Supervised Anomaly Detection for Generalized Deepfake Detection

Generalization to unseen forgeries remains a significant challenge in deepfake detection. Existing methods often focus on detecting specific forgery patterns, such as blending boundaries, identity leakage, etc. However, their overfitting to known fake samples during training restricts their generalization to unseen manipulations. To address this limitation, we propose Supervised Anomaly Detection for Generalized Deepfake Detection (SAD-GDD), a novel framework that treats real data as normal and fake data as anomalies, leveraging anomaly detection to improve generalization. At its core, we propose a Supervised Anomaly Detection Strategy (SADS), which constructs a refined Multivariate Gaussian Distribution (MGD) for real data and applies supervised guidance for fake data to strengthen the discrimination between real and fake. To further enhance performance, we design a two-stream network comprising a High-Frequency Texture Branch (HFT-B) and an RGB Branch (RGB-B), both optimized with SADS. Additionally, we propose a Multi-head Cross-Attention Enhancement Block (MCAEB) to enhance the capability of RGB-B in capturing high-frequency features. Extensive experiments show that SAD-GDD achieves superior generalization compared to state-of-the-art methods across multiple widely used datasets. The code will be made publicly available upon publication.