HARMONY: Bridging the Personalization-Generalization Gap by Mitigating Representation Skew in Heterogeneous Split Federated Learning

Researchers introduce HARMONY, a hybrid split federated learning framework that enables heterogeneous mobile devices to perform personalized on-device inference while maintaining a generalized server backend for fallback support. By using meta-learning and server-side contrastive learning, HARMONY addresses the representation skew problem that occurs when diverse device architectures extract features incompatibly, achieving up to 43% accuracy improvements without compromising privacy or increasing latency.

HARMONY tackles a critical challenge in federated learning infrastructure where mobile devices with varying computational capabilities and data distributions must collaborate without centralizing raw data. The framework acknowledges a practical reality: devices need fast local inference for common tasks while retaining remote fallback capability for edge cases, balancing personalization with generalization. This dual-path approach mirrors real-world deployment constraints where privacy, latency, and accuracy compete for optimization.

The core innovation addresses representation skew—a phenomenon where different device architectures extract features that don't align in shared embedding spaces, crippling server-side inference. Traditional hybrid split federated learning assumes architectural homogeneity, breaking when devices use different model families or sizes. HARMONY's meta-learning approach simulates diverse extractors during training, enabling the server to learn robust feature alignment strategies across heterogeneous clients without accessing raw labels or raw data.

For the distributed machine learning ecosystem, this research extends federated learning's applicability to realistic device landscapes with varying specifications. Organizations deploying edge AI across diverse hardware—smartphones, IoT devices, embedded systems—can now maintain both personalized performance and system-wide reliability. The 28-43% accuracy gains suggest meaningful practical impact for production systems handling both in-distribution and out-of-distribution inference scenarios.

The framework's privacy-preserving design, avoiding raw label sharing, strengthens its regulatory appeal in jurisdictions with data protection requirements. Future work likely involves scaling HARMONY across larger device heterogeneity and exploring its integration with emerging edge computing architectures.

• →HARMONY solves representation skew in heterogeneous federated learning by using meta-learning to simulate diverse device architectures during training.
• →The framework achieves 43% accuracy improvement for in-distribution tasks and 28.3% for out-of-distribution tasks compared to existing methods.
• →Server-side contrastive learning aligns features from different client extractors without requiring raw data or label sharing, preserving privacy.
• →The approach balances on-device personalization with server-side generalization, maintaining acceptable latency for mobile deployment.
• →HARMONY enables practical federated learning across diverse hardware landscapes with varying computational capabilities and model architectures.