Tackling Selfish Clients in Federated Learning

TL;DR

Selfish clients can bias the global model toward their local optima in non-iid federated learning. The authors propose RFL-Self, a server-side robust aggregation that detects potential selfish updates via a median-norm criterion, recovers their true updates through a convex combination with the median update, and aggregates them to mitigate selfish influence while preserving normal-client performance. The approach is supported by theoretical bounds on recovery error and empirical validation on MNIST and CIFAR-10, showing that even small percentages of selfish clients can significantly degrade accuracy without mitigation, whereas RFL-Self maintains performance and fairness. Overall, the work provides a practical, low-overhead defense against a non-malicious yet impactful failure mode in FL, with avenues for extensions to adaptive selfishness, collusion, and fairness.

Abstract

Federated Learning (FL) is a distributed machine learning paradigm facilitating participants to collaboratively train a model without revealing their local data. However, when FL is deployed into the wild, some intelligent clients can deliberately deviate from the standard training process to make the global model inclined toward their local model, thereby prioritizing their local data distribution. We refer to this novel category of misbehaving clients as selfish. In this paper, we propose a Robust aggregation strategy for FL server to mitigate the effect of Selfishness (in short RFL-Self). RFL-Self incorporates an innovative method to recover (or estimate) the true updates of selfish clients from the received ones, leveraging robust statistics (median of norms) of the updates at every round. By including the recovered updates in aggregation, our strategy offers strong robustness against selfishness. Our experimental results, obtained on MNIST and CIFAR-10 datasets, demonstrate that just 2% of clients behaving selfishly can decrease the accuracy by up to 36%, and RFL-Self can mitigate that effect without degrading the global model performance.

Figures (16)

• Figure 1: An example to show how selfish clients try to alter their updates to steer the global model toward their local optima.
• Figure 2: (a) A visualization for how a selfish client can estimate $\hat{\delta}_s$ through model replacement. (b) Effect of $\alpha$ on the aggregated model and selfish update vector of a single selfish client. The aggregated model $\hat{\mathbf{w}}$ is obtained when $\hat{\delta}_s$ is included in the aggregation.
• Figure 3: Illustrating the impact of selfishness on the test accuracy of the global model using CIFAR-10 dataset.
• Figure 4: Deviation of the global update when all clients are normal to the one with selfish clients (CIFAR-10 dataset).
• Figure 5: Distance of the recovered update $\delta_s'$ to $\delta_s$: (a) when the conditions 1-3 are met, and (b) when the three conditions do not hold.

Theorems & Definitions (6)

• Theorem 1
• proof
• Remark
• Theorem 2
• proof
• Remark