From Injection to Detection: Analyzing the Impact of Anomalous Data in Federated Learning

Abstract Federated Learning (FL) offers a privacy-preserving framework suitable for sensitive domains such as healthcare. This study aims to investigate how different types and intensities of anomalous data introduced by individual clients affect overall model performance in cross-silo FL environments. Additionally, it explores whether models analyzing gradient representations can detect such anomalies during training. We conduct systematic experiments injecting six types of anomalies at varying strengths into training data from two distinct datasets. Performance degradation is measured and statistically analyzed. Furthermore, we develop a Variational Autoencoder (VAE) trained on clean gradient representations to detect deviations caused by anomalies. Our findings indicate that the impact of anomalies on model accuracy varies significantly across datasets and anomaly types. CIFAR-10 data shows higher sensitivity compared to the biological cellular data derived from Quantitative Phase Imaging (QPI). The VAE-based gradient anomaly detection successfully identifies subtle shifts in gradient distributions, but effective differentiation is observed primarily for the QPI data. The results emphasize the importance of tailoring FL robustness and anomaly detection strategies to specific datasets and anomaly characteristics. Gradient-based detection methods show promise for enhancing FL security, but require further refinement. This work contributes critical insights for designing more reliable and secure FL systems, particularly in sensitive domains like healthcare.