Founded in 2021, Atlas Metrics makes the impact of companies on people and the environment fully transparent. Our mission is to build the tools that make it easy for companies to measure and share the impact of their activities. We’re here to support an increasingly active community of companies, investors, regulators, consumers, and non-profits in their efforts to bring about a more sustainable future.

Atlas Metrics

Authors:
(1) Maria Rigaki, Faculty of Electrical Engineering, Czech Technical University in Prague, Czech Republic and maria.rigaki@fel.cvut.cz;
(2) Sebastian Garcia, Faculty of Electrical Engineering, Czech Technical University in Prague, Czech Republic and sebastian.garcia@agents.fel.cvut.cz. Table of Links Abstract & Introduction Threat Model Background and Related Work Methodology Experiments Setup Results Discussion Conclusion, Acknowledgments, and References Appendix 4 Methodology The MEME algorithm combines model extraction and reinforcement learning (RL) techniques to improve the generation of evasive malware binaries. The algorithm takes advantage of the fact that the attacker fully controls the binary modifier part of the environment but does not control the target model, which is a black box. However, an attacker can use data collected during the training of the RL agent to create and train a surrogate model of the target. The algorithm is implemented as several training/testing rounds until a final reinforcement learning policy is learned that can produce adversarial binaries for a given target. Figure 2 presents a high-level description of the algorithm, and the detailed listing is presented in Algorithm 1. Following Algorithm 1, MEME initializes a Proximal Policy Optimization (PPO) RL algorithm [35 (lines 1 and 2) to perform a first train of the PPO policy using the target model and a limited amount of steps (using MalwareGym). The goal of this step is to store a small number of observations and labels in Dsur (line 3). Then, Dsur is combined with an auxiliary set of labeled data (Daux) that the attacker should possess, and this combined dataset is used to train a surrogate model for the model extraction attack (line 5). The improved surrogate replaces the target model in a new Malware-Gym environment that trains a better policy (agent) for evasion (line 6). Last, the improved policy is evaluated against the original target model (line 7) to obtain the final evasion metrics. In this last step, we take advantage of the queries done to the target to append their output to the Dsur dataset. This loop is repeated for k rounds. During the last round, the evaluation is performed using the malware binary test set to produce the final evasion metrics (instead of the malware binary evaluation set used in the inner loop). The total number of queries to the target during training is k ∗ n, where k is the total number of rounds. For learning the policy, we use the PPO algorithm, which is a model-free onpolicy gradient method that clips the objective function, limiting the updates done to the policy to avoid too large positive changes or a minimum of negative changes in order to stabilize it. We used the Stable Baselines3 [31] implementation of PPO that uses an Actor-Critic (A2C) architecture. To train the surrogate model, it is not enough to use the data collected during the learning or evaluation of the agent in Dsur. The number of samples is not enough since we aim to minimize the queries to the target. In addition, all observations come from the extracted features of the malicious binary dataset or their adversarial modifications. Therefore using an auxiliary dataset is necessary to learn a good surrogate. This dataset is assumed to be from a similar distribution as the training data distribution of the target, however, this is not easy to achieve when the target is an AV or an unknown system. Adaptations to the Gym Environment MEME was implemented around the Malware-Gym environment (Section 3.1) in its latest version1. Several modifications were applied to the environment to fit the assumptions and constraints for this work: – The ”modify machine type” action was removed because our tests showed that it produces invalid binaries for Windows 10 systems. – All targets were set to return hard labels (0 or 1), not scores. – Regarding the benign sections, the Malware-Gym implementation used only data from ”.text” sections. In our implementation, we use data from other sections if the ”.text” section is unavailable. – The latest environment supports the Ember