DDL: Effective and Comprehensible Interpretation Framework for Diverse Deepfake Detectors
Zekun Sun, Na Ruan, Jianhua Li · IEEE Transactions on Information Forensics and Security · 2025
In the context of escalating advancements in AI generative technologies, Deepfakes, the sophisticated face forgeries created using deep learning methods, have emerged as a significant security threat. The predominant countermeasures are Deepfake detectors based on deep learning (DL). However, due to the opaque nature of DL-model, they struggle to offer understandable explanations for their predictive decisions, which undermines their reliability and effectiveness in real-world applications. Existing mainstream DL-oriented interpretation approaches, the feature attribution methods, struggle to work on Deepfake detectors due to issues of low interpretation fidelity, poor intelligibility, and limited applicability across different types of detectors. This paper addresses these critical challenges by proposing the Deepfake Detector Lens (DDL), a novel framework designed to enhance the interpretability of diverse architectural Deepfake detectors, encompassing those based on image, frequency domain, and video.DDLemploys a heuristic algorithm to enhance interpretation efficacy and incorporates image segmentation and face parsing techniques to bridge the gap between the machine-generated interpretation saliency map and human understanding. Comprehensive evaluations ofDDLdemonstrate its superiority over existing feature attribution methods in terms of fidelity, intelligibility, and applicability. The proposedDDLsignificantly advances the interpretability of Deepfake detection technology, offering a more reliable and understandable tool for combating AI-generated face forgeries.