Explainable Label-flipping Attacks on Human Emotion Assessment System

Emotion recognition systems using brain signals are vulnerable to data poisoning attacks: The research demonstrates that machine learning models trained on EEG (brain signal) data to classify human emotions can be successfully attacked by flipping labels during training, regardless of which classifier is used—a critical finding for systems that might be deployed in healthcare, security, or human-computer interaction applications.

Different AI models have varying levels of robustness: The study found that while both AdaBoost and Random Forest classifiers were susceptible to label-flipping attacks, they showed "different degrees of resistance," suggesting that model choice matters when building systems that need to withstand adversarial manipulation of training data.

Explainability tools can help identify poisoning attacks: By applying Explainable AI (XAI) techniques to analyze the attacks, researchers were able to understand how and why the poisoned models made incorrect emotion classifications, which could help developers detect when training data has been compromised.

Biometric emotion systems require stronger security measures: Since EEG-based emotion assessment could be used in sensitive contexts (medical diagnosis, security screening, or user authentication), this research highlights the need for robust defenses against data poisoning—such as data validation protocols or anomaly detection during the training phase.