Liu. 2021. ForgeryNet: A Versatile Benchmark for Comprehensive Forgery Analysis. In IEEE, 4360 – 4369 Proceedings of the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition https://doi.org/10.1109/CVPR46437.2021.00434 [26] Liming Jiang, Zhengkui Guo, Wayne Wu, Zhaoyang Liu, Ziwei Liu, Chen Change Loy, Shuo Yang, Yuanjun Xiong, Wei Xia, Baoying Chen, Peiyu Zhuang, Sili Li, Shen Chen, Taiping Yao, Shouhong Ding, Jilin Li, Feiyue Huang, Liujuan Cao, Rongrong Ji, Changlei Lu, and Ganchao Tan. 2021. DeeperForensics Challenge 2020 on Real-World Face Forgery Detection: Methods and Results. arXiv:2102.09471. https://arxiv.org/pdf/2102.09471.pdf [27] Liming Jiang, Ren Li, Wayne Wu, Chen Qian, and Chen Change Loy. 2020. DeeperForensics-1.0: A Large-Scale Dataset for Real-World Face Forgery Detection. In . IEEE, 2886–2895. 2020 IEEE/CVF Conference on Computer Vision dan Pengiktirafan Pattern https://doi.org/ 10.1109/CVPR42600.2020.00296 [28] Nal Kalchbrenner, Erich Elsen, Karen Simonyan, Seb Noury, Norman Casagrande, Edward Lockhart, Florian Stimberg, Aaron van den Oord, Sander Dieleman, and Koray Kavukcuoglu. 2018. Efficient Neural Audio Synthesis. https://doi.org/10.48550/ARXIV.1802.08435 [29] Takuhiro Kaneko dan Hirokazu Kameoka. 2017. penukaran suara paralel tanpa data menggunakan rangkaian lawan yang konsisten dalam kitaran. https://doi.org/10.48550/ARXIV.1711.11293 [30] Tero Karras, Samuli Laine, and Timo Aila. 2019. A Style-based Generator Architecture for Generative Adversarial Networks. In . IEEE, 4401–4410. Proceedings of the 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition https://doi.org/10.1109/CVPR.2019.00453 [31] Ross Kelly. 2021. Digital Art Exhibition to Showcase Creative Potential of AI. DIGIT News. https://www.digit.fyi/digital-art-exhibition-to-showcase-creative-potential-of-ai/ [32] Ali Khodabakhsh, Raghavendra Ramachandra, Kiran Raja, Pankaj Wasnik, and Christoph Busch. 2018. Fake Face Detection Methods: Can They Be Generalized?. In . IEEE, 1–6. Proceedings of the 2018 International Conference of the Biometrics Special Interest Group https://doi.org/10.2 3919/BIOSIG.2018.8553251 [33] Hyeongwoo Kim, Mohamed Elgharib, Hans-Peter Zoll¨ofer, Michael Seidel, Thabo Beeler, Christian Richardt, and Christian Theobalt. 2019. Neural Style-Preserving Visual Dubbing. 38, 6, Article 178 (2019), 13 pages. ACM Transactions on Graphics https://doi.org/10.1145/3355089.3356500 [34] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick P ́erez, Christian Richardt, Michael Zollh ̈ofer, dan Christian Theobalt. 2018. 37, 4, Perkara 163 (2018), 14 halaman. Transaksi ACM pada Grafis https://doi.org/ 10.1145/3197517.3201283 [35] Pavel Korshunov and S´ebastien Marcel. 2019. Vulnerability Assessment and Detection of Deepfake Videos. In . IEEE, 1–6. Persidangan Antarabangsa Biometrik 2019 https://doi.org/10.1109/ICB45273.2019.8987375 [36] Patrick Kwon, Jaeseong You, Gyuhyeon Nam, Sungwoo Park, dan Gyeongsu Chae. 2021. KoDF: A Large-scale Korean DeepFake Detection Dataset IEEE, 10724–10733. Konvensyen Antarabangsa 2021 IEEE/CVF mengenai Penglihatan Komputer https://doi.org/10.1109/ICCV 48922.2021.01057 [37] Lingzhi Li, Jianmin Bao, Hao Yang, Dong Chen, and Fang Wen. 2020. FaceShifter: Towards High Fidelity And Occlusion Aware Face Swapping. arXiv:1912.13457. https://arxiv.org/abs/1912.13457 [38] Xurong Li, Shouling Ji, Chunming Wu, Zhenguang Liu, Shuiguang Deng, Peng Cheng, Min Yang, and Xiangwei Kong. 2021. Survey on Deepfakes and Detection Techniques. 