Federated Multimodal Unlearning via Entanglement-Aware Anchor Closure

The newly introduced method, EASE, tackles the issue of unlearning in federated multimodal learning (FML). In FML, models are trained using private image-text pairs across decentralized clients; however, joint embedding complicates the unlearning process by intertwining forgotten knowledge within modalities and client gradient subspaces. Current federated unlearning methods do not effectively disconnect the cross-modal reconstruction channel influenced by bilinear coupling or differentiate between update directions for forgotten and retained knowledge. The authors propose an Anchor Principle, noting that forgotten alignments are sustained by three residual anchors from bilinear coupling, subspace entanglement, and ongoing federated updates. EASE, which aims to enhance privacy in decentralized multimodal learning, is detailed in arXiv (2605.00733).