Federated learning-based cross-domain face forgery detection for consumer IOT systems

In this paper, we propose a novel federated learning framework for cross-domain face forgery detection in consumer IoT environments, addressing challenges such as data privacy, cross-domain generalization, and the utilization of unlabeled data. Our approach integrates a multi-modal feature extraction module leveraging both spatial and frequency domains, combined with adversarial distribution alignment to separate domain-specific and shared forgery features. Additionally, a federated distillation mechanism is employed to refine target domain adaptability while preserving privacy. This framework ensures effective separation of forgery-related features and improves robustness across unseen forgery types. Extensive experiments on benchmark datasets, including FF++, Celeb-DF, DFDC, and synthetic IoT-like environments, demonstrate superior detection accuracy, enhanced cross-domain performance, and compliance with privacy constraints.