32, 2 (2021), 496–518. Journal of Software http://www.jos.org.cn/1000-9825/6140.htm [39] Yuezun Li, Ming-Ching Chang, and Siwei Lyu. 2018. In Ictu Oculi: Exposing AI Created Fake Videos by Detecting Eye Blinking. In . IEEE, 1–7. Laporan Workshop Antarabangsa IEEE 2018 mengenai Forensik Maklumat dan Keselamatan https://doi.org/10.1109/WIFS.2018.8630787 [40] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. 2020. Celeb-DF: A Large-Scale Challenging Dataset for DeepFake Forensics. In IEEE 3204 – 3213. Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition https://doi.org/10.1109/CVPR42600. 2020.00327 [41] Ruigang Liang, Peizhuo Lv, Yue Zhao, Peng Chen, Hao Xing, Yingjun Zhang, Jizhong Han, Ran He, Xianfeng Zhao, Ming Li, and Kai Chen. 2020. A Survey of Audiovisual Deepfake Detection Techniques. 5, 2 (2020), 1 hingga 17. Journal of Cyber Security http://jcs.iie.ac.cn/xxaqxb/ch/re ader/view abstract.aspx?file no=20200202&flag=1 [42] Steven R. Livingstone and Frank A. Russo. 2018. The Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS): A Dynamic, Multimodal Set of Facial and Vocal Expressions in North American English. 13, 5 (2018), 35 pages. PloS one [43] Chen-Chou Lo, Szu-Wei Fu, Wen-Chin Huang, Xin Wang, Junichi Yamagishi, Yu Tsao, dan Hsin-Min Wang. 2021. MOSNet: Penilaian Objektif Berdasarkan Pembelajaran Mendalam untuk Penukaran Suara. arXiv:1904.08352. https://arxiv.org/pdf/1904.08352.pdf [44] Jaime Lorenzo-Trueba, Junichi Yamagishi, Tomoki Toda, Daisuke Saito, Fernando Villavicencio, Tomi Kinnunen, and Zhenhua Ling. 2018. The Voice Conversion Challenge 2018: Promoting Development of Parallel and Nonparallel Methods. In . International Speech Communication Association, 195–202. Proceedings of the Odyssey 2018 The Speaker and Language Recognition Workshop https://doi.org/10.21437/Odyssey.2018-28 [45] Siwei Lyu. 2020. pengesanan deepfake: cabaran semasa dan langkah seterusnya. dalam . IEEE, 6 pages. Proceedings of the 2020 IEEE International Conference on Multimedia Expo Workshops https://doi.org/10.1109/ICMEW46912.2020.9105991 [46] Yisroel Mirsky and Wenke Lee. 2021. The Creation and Detection of Deepfakes: A Survey. 54, 1, Article 7 (2021), 41 pages. ACM Computing Survey https://doi.org/10.1145/3425780 [47] Gautham J. Mysore. 2015. Can we Automatically Transform Speech Recorded on Common Consumer Devices in Real-World Environments into Professional Production Quality Speech?—A Dataset, Insights, and Challenges. 22, 8 (2015), 1006–1010. Surat pemprosesan isyarat IEEE https://doi.org/10.1109/LSP.2014.2379648 [48] Jo ̃ao C. Neves, Ruben Tolosana, Ruben Vera-Rodriguez, Vasco Lopes, Hugo Proen ̧ca, dan Julian Fierrez. 2020. GANprintR: Peningkatan Fake dan Penilaian Keadaan Seni dalam Pengesanan Manipulasi Wajah. 