and Sorel-LGB classifiers as targets. We added new environments to support the Sorel-FFNN, surrogate, and AV targets. The AV requires a web service in a virtual machine that invokes the AV static scanning capabilities. – For all target environments, we added support for saving the observations (features) and scores during training and evaluation runs so that they can be used for the training of the surrogate. Our version of the environment and the experiments performed as part of this work can be found in (https://github.com/stratosphereips/meme malware rl/releases/tag/v1.0). Please note that we are releasing the source code of our implementation for reproducibility and improvement; however, we do not release the trained models or agents to avoid potential misuse. 4.1 Evasion Evaluation To evaluate the performance of the evasion models reasonably and realistically, we limited how many queries they do to the target model and how long they run. The idea behind liming running time is that the ratio at which malware authors create new malware is high, and a method that takes too long to create evasive malware is impractical. Industry measurements such as the ones provided by AV-ATLAS [18] report that 180 new malware are generated per minute. Virus Total [41] provides an even higher measurement of 560 distinct new files uploaded per minute as of the 21st of May, 2023. These global measurements vary but show how quickly new malware variants appear, possibly due to the proliferation of Malware-as-a-Service frameworks. A recent and more conservative measurement from Blackberry mentions that their customers get attacked by 1.5 new malware per minute [39]. Given the above measurements, we decided to constrain the running time of each experiment to four hours in total. Given that the test set consists of 300 malware binaries, this corresponds to 1.25 processed binaries per minute which is lower than the most conservative reported metric we could find. The primary metric used to evaluate the malware evasion task was the evasion rate E, which is the fraction of malware that becomes evasive within the time window of four hours over the total number of malware that were initially detected by each target: E = nev/ndet . We also report the average number of binary modifications required for a malware binary to evade the target. For the Random, PPO, and MEME methods, this is equivalent to the mean episode length over all the detected malware binaries in the test set. The episode length for a non-evasive binary is equal to the maximum number of attempts, and for an evasive one is the number of changes required to become evasive. For the MAB framework, for the evasive binaries, we considered only the minimal binaries with the least amount of actions. The average number of binary modifications was impossible to measure for the GAMMA attack since the SecML framework (through which GAMMA was implemented) does not provide a straightforward way to measure this metric. 4.2 Surrogate Evaluation The surrogate models trained in MEME were evaluated using two different metrics: label agreement and explainability-based feature agreement [37]. The label agreement of two models f and ˆf, respectively, is defined as the average number of similar predictions over a test set Xtest, and it is a standard metric for model extraction fidelity attacks [19]: The feature agreement metric computes the fraction of common features between the sets of top-k features of two explanations. Given two explanations Et (target) and Es (surrogate), the feature agreement metric can be formally defined as: where top features(E, k) returns the set of top-k features of the explanation E based on the magnitude of the feature importance values. The maximum value of the feature agreement is 1. For this work, we measured the feature agreement for k = 10 and k = 20, and the explainability method used was SHAP (SHapley Additive exPlanations) [25]. SHAP employs game theory principles to explain machine learning model outputs by linking effective credit allocation with localized explanations via Shapley values that originated from game theory. SHAP is model-agnostic and has been used effectively in the past for adversarial malware creation [33]. For the LightGBM targets we used TreeSHAP which is designed specifically for tree models, while for Sorel FFNN we used the KernelSHAP variant. This paper is available on arxiv under CC BY-NC-SA 4.0 DEED license. Authors: (1) Maria Rigaki, Faculty of Electrical Engineering, Czech Technical University