Explore and Enhance the Generalization of Anomaly DeepFake Detection

In recent years, Anomaly DeepFake Detection (ADFD) has made significant breakthroughs in terms of generalization when meeting various unknown tampers. These detection methods primarily enhance generalization by constructing pseudo-fake samples, which involve three main steps: mask generation, source-target preprocessing, and blending. In this paper, we conducted a systematic analysis of some core factors in these steps. Based on the aforementioned observations at the mask generation step, we find that previous ADFD methods have limitations as they only consider specific tampering types, which is not representative of real-world scenarios, and generate noise samples that closely resemble real samples, causing confusion and hindering generalization. To alleviate these issues, we propose our new method, which consists of the Boundary Blur Mask Generator (BBMG) and the Noise Refinement Strategy (NRS) modules. BBMG leverages the inherent characteristics of boundary blur to simulate a comprehensive range of tampering techniques, enabling a more realistic representation of real-world scenarios. In conjunction with BBMG, the NRS module effectively mitigates the influence of noise samples. Extensive ablation experiments and comparative evaluations demonstrate the effectiveness of our method.