CDD-RFLNet: Continual Disagreement-Distilled Residual-Feature Learning for Open-World Deepfake-Video Detection

Abstract

Deepfake detectors are usually trained on a fixed collection of manipulation methods, although synthetic video generation continues to evolve after deployment. This paper presents CDD-RFLNet, a continual open-world framework that combines RGB, constrained-residual, temporal-change, and forensic-disagreement cues. A calibrated gate uses predictive energy, residual-prototype distance, and temporal disagreement to identify unfamiliar clips. When labeled examples from a new domain become available, a frozen stability teacher preserves earlier knowledge while a plasticity teacher learns the emerging manipulation. Reliability-weighted distillation transfers their shared evidence, and localized disagreement controls lightweight stage adapters. Prototype replay limits forgetting without retaining the full historical training set. Trained initially on FaceForensics++, the model is evaluated through a sequential protocol that introduces unseen manipulation families and datasets. CDD-RFLNet achieves 91.4% unknown-detection AUROC and, after four adaptation stages, 93.7% average AUC with a standard deviation of 0.4% and only 1.0 points of forgetting. These results show that a compact detector can recognize distribution change, learn of emerging attacks, and preserve earlier forensic knowledge.