in Prague, Czech Republic and maria.rigaki@fel.cvut.cz; (2) Sebastian Garcia, Faculty of Electrical Engineering, Czech Technical University in Prague, Czech Republic and sebastian.garcia@agents.fel.cvut.cz. Authors: Authors: (1) Maria Rigaki, Faculty of Electrical Engineering, Czech Technical University in Prague, Czech Republic and maria.rigaki@fel.cvut.cz; (2) Sebastian Garcia, Faculty of Electrical Engineering, Czech Technical University in Prague, Czech Republic and sebastian.garcia@agents.fel.cvut.cz. Table of Links Abstract & Introduction Abstract & Introduction Threat Model Threat Model Background and Related Work Background and Related Work Methodology Methodology Experiments Setup Experiments Setup Results Results Discussion Discussion Conclusion, Acknowledgments, and References Conclusion, Acknowledgments, and References Appendix Appendix 4 Methodology The MEME algorithm combines model extraction and reinforcement learning (RL) techniques to improve the generation of evasive malware binaries. The algorithm takes advantage of the fact that the attacker fully controls the binary modifier part of the environment but does not control the target model, which is a black box. However, an attacker can use data collected during the training of the RL agent to create and train a surrogate model of the target. The algorithm is implemented as several training/testing rounds until a final reinforcement learning policy is learned that can produce adversarial binaries for a given target. Figure 2 presents a high-level description of the algorithm, and the detailed listing is presented in Algorithm 1. Following Algorithm 1, MEME initializes a Proximal Policy Optimization (PPO) RL algorithm [35 (lines 1 and 2) to perform a first train of the PPO policy using the target model and a limited amount of steps (using MalwareGym). The goal of this step is to store a small number of observations and labels in Dsur (line 3). Then, Dsur is combined with an auxiliary set of labeled data (Daux) that the attacker should possess, and this combined dataset is used to train a surrogate model for the model extraction attack (line 5). The improved surrogate replaces the target model in a new Malware-Gym environment that trains a better policy (agent) for evasion (line 6). Last, the improved policy is evaluated against the original target model (line 7) to obtain the final evasion metrics. In this last step, we take advantage of the queries done to the target to append their output to the Dsur dataset. This loop is repeated for k rounds. During the last round, the evaluation is performed using the malware binary test set to produce the final evasion metrics (instead of the malware binary evaluation set used in the inner loop). The total number of queries to the target during training is k ∗ n, where k is the total number of rounds. For learning the policy, we use the PPO algorithm, which is a model-free onpolicy gradient method that clips the objective function, limiting the updates done to the policy to avoid too large positive changes or a minimum of negative changes in order to stabilize it. We used the Stable Baselines3 [31] implementation of PPO that uses an Actor-Critic (A2C) architecture. To train the surrogate model, it is not enough to use the data collected during the learning or evaluation of the agent in Dsur. The number of samples is not enough since we aim to minimize the queries to the target. In addition, all observations come from the extracted features of the malicious binary dataset or their adversarial modifications. Therefore using an auxiliary dataset is necessary to learn a good surrogate. This dataset is assumed to be from a similar distribution as the training data distribution of the target, however, this is not easy to achieve when the target is an AV or an unknown system. Adaptations to the Gym Environment MEME was implemented around the Malware-Gym environment (Section 3.1) in its latest version1. Several modifications were applied to the environment to fit the assumptions and constraints for this work: Adaptations to the Gym Environment – The ”modify machine type” action was removed because our tests showed that it produces invalid binaries for Windows 10 systems. – All targets were set to return hard labels (0 or 1), not scores. – Regarding the benign sections, the Malware-Gym implementation used only data from ”.text” sections. In our implementation, we use data from other sections if the ”.text” section is unavailable. – The latest environment supports the Ember