On the Push-Based Asynchronous Federated Learning: A Bias-Correction Aggregation Approach

Researchers propose PushCen-ADFL, a new framework for asynchronous decentralized federated learning that reduces communication overhead by over 80% while improving accuracy under data heterogeneity. The approach uses centroid-based message compression and bias-correction aggregation to enable stable model training across distributed systems without central coordination.

PushCen-ADFL addresses a fundamental scaling challenge in federated learning systems: how to train machine learning models across distributed participants without a central server while managing communication costs and data heterogeneity. Traditional federated learning requires global synchronization, creating bottlenecks in large-scale deployments. This research tackles the harder problem of asynchronous learning, where participants update at different rates and communicate directly with peers.

The framework's innovation centers on three interconnected mechanisms. First, centroid-based message compression reduces bandwidth requirements dramatically—an 80% communication reduction carries significant practical implications for edge computing and IoT deployments where bandwidth constraints are severe. Second, the push-sum mixing algorithm corrects aggregation bias that typically emerges when nodes update asynchronously and non-uniformly, a persistent problem in decentralized systems. Third, centroid regularization stabilizes model convergence despite participant staleness and data distribution skews, addressing the non-IID (non-independent and identically distributed) data challenge that plagues real-world federated settings.

For developers building distributed machine learning systems, this work offers a practical pathway to reduce infrastructure costs while maintaining model quality. The 6% accuracy improvement under heterogeneous data directly translates to better real-world performance in applications ranging from healthcare to mobile device networks. The bounded-buffer design for handling irregular asynchronous arrivals demonstrates thoughtful engineering beyond theoretical contributions.

The research signals growing maturity in decentralized AI systems, moving beyond centralized cloud approaches. Success in this domain could accelerate adoption of federated learning in privacy-sensitive applications and resource-constrained environments, though real-world validation at production scale remains the next critical milestone.

• →PushCen-ADFL reduces per-push communication costs by over 80% compared to baseline asynchronous federated learning approaches
• →Centroid-based compression and push-sum mixing enable stable training under asymmetric communication and delayed client participation without central coordination
• →Framework achieves up to 6% accuracy improvement on vision datasets under data heterogeneity, demonstrating practical viability
• →Bounded sender-deduplicated buffer design improves robustness when handling irregular asynchronous client arrivals
• →Approach couples communication efficiency with optimization stability, addressing the fundamental accuracy-communication tradeoff in distributed learning