14, 5 (2020), 1038–1048 IEEE Journal of Selected Topics in Signal Processing https://doi.org/10.1109/JSTSP.2020.3007250 [49] NIST Media Forensics Challenge Team. 2021. Open Media Forensics Challenge 2020-2021 Rancangan Penilaian. https://mig.nist.gov/MFC/Web/EvalPlan2020/OpenMFC2020EvaluationPlan.pdf [50] Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. 2016. WaveNet: A Generative Model for Raw Audio. https://doi.org/10.48550/ARXIV.1609.03499 [51] Debajyoti Pal and Tuul Triyason. 2018. A Survey of Standardized Approaches towards the Quality of Experience Evaluation for Video Services: An ITU Perspective. 2018, Article 1391724 (2018), 25 pages. Jurnal Antarabangsa Penyiaran Multimedia Digital https://doi.org/10.1155/20 18/1391724 [52] Bo Peng, Hongxing Fan, Wei Wang, Jing Dong, Yuezun Li, Siwei Lyu, Qi Li, Zhenan Sun, Han Chen, Baoying Chen, Yanjie Hu, Shenghai Luo, Junrui Huang, Yutong Yao, Boyuan Liu, Hefei Ling, Guosheng Zhang, Zhiliang Xu, Changtao Miao, Changlei Lu, Shan He, Xiaoyan Wu, dan Wanyi Zhuang. https: //arxiv.org/abs/2106.01217 [53] Ivan Perov, Daiheng Gao, Nikolay Chervoniy, Kunlin Liu, Sugasa Marangonda, Chris Um´e, Mr. Dpfks, Carl Shift Facenheim, Luis RP, Jian Jiang, Sheng Zhang, Pingyu Wu, Bo Zhou, and Weiming Zhang. 2020. DeepFaceLab: Integrated, Flexible and Extensible Face-swapping Framework. https://doi.org/10.48550/ARXIV.2005.05535 [54] Stanislav Pidhorskyi, Donald A. Adjeroh, dan Gianfranco Doretto. 2020. Latent Adversarial Au-to-coders. dalam . IEEE, 10 pages. Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition https://doi.org/10.1109/CVPR42600.2020.01411 [55] Andreas R ̈ossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, dan Matthias Nießner. 2018. FaceForensics: Set data video skala besar untuk pengesanan penipuan di muka manusia. https://doi.org/10.48550/ARXIV.1803.09179 [56] Andreas R ̈ossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, dan Matthias Nießner. 2019. FaceForensics++: Belajar untuk Mendeteksi Imej Memanipulasi Wajah. dalam . IEEE, 1–11. Proceedings of the 2019 International Conference on Computer Vision https://doi.org/10.110 9 / ICCV.2019.00009 [57] Anindya Sen. 2021. seni dan kecerdasan buatan: Bagaimana 'deepfakes' boleh membantu mewujudkan pengalaman muzium yang autentik! https://medium.com/art-world-zen/art-and-artificia l-intelligence-how-deepfakes-can-help-create-authentic-museum-experiences-65d8aa 7d9c [58] Kou Tanaka, Hirokazu Kameoka, Takuhiro Kaneko, and Nobukatsu Hojo. 2019. WaveCycleGAN2: Time-domain Neural Post-filter for Speech Waveform Generation. https://doi.org/10.48550/A RXIV.1904.02892 [59] Jianhua Tao, Ruibo Fu, Jiangyan Yi, Chenglong Wang, dan Tao Wang. 5, 2 (2020), 28–38. Jurnal Keselamatan Siber http://jcs.iie.ac.cn/xxaqxb/ch/reader/view abstract.aspx?file no=20200204&flag=1 [60] Tencent. 2020 Kecerdasan Buatan Buku Putih. https://tech.sina.com.cn/roll/2020-07-14/doc-iivhvpwx5201226.shtml [61] Justus Thies, Michael Zollh, dan Matthias Niessner. 2019. Rendering Neural ditangguhkan: Sintesis imej menggunakan Tekstur Neural. 38 Perkara 66 (2019), 12 halaman. ACM Transactions on Graphics https://doi.org/10.1145/3306346.3323035 [62] Tomoki Toda, Ling-Hui Chen, Daisuke Saito, Fernando Villavicencio, Mirjam Wester, Zhizheng Wu, and Junichi Yamagishi. 