TFD-Video: Threshold-Aware Federated Deepfake Detection for Video Forensics

Deepfakes and other AI-based manipulated and synthesized videos pose a rising threat, yet training robust detectors under federated settings is challenging, where data is siloed and communication is limited. Existing compression strategies for Federated Learning (FL) can reduce detection accuracy for forensics, and standard federated averaging (FedAvg) may weaken client-specific adaptations that are important in practice. To this end, we propose Threshold-Aware Federated Deepfake Detection (TFD), a framework that combines communication-efficient structured sparsification with client-side optimization. Each client augments a shared video backbone with learnable per-filter thresholds that gate computations and induce structured sparsity. During training, clients update local weights and thresholds but transmit only compact threshold vectors. The server aggregates these vectors to establish a common sparsity pattern, while dense model parameters and data remain local. This efficient threshold-sharing scheme dramatically reduces communication overhead by several orders of magnitude while enabling each client to maintain a model tailored to its local data. We evaluate TFD on deepfake video benchmarks (FaceForensics++ and Celeb-DF), reporting detection metrics (accuracy, AUROC/AUPRC) and efficiency indicators (per-round payload size and costs). The results indicate that TFD can significantly reduce costs and latency while maintaining competitive detection performance relative to state-of-the-art approaches.