MDP-FL: Multimodal Dynamic Purification Federated Learning

To address the challenge of defending against backdoor attacks in multimodal federated learning, this paper proposes a multimodal dynamic purification federated learning framework (MDP-FL). Traditional unimodal defense methods fail to effectively counter covert backdoor attacks in multimodal scenarios due to their neglect of modal interaction characteristics and privacy constraints. MDP-FL implements dynamic defense through a dual-layer architecture comprising a client behavior monitoring layer and a multimodal parameter purification layer. The client behavior monitoring layer quantifies the modal consistency of client parameter updates based on the modal contribution difference index, generating a suspicious client list via dynamic threshold detection. The multimodal parameter purification layer employs gradient projection correction and attention weight calibration techniques to eliminate backdoor parameter perturbations while preserving the model's semantic integrity. Experiments demonstrate that on the VQAv2 dataset, MDP-FL reduces the backdoor attack success rate from 90.11% to 24.26%, with model accuracy fluctuations kept below 2.1%, and reduces single-round aggregation time by 41% - 73% compared to traditional methods. This framework outperforms existing approaches in terms of defense effectiveness, computational efficiency, and cross-model adaptability, offering a novel perspective for the secure deployment of multimodal federated learning.