Automated detection is now essential for combating realistic synthetic media, yet scaling it requires training on sensitive data from multiple organizations, creating a tension between privacy and the transparency demanded by forensic review. Federated learning mitigates privacy concerns by sharing only model updates instead of raw data, but its impact on the trustworthiness of gradient‑based explanations used by analysts has not been systematically examined. We conduct a systematic study of Grad‑CAM faithfulness in federated deepfake detection, evaluating an EfficientNet‑B4 detector on the FaceForensics++ dataset across five federated clients with varying degrees of data heterogeneity. Results reveal a non‑monotonic relationship: under near‑IID or moderate heterogeneity, federated models achieve explanation quality comparable to or exceeding a centralized baseline, whereas extreme heterogeneity leads to a complete breakdown of explanations, producing spatially incoherent and causally misleading maps. This degradation is not an inherent cost of federated training but a threshold effect tied to client heterogeneity. We therefore offer practical guidance for identifying scenarios where privacy‑preserving deepfake detection can be safely deployed and when its explanations should no longer be trusted. Code and data partitions are released for reproducibility.
Review