and Sorel-LGB classifiers as targets. We added new environments to support the Sorel-FFNN, surrogate, and AV targets. The AV requires a web service in a virtual machine that invokes the AV static scanning capabilities. – For all target environments, we added support for saving the observations (features) and scores during training and evaluation runs so that they can be used for the training of the surrogate. Our version of the environment and the experiments performed as part of this work can be found in (https://github.com/stratosphereips/meme malware rl/releases/tag/v1.0). Please note that we are releasing the source code of our implementation for reproducibility and improvement; however, we do not release the trained models or agents to avoid potential misuse. 4.1 Evasion Evaluation To evaluate the performance of the evasion models reasonably and realistically, we limited how many queries they do to the target model and how long they run. The idea behind liming running time is that the ratio at which malware authors create new malware is high, and a method that takes too long to create evasive malware is impractical. Industry measurements such as the ones provided by AV-ATLAS [18] report that 180 new malware are generated per minute. Virus Total [41] provides an even higher measurement of 560 distinct new files uploaded per minute as of the 21st of May, 2023. These global measurements vary but show how quickly new malware variants appear, possibly due to the proliferation of Malware-as-a-Service frameworks. A recent and more conservative measurement from Blackberry mentions that their customers get attacked by 1.5 new malware per minute [39]. Given the above measurements, we decided to constrain the running time of each experiment to four hours in total. Given that the test set consists of 300 malware binaries, this corresponds to 1.25 processed binaries per minute which is lower than the most conservative reported metric we could find. The primary metric used to evaluate the malware evasion task was the evasion rate E, which is the fraction of malware that becomes evasive within the time window of four hours over the total number of malware that were initially detected by each target: E = nev/ndet . We also report the average number of binary modifications required for a malware binary to evade the target. For the Random, PPO, and MEME methods, this is equivalent to the mean episode length over all the detected malware binaries in the test set. The episode length for a non-evasive binary is equal to the maximum number of attempts, and for an evasive one is the number of changes required to become evasive. For the MAB framework, for the evasive binaries, we considered only the minimal binaries with the least amount of actions. The average number of binary modifications was impossible to measure for the GAMMA attack since the SecML framework (through which GAMMA was implemented) does not provide a straightforward way to measure this metric. 4.2 Surrogate Evaluation The surrogate models trained in MEME were evaluated using two different metrics: label agreement and explainability-based feature agreement [37]. The label agreement of two models f and ˆf, respectively, is defined as the average number of similar predictions over a test set Xtest, and it is a standard metric for model extraction fidelity attacks [19]: The feature agreement metric computes the fraction of common features between the sets of top-k features of two explanations. Given two explanations Et (target) and Es (surrogate), the feature agreement metric can be formally defined as: where top features(E, k) returns the set of top-k features of the explanation E based on the magnitude of the feature importance values. The maximum value of the feature agreement is 1. For this work, we measured the feature agreement for k = 10 and k = 20, and the explainability method used was SHAP (SHapley Additive exPlanations) [25]. SHAP employs game theory principles to explain machine learning model outputs by linking effective credit allocation with localized explanations via Shapley values that originated from game theory. SHAP is model-agnostic and has been used effectively in the past for adversarial malware creation [33]. For the LightGBM targets we used TreeSHAP which is designed specifically for tree models, while for Sorel FFNN we used the KernelSHAP variant. This paper is available on arxiv under CC BY-NC-SA 4.0 DEED license. This paper is available on arxiv under CC BY-NC-SA 4.0 DEED license. available on arxiv

