Federated training affects deepfake detector explanations with varied client data
Does Federated Training Degrade Explainability? A Comparative Analysis of Grad-CAM Faithfulness in Centralized vs. Federated Deepfake Detection
Artificial IntelligenceCryptography and SecurityComputer Vision and Pattern RecognitionMachine Learning
Summary
Detecting fake videos is important but sharing all data for training raises privacy concerns. Federated learning trains models by only sharing updates, not raw data, helping protect privacy. The authors study if this approach harms the explanations that show why the model flagged a video as fake. They find that when data among clients is similar or moderately different, explanation quality stays good or even improves compared to centralized training. However, when data differences are very high, explanations become unreliable and misleading.
What this means in practice
- •For ai engineers: Develop privacy-preserving deepfake detection systems that maintain trustworthy explanations under moderate client data differences.
- •For privacy teams: Assess when federated learning can be safely deployed for collaborative model training without losing explanation reliability.
Authors
Muhammad Uzair Shahid, Muhammad Zabil Mehboob
Abstract
Automated detection has become the essential safeguard against realistic synthetic media, but deploying it at scale requires training on sensitive data from multiple organizations, putting privacy and the transparency demanded by forensic review in direct conflict. Federated learning resolves the privacy side by sharing only model updates rather than raw data, but whether this comes at the cost of trustworthy gradient-based explanations, which analysts rely on to decide whether or not to trust a flagged clip, has not been studied. We present the systematic study of Grad-CAM explanation faithfulness under federated deepfake detection, evaluating an EfficientNet-B4 detector on FaceForensics++ across five federated clients under varying degrees of data heterogeneity. We find that the relationship between federation and explanation quality is non-monotonic: at near-IID (Independent and Identically Distributed Data) and moderate heterogeneity, federated models match or exceed the faithfulness of a centralized baseline, but at extreme heterogeneity, explanations break down entirely, becoming spatially incoherent and, on average, causally misleading. This degradation is not a general cost of federated training but a threshold effect tied to client heterogeneity, and we offer practical guidance for identifying when privacy-preserving deepfake detection can be deployed safely and when its explanations can no longer be trusted. Code and data partitions are available for reproducibility.