ARISE: Real-reference-guided stability enhancement for continual multimodal forgery detection

ARISE targets incremental continual multimodal forgery detection under attacks, improving history retention and final average performance over continual learning baselines across three attack evolution protocols and multiple detector architectures. The method uses a Gaussian mixture model fitted only on real samples as a historical reference; GRASP constrains historically stable directions with an anisotropic metric, and VAR assigns bounded replay weights to vulnerable stages; the accompanying Fable-Bench organizes attack evolution with joint-modality, text-side and image-side protocols.