Face forgery detection improves explanations using supported forensic cues
Not Every Cue Is Evidence: Cue Supportability for Forensic Distillation in Face Forgery Detection
Computer Vision and Pattern Recognition
Summary
Spotting fake face images is tricky because the fine details that reveal fakes are hard to see in low-quality pictures. The authors found that teaching AI models about these fine details can confuse them if the details can’t actually be seen in the images. To fix this, they created a method to check which clues the AI can really find in each image before teaching it. This makes the model not just accurate but also better at explaining the reasons behind its decisions.
What this means in practice
- •For digital forensic teams: Improve trustworthiness of fake face image detection by providing explanations backed by visible evidence in forensic images.
- •For security software developers: Build more reliable face forgery detection tools that avoid misleading explanations by filtering training cues for real image support.
Authors
Binjia Zhou, Dawei Luo, Shuai Chen, Feng Xu, Haoyuan Li, Jia wen Seow, Jiachi Wang, Jie Lei, Zunlei Feng, Yijun Bei
Abstract
Explainable face forgery detection requires not only correct judgments but also forensic evidence supported by the input image. However, training cues are typically derived from the forgery generation pipeline. Since images available to the model are of limited quality and resolution, it is difficult to observe the fine-grained traces these cues refer to. Training methods that further supply such cues as privileged supervision exacerbate the problem, as the student is taught to articulate evidence it cannot perceive. The student consequently resorts to templated or over-generalized expressions, giving rise to the correct-answer, wrong-evidence (CAWE) failure. We propose a dual supportability audit that filters privileged cues before distillation. Each candidate cue must be predictable from the remaining privileged context and causally tied to high-attribution visual regions in the current image. Only cues satisfying both conditions receive substantial distillation weight, preventing unsupported details from leaking into the student. Experiments across eight benchmarks show that our method maintains or improves detection accuracy while significantly raising explanation faithfulness compared to existing paradigms.