MEME Algorithm: Optimizing Malware Evasion Through Model Extraction and Reinforcement Learning

96 Stories To Learn About Data Engineering

361 Stories To Learn About Big Data

86 Stories To Learn About Startup

178 Stories To Learn About Entrepreneurship

A few years ago, I participated as a mentor in a game jam. After the announcement of the winners, one of my mentees asked me: "The winners created their game by assembling assets from the Asset Store. They didn't even write three lines of code. Is that fair?" At that moment, I became convinced that the Unity Asset Store and third-party SDKs (Software Development Kit) are the best things the internet gods have sent us. Imagine a ready-to-use, tested solution that can save your project time and money. What could be better? Only a list of ten solutions that developers will surely be thrilled about. The order in the list is arbitrary. You won't find in-app purchases or advertisement SDKs here because they deserve a separate article. The same situation is with content assets because their necessity depends on specific projects. Mesh Baker This tool is essential for optimizing a 3D project. It becomes even more necessary when you don't have an experienced full-time 3D artist ready to lend a hand at any moment. Working with stock or freelance models often involves excessive materials and sub-meshes. When you insert such a model into your project, the number of draw calls overgrows. Mesh Baker takes care of the optimization process entirely. Unlike other similar programs, Mesh Baker can do everything. Here are some of its capabilities: Creating texture atlases or arrays for prop sets Finding, grouping, and combining meshes in large and complex scenes Modifying model prefabs for atlas materials usage Supporting models with multiple materials Processing tiled textures Combining, customizing, and adding objects to skinned meshes Combining dynamic meshes into a single-skinned mesh This asset also has well-documented instructions, regularly updated video tutorials, and good developer support. Amplify Shader Editor If we put it briefly, Amplify Shader Editor is the Shader Graph of a health-conscious person. Here are its advantages over other in Asset Store: shader editors Super-detailed documentation describing every piece of functionality. The developers have created their own Wikipedia Shader templates Functions for shaders Open-source code. Support for Xbox One/Xbox Series X/PS4/PS5/Switch consoles Grouping shaders in post-processing Extensive collection of samples Spine Do you continue to use frame-by-frame animation in your project? If you're tired of constantly changing Unity's toolkit for working with skeletal animation, Spine is specifically designed for you. It's not an asset from the Unity Asset Store. Still, it is the best solution for working with 2D character animation in . Unity Spine operates based on skeletal animation principles: the character is divided into parts, and the animation is created by changing the state of those individual parts. As a result, a set of 10-20 character animations can fit into a 1024x1024 pixel atlas, along with a .txt file containing the atlas layout and a JSON file with the saved animations. The animations for such a character will take up a maximum of 2MB. Now, how much space would files for 20 frames of frame-by-frame animation take? Of course, you can use one of the 10 Unity solutions for skeletal animation or find something on the Asset Store, but Spine has several advantages over them: Spine works with any 2D game engine and is suitable for creating animation clips. Finding an animator who works with Spine is much easier than finding someone who can create high-quality 2D animations in Unity. Spine bypasses the work with Unity Animator. Honestly, I can't stand working with it. If you've ever tried to set up complex transitions between animation states in Unity, you'll understand me. Let's be silent about animation events and working with animation parameters via strings. Spine allows you to control animations through code. Additionally, Spine has: a. Custom serialization of animation clips and skins b. Animation events at the end of playing an animation clip or animation c. Easy integration of animation events within the animation itself d. The ability to attach sprites to your animation skeleton that participate in the animation — creating an equipment system with Spine is a breeze e. One file for game and UI animations f. Well-documented instructions and an active community Overall, you won't want to work with other 2D animators once you try Spine. Cinemachine If your project requires something more than a single-plane static camera, use Cinemachine. When I last created a custom controller for a 2D camera, it took me several days. What I ended up with was less flexible than Cinemachine. Installing Cinemachine is the first thing I do when I need a camera controller. Here are its advantages: Support for 2D functionality Dozens of different camera controller settings Blending between different cameras using easy functions Camera shake system Ability to add up to 8 virtual cameras Obstacle avoidance system with colliders Split-screen and picture-in-picture modes Automatic zoom Post-processing effects Firebase Here I cheated a bit by grouping together a whole set of products from one developer. Firebase is irreplaceable, and the entire product group is much easier to work with than individual assets. It can offer the following: How do users respond to your project's features? Add the Firebase Analytics SDK, track events from the desired code locations, and monitor them in the web console. This way, you can see where users drop off in your app or how much time they spend on each game level. Project analytics: They allow you to change the appearance of your app without releasing an update. Remote configs: In short, it's a tool for testing the impact of functionality on key metrics of your app. For example, if you want to change the main menu design but are concerned that it might deter users, A/B testing can help. You allocate a group of users, let's say 20% of all users, and show half of them the old menu and the other half the new one. Then, you observe which group shows better metrics in terms of purchases, time spent in the app, etc., and conclude the need for a