Revealing the compactness of real samples via image reconstruction for deepfake detection

The escalating threats posed by deepfakes to society and cybersecurity have triggered public anxiety, and growing efforts have been devoted to this pivotal research on deepfake detection. The generalization capability of existing models encounters a serious challenge. A prevailing explanation is that models tend to overfit artifacts in fake samples, thereby neglecting the exploration of available real ones. Prior studies have indicated that real images exhibit intra-class clustering and inter-class uniformity in the latent feature space, termed as compactness. Since deepfakes disrupt this property, exploring the common compactness of real samples may boost the generalization of models. In light of this, this paper proposes a targeted Compact Reconstruction Learning (CRL) strategy. It applies an enhanced Multi-View Reconstruction Loss (for self-compactness) to reconstruct only real images and a new Real-Sample Compactness Loss (for other-compactness) to bolster ties across real samples. Besides, a novel Joint-Guided Reasoning (JointGR) module is introduced, which richly fuses features from the encoder-decoder and reconstructed differences. It fully capitalizes on multi-source features from CRL while improving the representational ability of our model. Under the latest benchmark, extensive experiments show our model keeps the competitive performance on most challenging datasets, even achieving state-of-the-art results on some. The code will be open-sourced at https://github.com/Dongyu-Han/CRL.