抗深度取证的多粒度融合图像修复网络

数字图像的真伪判别是图像安全领域中的基础问题,因数字媒体极易被攻击篡改,针对图像的取证技术得到了广泛的研究.另一方面,对图像篡改反取证技术的研究,不仅追求更逼真的图像篡改操作,也从相反的方向促进了取证技术的发展.图像修复作为基础的图像篡改操作,一直是国内外学者的研究热点.针对被修复篡改后的图像会被深度取证网络取证的问题,提出了一种抗深度取证的多粒度融合图像修复(multi-granularity fusion-based image inpainting network resistant to deep forensics,MGFR)网络.MGFR网络包括编解码器、多粒度生成模块以及多粒度注意力模块.首先,输入的破损图像被编码器编码成深度特征,深度特征通过多粒度生成模块生成3个不同粒度中间特征;然后,采用多粒度注意力模块来计算不同粒度中间特征之间的相关性并将其融合;最后,融合特征通过解码器生成输出结果.另外,所提出的MGFR网络被重建损失、模式噪声损失、深度取证损失以及对抗损失联合监督.研究结果显示,所提出的MGFR网络在拥有较好的修复性能的同时能成功规避深度取证网络的取证.

Multi-granularity fusion-based image inpainting network resistant to deep forensics

Considering that it is relatively easy to use technological tools to attack and tamper with digital media, image forensics technology has been studied extensively in the field of image security. In addition to developing more realistic image forgery operations, research on image forgery anti-forensics technology also promotes forensics technology development in the opposite direction. The process of image inpainting has always been a research hotspot. This paper proposes a multi-granularity fusion-based image inpainting network resistant to deep forensics (MGFR) network. MGFR network comprises three parts: a codec, multi-granularity generation module, and a multi-granularity attention module. First, the encoder encodes the input damaged image into depth features, and then the depth features are generated by the multi-granularity generation module into three intermediate features. Subsequently, we use the multi-granularity attention module to fuse the intermediate features of different granularities. As a final step, the fused features are passed through the decoder to produce the output. Additionally, the proposed MGFR is jointly supervised by reconstruction loss, pattern noise loss, deep forensic loss, and adversarial loss. Experimental results reveal that the proposed MGFR avoids the forensics of deep forensics networks while maintaining decent inpainting performance.