Authors:
(1) Xueying Mao, School of Computer Science, Fudan University, China (xymao22@[email protected]);
(2) Xiaoxiao Hu, School of Computer Science, Fudan University, China ([email protected]);
(3) Wanli Peng, School of Computer Science, Fudan University, China ([email protected]);
(4) Zhenliang Gan, School of Computer Science, Fudan University, China (zlgan23@[email protected]);
(5) Qichao Ying, School of Computer Science, Fudan University, China ([email protected]);
(6) Zhenxing Qian, School of Computer Science, Fudan University, China and a Corresponding Author ([email protected]);
(7) Sheng Li, School of Computer Science, Fudan University, China ([email protected]);
(8) Xinpeng Zhang, School of Computer Science, Fudan University, China ([email protected]).
Editor’s note: This is Part 6 of 7 of a study describing the development of a new method to hide secret messages in semantic features of videos, making it more secure and resistant to distortion during online sharing. Read the rest below.
Table of Links
4.3. Ablation Study
Embedding Position of Secret Message. In our generation network with 9 Secret-ID blocks, we explore different positions for embedding the secret message. We divide the secret message into two 9-bit segments and allocate their positions. In detail, Setting (a): 1st-4th blocks and 5th-9th blocks.
Setting (b): 1st-2nd blocks and 3rd-4th blocks. Setting (c): 5th-6th blocks and 7th-8th blocks. They are in comparison of the standard setting of RoGVS: 1st-3rd blocks and 4th-6th blocks.
Table 2 displays the performance for these four setups. Both Settings b and c show a considerable decrease compared to Settings a and d, suggesting that adding more Secret-ID blocks improves performance. Notably, Setting c outperforms Setting b, indicating the higher influence of subsequent blocks on the generated image.
Ablation on Attacking Layer, λ & Discriminator. Fig 5 shows even without the module, our method demonstrates considerable robustness, surpassing the three comparative methods. The addition of attacking layer improves accuracy by an average of 6%. Table 3 presents the impact of λ on the extraction accuracy. More ablation results on λ and the discriminator are displayed in the supplement.
