| 摘要: |
| 目的 针对 JPEG 压缩反取证方法在生成图像质量与反取证性能之间平衡不足的问题,设计了一种融合多级
残差块和通道注意力机制的反取证模型 RBAM-JAF,以提高生成图像的质量,促使反取证性能和图像质量之间达
到更好的平衡。 方法 采用基于生成对抗网络的框架,包括生成器和鉴别器两部分,其中,生成器引入多级残差块和
通道注意力机制,使模型具有更好的泛化性,增强图像特征表示的能力;添加特征融合模块,以充分利用所有卷积
层的特征,提高生成图像的质量。 结果 实验结果显示,与现有的 4 种反取证方法 M1 、M2 、M3 和 M4 相比,在 QF = 25
的情况下,PSNR 值分别增长了 8. 52%、3. 31%、1. 52% 和 0. 07%,SSIM 值分别增长了 12. 89%、2. 46%、1. 90% 和
0. 55%;在 QF = 50 的情况下,PSNR 值分别增长了 10. 22%、2. 21%、0. 88% 和 0. 19%,SSIM 值分别增长了 9. 71%、
1. 52%、0. 64%和 0. 21%;在 QF = 75 的情况下,PSNR 值分别增长了 18. 26%、3. 56%、3. 80%和 2. 96%,SSIM 值分别
增长了 10. 83%、1. 58%、1. 16%和 0. 52%。 此外,通过 4 种取证检测器的检测,QF = 25、50 和 75 时,AUC 值均接近
或低于 0. 5。 结论 实验结果显示:方法 M5 与现有的反取证方法相比,提高了生成图像的视觉质量且能够有效地欺
骗现有的取证检测器,使反取证性能和生成图像质量之间达到了更好的平衡。 |
| 关键词: JPEG 压缩 JPEG 反取证 生成对抗网络 残差块 通道注意力机制 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| JPEG Compression Anti-forensics Integrating Residual Blocks and Attention Mechanism |
|
TANG Beibei1, CHEN Lei2, LI Ruoyu2
|
|
1. School of Artificial Intelligence Anhui University of Science and Technology Anhui Huainan 232001 China
2. School of Computer Science Huainan Normal University Anhui Huainan 232001 China
|
| Abstract: |
| Objective To address the lack of balance between the quality of generated images and anti-forensics
performance in JPEG compression anti-forensics methods an anti-forensics model RBAM-JAF combining multi-level
residual blocks and channel attention mechanism was designed. This model aimed to improve the quality of generated
images and achieve a better balance between anti-forensics performance and image quality. Methods A framework based
on generative adversarial networks GANs was employed including a generator and a discriminator. The generator
incorporated multi-level residual blocks and channel attention mechanisms to enhance the model?? s generalization capability
and improve the representation of image features. Additionally a feature fusion module was introduced to fully utilize features from all convolutional layers in order to enhance the quality of generated images. Results According to the
experimental results compared with four existing anti-forensic methods M1 M2 M3 and M4 the proposed method
showed significant improvements. At QF = 25 the PSNR values increased by 8. 52% 3. 31% 1. 52% and 0. 07%
respectively and the SSIM values increased by 12. 89% 2. 46% 1. 90% and 0. 55% respectively. At QF = 50 the
PSNR values increased by 10. 22% 2. 21% 0. 88% and 0. 19% respectively and the SSIM values increased by
9. 71% 1. 52% 0. 64% and 0. 21% respectively. At QF = 75 the PSNR values increased by 18. 26% 3. 56%
3. 80% and 2. 96% respectively and the SSIM values increased by 10. 83% 1. 58% 1. 16% and 0. 52%
respectively. Additionally the AUC values of the four detectors for QF = 25、 50 and 75 were close to or below 0. 5.
Conclusion Experimental results demonstrate that method M5 improves the visual quality of generated images compared
with existing methods while effectively deceiving forensic detectors achieving a better balance between anti-forensic
performance and the quality of the generated images. |
| Key words: JPEG compression JPEG anti-forensics generative adversarial networks residual blocks channel attention
mechanism |