2016. The Voice Conversion Challenge 2016. In . International Speech Communication Association, 1632–1636. Proceedings of Interspeech 2016 https://doi.org/10.21437/Interspeech.2016-1066 [63] Massimiliano Todisco, Xin Wang, Ville Vestman, Md Sahidullah, Hector Delgado, Andreas Nautsch, Junichi Yamagishi, Nicholas Evans, Tomi Kinnunen, dan Kong Aik Lee. 2019. ASVspoof 2019: Horizons Masa Depan dalam Pengesanan Audio yang Dipalsukan dan Palsu. arXiv:1904.05441. https://arxiv.org/ pdf/1904.05441.pdf [64] Ruben Tolosana, Ruben Vera-Rodriguez, Julian Fierrez, Aythami Morales, and Javier Ortega-Garcia. 2020. Deepfakes and beyond: A Survey of face manipulation and fake detection. 64 (2020), 131 hingga 148. Pengumpulan maklumat Fusion https://doi.org/10.1016/j.inffus.2020.06.014 [65] Xin Tong, Luona Wang, Xiaoqin Pan, dan Jingya Wang. 2020. gambaran keseluruhan Deepfake: Pedang Damocles dalam AI. IEEE, 265 – 273 Persidangan Antarabangsa 2020 mengenai Penglihatan Komputer, Imej dan Pembelajaran Mendalam https://doi.org/10.1109/CVIDL51233.2020.00-88 [66] Luisa Verdoliva. 2020. Media Forensics and DeepFakes: An Overview. 14, 5 (2020), 910–932. IEEE Journal of Selected Topics in Signal Processing https://doi.org/10.1109/JSTSP.2020.3002101 [67] Xin Wang, Junichi Yamagishi, Massimiliano Todisco, H´ector Delgado, Andreas Nautsch, Nicholas Evans, Md Sahidullah, Ville Vestman, Tomi Kinnunen, Kong Aik Lee, Lauri Juvela, Paavo Alku, Yu-Huai Peng, Hsin-Te Hwang, Yu Tsao, Hsin-Min Wang, S´ebastien Le Maguer, Markus Becker, Fergus Henderson, Rob Clark, Yu Zhang, Quan Wang, Ye Jia, Kai Onuma, Koji Mushika, Takashi Kaneda, Yuan Jiang, Li-Juan Liu, Yi-Chiao Wu, Wen-Chin Huang, Tomoki Toda, Kou Tanaka, Hirokazu Kameoka, Ingmar Steiner, Driss Matrouf, Jean-Fran¸cois Bonastre, Avashna Govender, Srikanth Ronanki, Jing-Xuan Zhang, and Zhen-Hua Ling. 2020. ASVspoof 2019: A Large-scale Public Database of Synthesized, Converted and Replayed Speech. 64 (2020), 27 halaman Computer Speech & Language https://doi.org/10.1016/j.csl.2020.101114 [68] Mirjam Wester, Zhizheng Wu, and Junichi Yamagishi. 2016. Analysis of the Voice Conversion Challenge 2016 Evaluation Results. In Persatuan Komunikasi Ucapan Antarabangsa, 1637–1641. Prosedur Persidangan Interspeech 2016 https://doi.org/10.21437/Interspeech.201 6 daripada 1331 [69] Junichi Yamagishi, Massimiliano Todisco, Md Sahidullah, H´ector Delgado, Xin Wang, Nicholas Evans, Tomi Kinnunen, Kong Aik Lee, Ville Vestman, and Andreas Nautsch. 2019. ASVspoof 2019: Automatic Speaker Verification Spoofing and Countermeasures Challenge Evaluation Plan. https://www.asvspoof.org/asvspoof2019/asvspoof2019 evaluation plan.pdf [70] Zhao Yi, Wen-Chin Huang, Xiaohai Tian, Junichi Yamagishi, Rohan Kumar Das, Tomi Kinnunen, Zhen-Hua Ling, and Tomoki Toda. 2020. Voice Conversion Challenge 2020 – Intra-lingual Semi-parallel and Cross-lingual Voice Conversion –. In Persatuan Komunikasi Bercakap Antarabangsa, 80–98. Proceedings of the Joint Workshop for the Blizzard Challenge and Voice Conversion Challenge 2020 https://doi.org/10.21437/VCC BC.2020-14 [71] Mohammed A. Younus dan Taha M. Hasan. 