PushCen-ADFL: Bias-Correction for Asynchronous Federated Learning
A recent study published on arXiv introduces PushCen-ADFL, an innovative framework for asynchronous decentralized federated learning (ADFL) that enhances communication efficiency. This framework tackles issues such as high communication costs, biased aggregation, and significant model drift, which arise from frequent peer-to-peer interactions, asynchronous updates on directed networks, and non-IID data. By integrating communication, aggregation, and local stabilization within a shared centroid representation space, it creates a feedback loop between compression and optimization. Clients share messages in centroid form, utilize average-preserving push-sum mixing to address aggregation bias, and implement lightweight centroid regularization to reduce drift amid heterogeneity and uneven communication. The full paper can be accessed at https://arxiv.org/abs/2605.26162.
Key facts
- Paper proposes PushCen-ADFL for asynchronous decentralized federated learning.
- Addresses communication overhead, biased aggregation, and model drift.
- Uses shared centroid representation space for coupling communication, aggregation, and stabilization.
- Clients exchange centroid-form messages.
- Applies average-preserving push-sum mixing to correct aggregation bias.
- Uses centroid regularization to mitigate drift.
- arXiv paper ID: 2605.26162.
- Published on arXiv.
Entities
Institutions
- arXiv