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

A new family of federated learning protocols called FIRMA proposes to resolve a structural trilemma that has constrained privacy-preserving model training, according to research published on arXiv. The protocols operate without a central server, keep classification heads permanently private, and use a ring topology with asymmetric neighbour weighting [1]. The work, titled "FIRMA: FIbonacci Ring Model Aggregation for Privacy-preserving Federated Learning," introduces three progressively enhanced protocols [1]. The first, referred to as FIBFL, establishes a server-free ring aggregation framework that blends neighbour models using Fibonacci-weighted coefficients while ensuring classification heads remain private to each client [2]. The second protocol, FIBFLP, builds on this by adding accuracy-gated neighbour suppression, which selectively down-weights peers that have not yet converged well, preserving the Fibonacci directional bias [2]. The full system, FIBFLPP, completes the family with a 2-opt ring permutation designed to maximize adjacent-client class diversity and global ring coverage through a number of gossip passes equal to half the client count, rounded up, alongside cosine-annealed self-retention calibration [2]. The authors also establish a convergence rate bound and three supporting propositions that govern normalisation, coverage, retention, and diversity optimality [2]. Systematic experiments were conducted across 28 configurations, combining four benchmark datasets with seven heterogeneity regimes [1]. The results show that FIBFLPP surpasses the canonical FedAvg protocol in all 12 label-skew configurations, recording a peak advantage of 20.7 percentage points on CIFAR-10 at K=1 [1]. Under Dirichlet heterogeneity, FIBFLPP emerged as the Pareto-dominant method among all server-free protocols tested, achieving the highest accuracy in 17 of the 28 configurations [1]. The research addresses a gap identified by the authors: existing server-based aggregation creates a single point of failure and gradient inversion risk, while decentralised ring-gossip alternatives expose classification heads to semi-honest peers through uninformed uniform weights, and personalised methods reintroduce central aggregation [2]. No prior protocol had simultaneously combined server-free operation, permanently private heads, a ring topology, and principled asymmetric neighbour weighting [2].