redesign. A/B testing: Firebase can handle authentication for various types, including email and password, Facebook, Google, Apple Account, and even GitHub. If a user still needs to authenticate, Firebase creates a ghost account for them, and you can utilize the data from that account after authentication. Authentication: Real-time database: A database based on a JSON file that allows you to read updates in real-time with minimal delay. And this is probably just a small part of what Firebase is capable of. DOTween DOTween is the best solution for creating animation sequences in Unity. It allows you to develop component-based animations from code, combine them into sequences, receive callbacks when they end, and launch them in asynchronous methods. DOTween also supports the use of , unlike similar assets. easing functions DOTween creates animations based on . For Rigidbody, for example, you can control position, rotation, and velocity. This means that DOTween can even be used to create movement controllers for 2D and 3D. Additionally, DOTween works great with UI animations. Unity components I'm providing a link only to the free version, as I have never encountered a moment when this functionality was insufficient throughout my career. Of course, you can tween TMPro in the pro version, but it's easy to work without it. / . Dependency Hunter Finder This asset's paid, and free versions are essential for organizing a project. You have encountered files lying in your project since its creation multiple times for 100%. They seem to be unused, but deleting them is still scary. These two assets will show you a complete list of dependencies between files, even if they are embedded in the addressable system. Odin Serializer Experience the power of Scandinavian gods working with the Unity inspector: Static Inspector allows you to select any type and view all its static members. Properties, fields, and even methods are displayed with all their arguments. You can filter the output by type or member name. Enumeration will have a proper appearance in the inspector, plus you'll be able to assign its value through search. Odin Editor Window is an extension of Unity Editor Window functionality that allows you to forget about writing GUI code. Many attributes for working with GUI elements will satisfy all your desires. You can override the GetTarget method and pass an instance of any type for display. It doesn't have to be serialized or even a Unity object. This allows you to serialize interfaces, abstract classes, and even collections of such objects. 100+ attributes that allow you to customize the inspector's behavior. You don't need to inherit from anything — mark the object with the necessary attribute, and you're done. The complete list is available at the . following link I have selected a few of the most superficial attributes and described their capabilities: a. [ChildGameObjectsOnly] — can be used on fields and properties of type Component. The compiler will iterate through child objects and return objects of the corresponding type in a dropdown. b. [SceneObjectsOnly] — works similarly to the previous attribute, but the search is performed throughout the scene. c. [AssetsOnly] — can only be used on fields and properties of type Unity.Object. This attribute allows serialization only for objects not present in the scene. d. [Required] — can be used on a field or property of type Unity.Object. It will display an error icon and a message in the inspector if the object is empty. e. [ValidateInput("ValidationMethod")] — a dropdown based on the provided IEnumerable with your specified values. The cherry on top can be another asset that pairs well with the serializer: . It will save you from broken objects, missing references, and similar issues. Combining all these features makes working with it extremely comfortable compared to similar assets. Odin Validator + Procedural UI Image UI Gradient Let's start with its functionality. Procedural UI Image is an . It can do everything the original can and even more: alternative to Unity Image Adjust the border radius and width of the image High-quality edge smoothing without impacting performance or causing jagged edges Support for border blurring, which can be used for soft shadow or glow effects UI Gradient is an extension of the previous asset: Gradients are procedural and can be animated Support for linear and quadrilateral gradients Gradients can be rotated at any angle High performance After familiarizing yourself with the functionality of this pair, you immediately realize the space saved. You can say "goodbye" to hundreds of sprites with different radii or gradients that only UI designers can distinguish. As a bonus, a similar solution is the . This asset offers more functionality but may have lower performance than the previous ones. Modern Procedural UI Kit SR-debugger If the use of previous assets can still be questioned, this particular asset, or its equivalent, must be in every project. It's pretty simple — a debug window for on-device builds. If you are developing for platforms like Standalone, iOS, Android, Windows Store, WebGL, or Web Player and you're not using something similar, I feel sorry for you. And I firmly recommend installing this asset quickly. This asset should be the first thing you add to your project after its creation. So, what can it do? Display Unity logs in a custom UI on the device with filtering, clearing, and copying capabilities. Built-in Bug Reporter that allows you to send logs via email. Built-in profiler for monitoring device resources. Open-source code that allows you to embed your dev tools into the debugger panel. These are the assets that I use most often, but there are many more of them. Nowadays, before starting a task, I increasingly check if there is a ready-made solution in the Asset Store. If I need water effect, I go to the Asset Store, realistic ropes — Asset Store, 2D pathfinding — Asset Store. The answer to "Why do companies need you if there are so many ready-made solutions?" is quite simple. A good specialist is not only someone who can do everything from scratch but also someone who can provide the most beneficial business solution in the shortest possible time. So make full use of the asset store! It will help optimize your work and allow you to spend more time on complex tasks.

