Data Poisoning Against Federated Learning: Comparative Analysis Under Label-Flipping Attacks and GAN-Generated EEG Data

Federated Learning (FL) has emerged as a privacy-preserving machine learning approach, enabling collaborative model training across devices while maintaining the decentralization of raw data. This paper investigates the application of FL in insider threat detection, a critical aspect of organizational security that addresses potential risks posed by individuals with access to sensitive data. We focus on using Electroencephalogram (EEG) data to identify malicious intentions, which consists of highly sensitive brain signals that we aim to safeguard by employing FL. Despite its advantages, FL still encounters a rising threat from data poisoning attacks. This study investigates the resilience of our FL model against label-flipping attacks, utilizing three classifiers: Multiplayer Perceptron (MLP), Convolutional Neural Network (CNN), and an ensemble learning classifier, the Voting Classifier (VC). Due to insufficient EEG data, a Generative Adversarial Network (GAN) model is utilized to augment and increase the size of the data. Our findings reveal that VC demonstrates the highest performance with an accuracy of 95% for the original dataset. In contrast, CNN is the sole classifier that outperformed others with the GAN-generated dataset, achieving an accuracy of 93.5%. Furthermore, we examine various cases and scenarios of label-flipping, demonstrating that compromising one client (device) in an FL framework has the least overall performance degradation on the model, emphasizing the efficacy of FL in fostering collaborative learning.