2020. Pandangan singkat teknik pengesanan video Deepfake. dalam . IEEE, 115–120. Proceedings of the 2020 6th International Engineering Conference https://doi.org/10.1109/IEC49899.2020.9122916 [72] Guangtao Zhai and Xiongkuo Min. 2020. Perceptual Image Quality Assessment: A Survey. 63, Perkara 211301 (2020), 52 halaman. China Sains Maklumat https://doi.org/10.1007/s11432-019-2757-1 [73] Teng Zhang, Lirui Deng, Liang Zhang, dan Xianglei Dang. 2020. Pembelajaran mendalam dalam sintesis muka: Survey pada Deepfakes. dalam IEEE, 67 hingga 70 Proceedings of the 2020 IEEE 3rd International Conference on Computer and Communication Engineering Technology https://doi.org/10.1109/CCET5090 1.2020.9213159 daripada [74] Yuxuan Zhang, Huan Ling, Jun Gao, Kangxue Yin, Jean-Francois Lafleche, Adela Barriuso, Antonio Torralba, dan Sanja Fidler. 2021. DatasetGAN: Pabrik Data Label yang berkesan dengan usaha manusia yang minimum. IEEE, 10140 hingga 10150 Konvensyen IEEE/CVF 2021 mengenai Pengiktirafan Penglihatan Komputer dan Pattern https://doi.org/10.1109/CVPR46437.2021.01001 [75] ![](file:///C:/Users/user/AppData/Local/Temp/msohtmlclip1/01/clip_image051.gif)Yuanhan Zhang, ZhenFei Yin, Yidong Li, Guojun Yin, Junjie Yan, Jing Shao, dan Ziwei Liu. . Springer, 70–85. Konferensi Eropah Penglihatan Komputer 2020 https://doi.org/10.1007/97 8-3-030-58610-2 5 [76] Yuanhan Zhang, Zhenfei Yin, Jing Shao, Ziwei Liu, Shuo Yang, Yuanjun Xiong, Wei Xia, Yan Xu, Man Luo, Jian Liu, Jianshu Li, Zhijun Chen, Mingyu Guo, Hui Li, Junfu Liu, Pengfei Gao, Tianqi Hong, Hao Han, Shijie Liu, Xinhua Chen, Di Qiu, Cheng Zhen, Dashuang Liang, Yufeng Jin, dan Zhanlong Hao. https://arxiv.org/pdf/2102.12642.pdf [77] Bo Zhao, Shaozeng Zhang, Chunxue Xu, Yifan Sun, dan Chengbin Deng. 2021. Deep Fake Geography? apabila data geospatial bertemu kecerdasan buatan. 48, 4 (2021), 338–352. Cartography and Geographic Information Science https://doi.org/10.1080/15230406.2021.1910075 [78] Peng Zhou, Xintong Han, Vlad I. Morariu, dan Larry S. Davis. 2017. Rangkaian saraf dua aliran untuk pengesanan muka tampered. dalam IEEE pada tahun 1831-1839. Persidangan 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshop https://doi.org/10.1109/CVPRW.2017.229 [79] Tianfei Zhou, Wenguan Wang, Zhiyuan Liang, dan Jianbing Shen. 2021. IEEE, 5774 sehingga 5784 Konvensyen IEEE/CVF 2021 mengenai Pengiktirafan Penglihatan Komputer dan Pattern https://doi.org/10.1109/CVPR46437.2021.00572 [80] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. 2017. Unpaired Image-to-Image Translation Using Cycle-Consistent Adversarial Networks. In IEEE, 2242 sehingga 2251 Konferensi Antarabangsa IEEE 2017 mengenai Penglihatan Komputer https://doi.org/10.1109/ICCV.2 017.244 [81] Bojia Zi, Minghao Chang, Jingjing Chen, Xingjun Ma, dan Yu-Gang Jiang. 2020. WildDeepfake: Set data dunia sebenar yang mencabar untuk pengesanan Deepfake. ACM, 2382 – 2390 Prosedur Konferensi Antarabangsa ACM Multimedia ke-28 2020 https://doi.org/10.1145/339417 1.3413769 Dokumen ini boleh didapati di archiv di bawah lesen CC by 4.0 Deed (Attribution 4.0 International). This paper is di bawah lesen CC by 4.0 Deed (Attribution 4.0 International). Tersedia dalam Archive