10 Useful Assets and Third-party Kits for Unity Developers of Any Level

Optimize Data Migration in MongoDB: Resharding Techniques for Speed and Scalability

মঙ্গোডিবিতে ডেটা মাইগ্রেশন অপ্টিমাইজ করুন: গতি এবং স্কেলেবিলিটির জন্য রিশার্ডিং টেকনিক

Optimieren Sie die Datenmigration in MongoDB: Resharding-Techniken für Geschwindigkeit und Skalierbarkeit

Optimice la migración de datos en MongoDB: técnicas de fragmentación para velocidad y escalabilidad

Optimiser la migration des données dans MongoDB : techniques de repartitionnement pour la vitesse et l'évolutivité

MongoDB में डेटा माइग्रेशन को ऑप्टिमाइज़ करें: स्पीड और स्केलेबिलिटी के लिए रीशेयरिंग तकनीक

MongoDB でのデータ移行の最適化: 速度とスケーラビリティのためのリシャーディング手法

MongoDB에서 데이터 마이그레이션 최적화: 속도와 확장성을 위한 리샤딩 기술

Otimize a migração de dados no MongoDB: técnicas de resharding para velocidade e escalabilidade

MongoDB'de Veri Geçişini Optimize Edin: Hız ve Ölçeklenebilirlik için Yeniden Paylaşım Teknikleri

Tối ưu hóa quá trình di chuyển dữ liệu trong MongoDB: Kỹ thuật chia lại cấu trúc cho tốc độ và khả năng mở rộng

