A Multi-Domain Camera Model Identification Feature Restoration Network to Counter AI Compression Attacks

Camera model identification(CMI) has been a well-established passive forensic technique in recent decades. Previous research primarily focused on unaltered images, while real-world scenarios often involve malicious attacks. With the rapid advancement of AI compression technology, it is becoming increasingly popular and prevalent in various applications. However, our experimental study demonstrates that the camera model identification drops dramatically when applied to AI-compressed images without any processing. In order to enhance the robustness of the camera model identification algorithm against AI compression attacks, we propose a multi-domain camera model identification feature restoration Network. Firstly, a spatial and frequency domain dual branch architecture is proposed to restore forensic features from AI compressed image patches. The frequency feature restoration module employs an unbalanced restoration strategy in amplitude and phase for efficient network training and improved identification robustness. Afterwards, a feature fusion and selection module are utilized to fuse and filter restored features, where double pooling cross fusion strategy is adopted in channel attention mechanism for making full use of channel statics. Experimental results show that the proposed method can efficiently restore camera model related forensic features and significantly improve the performance of the CMI model on AI compressed images.