Chinese Researchers Embed Secret Messages in Videos That Survive Social Media Distortions

Table of Links

• Abstract and 1. Introduction
• 2. Related Work
• 3. Proposed Approach
• 4. Experiments
• 4.1. Experimental Setups
• 4.2. Performance Analysis
• 4.3. Ablation Study
• Conclusions and References

3. PROPOSED APPROACH

3.1. Information Encoding Module

3.2. Secret Message Embedding and Extraction Module

This module aims to embed the secret message during face swapping. The key problem is how to implement face swapping under the guidance of secret message. To our understanding, the latent features of the cover video encompass both identity and attribute feature. Face swapping essentially involves replacing the cover video’s identity with that of the reference image. Consequently, we embed the secret message into the identity feature of the reference image, formulated as follows:

where λ is a hyper-parameter adjusting the influence of secret message on identity feature.

3.3. Attacking Layer

To bolster the robustness of our method for face-swapping videos in real-world scenarios, we design a attacking layer. This module simulates prevalent distortions encountered across social network platforms.

JPEG Compression. JPEG compression involves a nondifferentiable quantization step due to rounding. To mitigate this, we apply Shin et al.’s method [53] to approximate the near-zero quantization step using function Eq. (6):

where x denotes pixels of the input image. We uniformly sample the JPEG quality from within the range of [50, 100].

Color Distortions. We consider two general color distortions: brightness and contrast. We perform a linear transformation on the pixels of each channel as the formula Eq. (7):

where p(x) and f(x) refers to the distorted and the original image. The parameters a and c regulate contrast and brightness, respectively.

Color Saturation. We perform random linear interpolation between RGB and gray images equivalent to simulate the distortion.

Additive Noise. We use Gaussian noise to simulate any other distortions that are not considered in the attacking layer. We employ a Gaussian noise model (sampling the standard deviation δ ∼ U[0, 0.2]) to simulate imaging noise.

3.4. Loss Function

The proposed method ensures both high stego video quality and precise extraction of secret message. We achieve this by training the modules using the following losses.

Table 1. Comparison Results on Extraction Accuracy. “-” means “Without Distortion”. (·) represents Bits Per Frame (BPF).Under different distortion scenarios, our method demonstrates superior performance in comparison.

Attribute Loss. We use the weak feature matching loss [26] to constrain attribute difference before and after embedding secret message. The loss function is defined as follows:

where Dj refers to the feature extractor of Discriminator D for the j-th layer, Nj is the number of elements in the j-th layer, and H is the total number of layers. Additionally, h represents the starting layer for computing the weak feature matching loss.

Adversarial Loss. To enhance performance, we use multiscale Discriminator with gradient penalty. We adopt the Hinge version of adversarial loss defined as follows:

Secret Loss. To address this, we use the Binary Cross-Entropy loss (BCE) as defined in Eq. (11).

Total loss. The total loss is defined as follows: