Federated Behavioural Planes: Explaining the Evolution of Client Behaviour in Federated Learning

TL;DR

This paper addresses the challenge of understanding evolving client behavior in Federated Learning by introducing Federated Behavioural Planes (FBPs), consisting of the Error Behavioural Plane (EBP) for predictive errors and the Counterfactual Behavioural Plane (CBP) for decision-making via counterfactuals. It also proposes Federated Behavioural Shields (FBS), a robust aggregation mechanism that leverages FBPs to weight client updates and defend against malicious or noisy clients without prior attacker knowledge. Empirically, counterfactual generators trained alongside FL maintain predictive performance, FBPs reveal meaningful client clusters and trajectories, and FBS consistently outperforms traditional defenses across diverse datasets and attack types. The framework enhances transparency, trust, and security in FL, offering a practical approach to monitor, analyze, and control client dynamics in non-IID, privacy-preserving settings.

Abstract

Federated Learning (FL), a privacy-aware approach in distributed deep learning environments, enables many clients to collaboratively train a model without sharing sensitive data, thereby reducing privacy risks. However, enabling human trust and control over FL systems requires understanding the evolving behaviour of clients, whether beneficial or detrimental for the training, which still represents a key challenge in the current literature. To address this challenge, we introduce Federated Behavioural Planes (FBPs), a novel method to analyse, visualise, and explain the dynamics of FL systems, showing how clients behave under two different lenses: predictive performance (error behavioural space) and decision-making processes (counterfactual behavioural space). Our experiments demonstrate that FBPs provide informative trajectories describing the evolving states of clients and their contributions to the global model, thereby enabling the identification of clusters of clients with similar behaviours. Leveraging the patterns identified by FBPs, we propose a robust aggregation technique named Federated Behavioural Shields to detect malicious or noisy client models, thereby enhancing security and surpassing the efficacy of existing state-of-the-art FL defense mechanisms. Our code is publicly available on GitHub.

Figures (16)

• Figure 1: The Federated Behavioural Planes framework enables the visualization of client behaviour in FL from two perspectives: predictive performance (Error Behavioural Plane) and decision-making processes (Counterfactuals Behavioural Plane). It highlights client trajectories and similarities, offering insights into client interactions and supporting the introduction of a new and effective robust aggregation mechanism with performance that surpasses state-of-the-art baselines.
• Figure 2: Relative variation of client-proximity across datasets.
• Figure 3: Client trajectories on Counterfactuals and Error Behavioural Planes for Synthetic, Breast Cancer, and small-MNIST datasets, corresponding to Inverted-loss, Crafted-noise, and Inverted-gradient attacks, respectively. The figure highlights the deviation of the malicious client (red) from honest clients, who tend to cluster together over time, along with the previous-round global model (S)
• Figure 4: Comparative analysis of Federated Behavioural Shields and its simpler version with only counterfactuals (cf) or predictive-performance (error) versus Krum, Median, and Trimmed-mean defenses across five attack types—No attack, Crafted-noise, Inverted-gradient, Label-flipping, Inverted-loss—on three distinct datasets. Red dashed lines represent the accuracy achieved using FedAvg without attackers.
• Figure 5: Comparison of the computational time of our proposed method against the traditional FedAvg, and the robust aggregation Krum per round of training, across (a) different model parameters and (b) different numbers of clients.

Theorems & Definitions (2)

• Definition 3.1: Error Behavioural Plane
• Definition 3.2: Counterfactual Behavioural Plane