FedAtlas: a Privacy-Enhanced Federated Multi-Task Learning Framework Driven by Semantic Communication for Dual Heterogeneity

Federated Multi-Task Learning (FMTL) faces fundamental challenges arising from dual heterogeneity, including intrinsic task heterogeneity and communication challenges induced by heterogeneous system environments. Task heterogeneity often leads to negative knowledge transfer due to complex non-linear task relationships, while communication inefficiency constrains system scalability and training efficiency. Existing FMTL methods typically rely on shallow task relation modeling and overlook the structural nature of task dependencies. To address these challenges, we propose FedAtlas, a structure aware and privacy-enhanced FMTL framework that explicitly constructs a global non-linear task relationship atlas through an adaptive structure inference module. Guided by the inferred atlas, FedAtlas adopts a structure-aware dual-level aggregation strategy to jointly optimize shared encoders and personalized decoders, effectively mitigating negative transfer while preserving task-specific specialization. Furthermore, FedAtlas integrates a semantic communication mechanism to alleviate communication inefficiency under heterogeneous system conditions. By leveraging client-specific vector quantization codebook, the proposed design enables compact semantic feature transmission while providing intrinsic privacy protection without explicit noise injection. Extensive experiments demonstrate that FedAtlas outperforms state-of-the-art FMTL methods, showing 4.82% improvement in performance and efficiency trade-off.