优化 MongoDB 中的数据迁移：用于速度和可扩展性的重新分片技术

Hi there, I’m Olena, the Head of Product for B2C at Atlas, a biomedical holding company that specializes in genetic technologies. Today, I’d like to share the story of our search for growth opportunities for our products and the results we encountered along the way. Let’s dive in! How It All Began As the Head of Product, I'm in charge of the user experience for our DNA and microbiota testing products across both desktop and mobile apps. When I joined Atlas, the company had been around for seven years and the average employee had been there for 5-6 years. The team needed a fresh set of eyes to take a look at our product line, so I was tasked with evaluating the current offer and finding ways to boost B2C sales. We started out by analyzing our sales funnel. The results were passable—we had average conversion rates and our acquisition strategy worked. Tweaking the sales funnel alone wasn't going to make a significant impact. We had to look into other growth hypotheses. Search for Opportunities and Early Insights Our next move was to analyze the and upselling opportunities. We put together a long list of ideas for expanding our product line but quickly realized that they were either too expensive to implement or had little profit margins, and sometimes both. So, we decided to take another look at our existing products and search for growth opportunities there. LTV We began with customer development and analytics. The first important insight we uncovered was that people talk about DNA testing a lot—during coffee breaks with colleagues or when having lunch with friends. Alright, so what we have is a fun, trendy product. Great. Armed with this knowledge, we delved into the data—studied our sales numbers, interviewed our users, and went over the 2,000 feedback forms we've gathered from customers over the past two years. As a result, we found out that 90% of people share the results of their genetic tests with family and friends. Moreover, 60% of our surveyed customers first learned about our products from a relative or someone they know. It also turned out that 10% of all sales were purchased as gifts after a successful personal experience using the product, 10% of customers bought microbiota after buying a genetic test, and another 20% purchased a genetic test after buying microbiota. Key Issues The research provided valuable insights. We learned that our products were going viral. The next step was to figure out how to improve them. Yes, people did talk about genetic testing, but we needed to know exactly what our customers were saying. Were they satisfied? Disappointed? So, we ran surveys to measure the number of satisfied and dissatisfied customers and discovered that the overall for our products was -20%. As we looked at the results from different perspectives, we discovered additional insights that we later used in product development and marketing. NPS Take the most dissatisfied customers, for example, with an NPS of -60%. They took a genetic test in order to get recommendations on the most suitable type of sport or nutrition plan. In our ads, we promised to do just that, but in reality, genetics cannot be a reliable source of information in this regard. As a result, customers' expectations were not met. We also found that customers with health-related concerns were more dissatisfied with the product than those interested in learning about their ancestry. After ranking the issues by frequency and importance, we identified two primary concerns: the test results were either not informative enough or too complicated to understand. We tried to tackle the first problem by engaging a copywriter to rewrite our text. The thing is, the information we offered was just as detailed as our competitors’, give or take a few traits, which were not critical. Still, customers remained unhappy, and we wanted to know why. Poor UX Design: The Root of All Problems In search of a solution, I tried to reproduce the initial user experience during a . I asked potential customers to log in to a test account and give me feedback. In 8 out of 12 interviews, users displayed identical patterns—they were skipping whole sections of information because of poor navigation. UX interview Our "Health" section, for example, had three subsections: multifactorial risks (20), hereditary diseases (300), and other health-related factors (20). Clicking on "Health" in the main menu led to this , but the most clickable card on the homepage directed users straight to the subsection on risks. fork After checking it out, users kept following the side menu to the sections on nutrition, sports, and origin. This way, they skipped over the entire section on hereditary diseases and other health-related factors. The same thing happened to “Origins”. Users saw a high-level list card but didn't notice the button leading to a page with detailed information. We analyzed the metrics and confirmed our suspicions: one in five clients didn't see the details of their origin, and every third user skipped the health-related sections. Results As a result, we set out to revamp the navigation in our users' personal accounts. The initiative paid off: our NPS increased from a negative average of -20% to a positive score. At the same time, we improved the visibility of upselling blocks. These combined efforts led to a 33% increase in upselling and a 20% increase in half-year LTV. During the process, we had to deal with many unknowns. To some extent, NPS is subjective and slow to respond to changes; it is difficult to establish a correlation between sales numbers and the viral effect, and cross-selling one-time products doesn't address the issue of long-term LTV. Despite all that, this story aims to show how important it is to keep searching for patterns using all available means, even when you're unsure of what you're looking for. The featured image for this article was generated with Kadinsky 2. Prompt: Illustrate a classroom with suited-up students.

Boosting Customer Loyalty: Key Lessons Learned from a 20% NPS Drop

Atlas Metrics

atlasmetrics.io

11-50 employees

Since 2021

COMPANY RANKING

ATLAS METRICS

EVERGREEN INDEX #1158

Atlas Metrics WIKI

Apply to get your Tech Company News Page live on HackerNoon.com today!

Atlas Metrics

atlasmetrics.io

11-50 employees

Since 2021

COMPANY RANKING

ATLAS METRICS

EVERGREEN INDEX #1158

RELATED COMPANIESRANK

STORIES ABOUT ATLAS METRICS

Atlas Metrics WIKI

Apply to get your Tech Company News Page live on HackerNoon.com today!