Adaptive Client Selection in Federated Learning: A Network Anomaly Detection Use Case

TL;DR

This work addresses privacy-preserving, fault-tolerant federated learning for network anomaly detection by introducing an adaptive client selection framework that applies differential privacy to model updates and employs checkpointing for resilience. The method optimizes client selection via a utility-based objective, perturbs gradients with Gaussian noise to satisfy $(\epsilon,\delta)$-DP, and models failure probabilities with a Weibull-based checkpointing strategy to derive an optimal interval $t_c^*$. Empirical evaluation on UNSW-NB15 and ROAD shows up to 7% higher accuracy and 25% faster training than baselines like ACFL and FedL2P, with statistically significant improvements ($p<0.05$) in Mann-Whitney tests; DP budgets reveal a privacy-utility trade-off, and fault-tolerance incurs modest accuracy loss but improves robustness. The results demonstrate practical impact for privacy-conscious, large-scale FL in real-world network environments and highlight avenues for adaptive hyperparameter optimization and cryptographic alternatives in future work.

Abstract

Federated Learning (FL) has become a widely used approach for training machine learning models on decentralized data, addressing the significant privacy concerns associated with traditional centralized methods. However, the efficiency of FL relies on effective client selection and robust privacy preservation mechanisms. Ineffective client selection can result in suboptimal model performance, while inadequate privacy measures risk exposing sensitive data. This paper introduces a client selection framework for FL that incorporates differential privacy and fault tolerance. The proposed adaptive approach dynamically adjusts the number of selected clients based on model performance and system constraints, ensuring privacy through the addition of calibrated noise. The method is evaluated on a network anomaly detection use case using the UNSW-NB15 and ROAD datasets. Results demonstrate up to a 7% improvement in accuracy and a 25% reduction in training time compared to the FedL2P approach. Additionally, the study highlights trade-offs between privacy budgets and model performance, with higher privacy budgets leading to reduced noise and improved accuracy. While the fault tolerance mechanism introduces a slight performance decrease, it enhances robustness against client failures. Statistical validation using the Mann-Whitney U test confirms the significance of these improvements, with results achieving a p-value of less than 0.05.

Figures (3)

• Figure 1: FL architecture showing global server coordination and local client training with checkpointing and fault tolerance.
• Figure 2: Performance comparison of the proposed method, ACFL, and FedL2P across accuracy, AUC-ROC, and training time metrics for the UNSW-NB15 and ROAD datasets.
• Figure 3: Impact of privacy budgets ($\epsilon$ values) on global test accuracy and loss for UNSW-NB15 and ROAD datasets.