OC-SAN: Unsupervised Deepfake Detection for Specific Individual Protection Based on Deep One-Class Classification
Abstract: In recent years, the misuse of Deepfake technology has become a critical security concern, especially for high-profile individuals such as politicians and celebrities. The key figure targeted forgeries are usually been carefully crafted with a sophisticated process, leaving behind no manipulation-specific artifacts. The inadequacy of current supervised binary classification methods for addressing these specific individual protection (SIP) challenges is evident for two primary reasons. Firstly, sufficient fake samples cannot be ensured to train generic and robust classifiers. Secondly, the efficacy of detection varies significantly across different identities. To address these problems, we formulate forgery detection as a one-class anomaly detection problem and propose a detection network, OC-SAN (One-Class Style Auto-encoder Network), to offer tailored protection for each individual. Borrowing from style transfer literature, we conceptualize facial features as a fusion of coarse-grained (identity) and fine-grained (personalized/style) elements. The fundamental concept driving OC-SAN is that a facial image can be well restored when its two features belong to the same identity. As a one-class method, OC-SAN can be trained only with authentic samples, indicating good generalization performance in real-world scenarios without relying on prior knowledge of forgery methods. Extensive experiments have been conducted to demonstrate the superiority of OC-SAN in SIP tasks, when compared to other state-of-the-art forgery detection methods.
Loading