FilFL: Client Filtering for Optimized Client Participation in Federated Learning

TL;DR

FilFL introduces a novel client-filtering layer for federated learning that selects a cooperative subset of clients via a central, non-monotone combinatorial objective solved with a greedy algorithm. The approach, implemented as χGF with deterministic and randomized variants, is followed by standard FL client selection on the filtered set, yielding faster convergence and up to about 10 percentage points higher test accuracy across vision and language tasks under time-varying client availability. The authors provide convergence guarantees showing a $\mathcal{O}(\tfrac{1}{t})$ rate up to a filtering-dependent neighborhood, and demonstrate strong empirical gains along with robustness through ablations. This technique reduces communication and computation while improving generalization, with open-source code and a companion technical report for deeper theory.

Abstract

Federated learning, an emerging machine learning paradigm, enables clients to collaboratively train a model without exchanging local data. Clients participating in the training process significantly impact the convergence rate, learning efficiency, and model generalization. We propose a novel approach, client filtering, to improve model generalization and optimize client participation and training. The proposed method periodically filters available clients to identify a subset that maximizes a combinatorial objective function with an efficient greedy filtering algorithm. Thus, the clients are assessed as a combination rather than individually. We theoretically analyze the convergence of federated learning with client filtering in heterogeneous settings and evaluate its performance across diverse vision and language tasks, including realistic scenarios with time-varying client availability. Our empirical results demonstrate several benefits of our approach, including improved learning efficiency, faster convergence, and up to 10% higher test accuracy than training without client filtering.

Figures (22)

• Figure 1: Visualization of two scenarios with different suggested descent directions from different clients. Arrows are color-coded to indicate the quality of direction: blue for optimal, orange for favorable outlier, green for majority consensus, and red for unfavorable outlier.
• Figure 2: FilFL incorporates client filtering in FL, which is activated when the boolean condition 'Bool' becomes true, either when new clients become available or when $h$ rounds have elapsed since the last filtering call. Otherwise, the condition remains false. In both scenarios, clients are selected from the filtered-in subset of clients, denoted as $\mathcal{S}^f$.
• Figure 3: FilFL (FedAvg with $\chi$GF) vs FedAvg (w/o filtering) both with PoC on Shakespeare dataset with $N = 143$, $n=100$, $K=10$, $m=34$, and $h=5$.
• Figure 4: FilFL (FedProx with $\chi$GF) vs FedProx (w/o filtering) both with RS on FEMNIST dataset with $N = 190$, $n=50$, $K=5$, $m=2000$, and $h=5$.
• Figure 5: FilFL (FedProx + $\chi$GF + RS) vs FedProx (RS) without filtering on Shakespeare dataset.

Theorems & Definitions (17)

• Remark 1
• Remark 2
• Remark 3
• Remark 4
• Remark 5
• Theorem 1
• proof
• Lemma 1
• proof
• Lemma 2