Cross-Modal Multi-Level Backdoor Poisioning in Multimodal Federated Learning

Multimodal federated learning (MFL) enables privacy-preserving collaborative training across distributed clients with heterogeneous data modalities by leveraging cross-modal representation alignment and fusion to integrate modality-specific information from the same physical process. However, these mechanisms are also prone to emerging security vulnerabilities, which can be exploited to implant stealthy backdoors into the global model. Existing attack methods either focus on unimodal settings or rely on input-level triggers, overlooking coordinated cross-modal backdoors in MFL environments. To address this gap, we introduce an attack model named CMBA that implants a coordinated cross-modal multi-level backdoor attack by manipulating multiple learning components, including the classification objective, representation alignment, and latent structure. This design allows the backdoor to remain semantically consistent across modalities while manipulating the underlying multimodal representation space, enabling the attack to persist under federated aggregation.