Taming Fat-Tailed ("Heavier-Tailed'' with Potentially Infinite Variance) Noise in Federated Learning
This addresses a key gap in federated learning theory for handling fat-tailed noise, which is incremental but important for robust algorithm design in distributed systems.
The paper tackles the problem of federated learning with fat-tailed noise that may have infinite variance, proposing FAT-Clipping algorithms that achieve convergence rates of O((mT)^((2-α)/α)) for strongly-convex and O((mT)^((1-α)/(3α-2))) for non-convex settings, where α is the bounded moment parameter.
A key assumption in most existing works on FL algorithms' convergence analysis is that the noise in stochastic first-order information has a finite variance. Although this assumption covers all light-tailed (i.e., sub-exponential) and some heavy-tailed noise distributions (e.g., log-normal, Weibull, and some Pareto distributions), it fails for many fat-tailed noise distributions (i.e., ``heavier-tailed'' with potentially infinite variance) that have been empirically observed in the FL literature. To date, it remains unclear whether one can design convergent algorithms for FL systems that experience fat-tailed noise. This motivates us to fill this gap in this paper by proposing an algorithmic framework called FAT-Clipping (\ul{f}ederated \ul{a}veraging with \ul{t}wo-sided learning rates and \ul{clipping}), which contains two variants: FAT-Clipping per-round (FAT-Clipping-PR) and FAT-Clipping per-iteration (FAT-Clipping-PI). Specifically, for the largest $α\in (1,2]$ such that the fat-tailed noise in FL still has a bounded $α$-moment, we show that both variants achieve $\mathcal{O}((mT)^{\frac{2-α}α})$ and $\mathcal{O}((mT)^{\frac{1-α}{3α-2}})$ convergence rates in the strongly-convex and general non-convex settings, respectively, where $m$ and $T$ are the numbers of clients and communication rounds. Moreover, at the expense of more clipping operations compared to FAT-Clipping-PR, FAT-Clipping-PI further enjoys a linear speedup effect with respect to the number of local updates at each client and being lower-bound-matching (i.e., order-optimal). Collectively, our results advance the understanding of designing efficient algorithms for FL systems that exhibit fat-tailed first-order oracle information.