Explainable Data Poison Attacks on Human Emotion Evaluation Systems based on EEG Signals

EEG-based emotion detection systems are vulnerable to data poisoning attacks: Researchers demonstrated that attackers can manipulate training data by flipping emotion labels, successfully degrading the performance of machine learning models used to evaluate human emotional states from brain signals—a concerning finding given these systems' potential use in security screening.

These attacks work across different AI models: The study found that data poisoning attacks succeed regardless of which machine learning model is used, though different models show varying levels of vulnerability, suggesting no single technical approach provides complete protection.

Explainability tools can reveal how attacks work: By applying interpretability methods like SHAP values and LIME, researchers were able to explain which specific EEG signal features attackers manipulated, making the attacks more transparent and potentially helping defenders identify suspicious patterns.

Insider threats in emotion detection systems require attention: The paper highlights that EEG-based emotion evaluation systems designed to detect internal security threats themselves face risks from malicious actors who could poison training data, creating a potential blind spot in security infrastructure.

Transparency supports defensive measures: Making attack mechanisms publicly understandable through explainable AI methods and releasing code openly enables the research community to develop better safeguards for emotion detection systems before they're widely deployed in security-critical applications.