EASE: Federated Multimodal Unlearning via Entanglement-Aware Anchor Closure

Federated Multimodal Unlearning (FMU) eliminates the influence of a compromised client, a target class, or specified samples from a jointly trained multimodal model without accessing raw data. FMU faces significant challenges due to the entanglement of modalities and client gradient subspaces. Current methods cannot break the cross-modal reconstruction pipeline caused by bilinear coupling, and they do not separate unlearn-exclusive update directions from those shared with remaining clients. To address these, we first introduce an Anchor Principle to formalize bilinear cross-modal coupling, principal-angle subspace entanglement, and ongoing federated updates. We then design EASE, an Entanglement-Aware Subspace Excision framework. Within this framework, both the visual and language branches are updated simultaneously to prevent either branch from independently reconstructing the cross-modal alignment. Cosine–Sine decomposition is then employed to filter update directions unique to the unlearning process while preserving those shared among the remaining clients. In addition, a direction-selective unlearn lock is applied to prevent subsequent training from reverting toward the removed directions, thereby reinforcing the unlearning effect. Across three datasets, three backbone architectures, and three unlearning scenarios, EASE most consistently narrows the deviation from the retrain reference on both forget and retain metrics among nine baselines. For example, on Flickr30K with CLIP-B/32 under client unlearning, EASE matches the retrain reference within $0.2$ R@1 points on the forget set while staying within $4.2$ R@1 points on the retain set.