New AI Beats Rivals in Hiding Secret Messages in Videos

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

4.2. Performance Analysis

We compare the performance of our RoGVS with image-level steganography including HiDDeN [12] and LSB [80] and video-level steganography including PWRN [35].

Video Quality Assessment. Fig 4 shows qualitative results on the integrity of generated video frames. We perform tests within and across datasets, each containing 16 test samples.

The generated faces effectively change individual identities while retaining attributes like expressions and poses. More findings are available in the supplementary materials. Fig 3 illustrates the visual effects of certain intermittent frames within the stego videos.

Comparisons on Extraction Accuracy & Robustness. We conduct extensive experiments with multiple types of distortions. Detailed distortion implementations are provided in the supplement.

The quantitative comparison results in terms of accuracy are reported in Table 1. The results show that our method can successfully extract secret message with high accuracy even after severe distortions. LSB [80] struggles even with PNG (quantization) and HiDDeN [12], though trained with a distortion module, can not generalize well to video-level distortions. PWRN [35] demonstrates robustness across numerous distortions, yet its performance remains constrained under operations such as motion blur or contrast adjustment. The proposed RoGVS method shows superior robustness to these distortions while maintaining high extraction accuracy.

Security Analysis. We use three video steganalysis tools to evaluate the security of our method. The detection performance of these three steganalysis schemes is presented in Table 4. Table 4 demonstrates that our method exhibits slightly superior security compared to the three counterparts.