MESH-FL Boosts Multimodal Federated Learning on Edge Devices

Federated learning (FL) is increasingly used for multimodal models on mobile and edge devices, but these environments often have varying sensing capabilities and computational power. Existing compression methods for model updates typically apply uniform policies, failing to account for the unique spectral structure and compressibility of different modalities and layers. To address this, MESH-FL (Multimodal Entropy-guided SVD-based Heterogeneous FL) has been developed. This framework uses an entropy-guided matrix product state (MPS) compression technique that adaptively allocates compression ranks. It estimates the spectral entropy of each layer-wise update and assigns ranks based on this entropy, considering modality-specific differences and per-client payload budgets. The research demonstrates that higher spectral entropy necessitates a higher reconstruction rank, and MESH-FL's allocation strategy solves a convex surrogate rank-allocation problem. Experiments on a heterogeneous Raspberry Pi cluster showed MESH-FL achieving up to 56.8x compression, surpassing uncompressed FedAvg in accuracy by up to 2.01%, and reducing total transmitted data by up to 66x to reach convergence.