Interpretable DeepFake Detection Based on Frequency Spatial Transformer

Abstract

In recent years, the rapid development of DeepFake has garnered significant attention. Traditional DeepFake detection methods have achieved 100% accuracy on certain corresponding datasets, however, these methods lack interpretability. Existing methods for learning forgery traces often rely on pre-annotated data based on supervised learning, which limits their abilities in non-corresponding detection scenarios. To address this issue, we propose an interpretable DeepFake detection approach based on unsupervised learning called Find-X. The Find-X network consists of two components: forgery trace generation network (FTG) and forgery trace discrimination network (FTD). FTG is used to extract more general inconsistent forgery traces from frequency and spatial domains. Then input the extracted forgery traces into FTD to classify real/fake. By obtaining feedback from FTD, FTG can generate more effective forgery traces. As inconsistent features are prevalent in DeepFake videos, our detection approach improves the generalization of detecting unknown forgeries. Extensive experiments show that our method outperforms state-of-the-art methods on popular benchmarks, and the visual forgery traces provide meaningful explanations for DeepFake detection.