FIRMA: FIbonacci Ring Model Aggregation for Privacy-preserving Federated Learning

Background

Federated learning faces significant challenges, including the risk of central server failures and gradient inversion attacks in centralized approaches; exposure to semi-honest peers via uniform weights in decentralized ring-gossip methods; and reintroduction of central aggregation in personalized learning. No existing protocol simultaneously solves these issues while maintaining a ring topology and asymmetric neighbour weighting.

Key Points

1. FIRMA Protocols: The proposed protocols, \fibfl, \fibflp, and \fibflpp, are designed to address the trilemma by ensuring server-free operation, privacy, ring topology, and optimal neighbour blending.

2. Protocol Progression:

   • \fibfl: Establishes a foundation with server-free ring aggregation using Fibonacci-weighted neighbour blending and permanently private heads.
   • \fibflp: Augments \fibfl\ by introducing accuracy-gated neighbour suppression to selectively down-weight poorly-converged peers while maintaining directional bias.
   • \fibflpp: Completes the family with a 2-opt ring permutation, maximizing class diversity, global coverage via $K_g{=}\lceil N/2\rceil$ gossip passes, and cosine-annealed self-retention calibration.

3. Theoretical Foundation: The authors establish convergence rate bounds and supporting propositions governing normalisation, coverage, retention, and diversity optimality.

4. Experimental Evaluations:

   • Across 28 configurations involving four benchmarks and seven heterogeneity regimes, \fibflpp\ consistently outperforms existing methods like \fedavg\ in all 12 label-skew scenarios.
   • At $K{=}1$, \fibflpp\ achieves a peak advantage of (+20.7)% on the CIFAR-10 dataset.
   • Under Dirichlet heterogeneity, \fibflpp\ is Pareto-dominant among all server-free protocols, achieving the highest accuracy in 17 out of 28 configurations.

Significance

The significance of FIRMA lies in its comprehensive solution to the trilemma in federated learning. By offering a scalable and robust framework that maintains privacy and security while ensuring high performance, \fibflpp\ significantly advances the state-of-the-art in distributed machine learning. The detailed theoretical analysis and extensive experimental validation underscore its practical utility and potential impact on real-world applications where data privacy is paramount.