FRoCL: Robust Continual Defense Against Deepfakes in Facial Identification Systems

Facial identification in consumer electronics is increasingly threatened by deepfakes, requiring robust and adaptive defenses. Modern generative adversarial networks (GANs) and diffusion models generate highly realistic forgeries that bypass conventional verification, particularly in resource-constrained and compressed-input settings. Current deepfake detectors face two critical challenges: i) they suffer from catastrophic forgetting when learning sequential forgery types, and ii) they remain vulnerable to adversarial perturbations. These issues are addressed in isolation with continual learning that reduces forgetting but neglects robustness. Conversely, adversarial defenses improve resilience but fail to retain prior knowledge. This paper proposes a novel framework, Forgery-Robust Continual Learning (FRoCL), a unified approach that balances plasticity, stability, and robustness. FRoCL integrates supervised learning with knowledge distillation, feature alignment, contrastive objectives, and adversarial consistency, supported by an uncertainty-driven replay buffer under strict memory constraints. Extensive experiments on FaceForensics++, Celeb-DF, and DFDC under sequential clean and adversarial task protocols show that FRoCL consistently outperforms state-of-the-art baselines, achieving less than or equal to 5% forgetting and up to 40% improvements in adversarial robustness, while preserving efficiency at inference. These findings confirm the potential of FRoCL as a practical and future-proof solution for integrating forgery detection into consumer electronics.