Poisoning Attack Defense of Federated Emotional Generation Models for Mental Health Data Protection
Tingting Liu, Ameerah Muhsinah Binti Jamil, Ahmad Firdaus, Keke Zhao, Jinli WangAbstract
Objective
Poisoning attacks against emotion generation models in federated learning will distort output emotional data, interfere with accurate identification of users’ psychological characteristics, bring misjudgment pressure to mental health assessment, disrupt users’ objective mental state monitoring, and hinder reliable auxiliary diagnosis of public mental health. This paper proposes a targeted defense method combining differential privacy perception, to reduce psychological assessment errors caused by malicious attacks and ensure stable and accurate mental health state detection.
Subjects and Methods
This study builds a federated learning experimental platform for emotion generation models. It adopts adversarial attack simulation, differential privacy noise optimization and model anomaly detection algorithms. It compares changes of model output emotion data under attack and defense states, and analyzes the interference degree of malicious attacks on user psychological pressure identification, mental state classification and mental health evaluation results.
Results
The proposed defense method effectively resists various model poisoning attacks, restores accurate emotional feature output, eliminates data deviation-induced misjudgment of individual psychological pressure, avoids wrong identification of abnormal mental states, and greatly improves the credibility of intelligent mental health monitoring based on emotion generation models.
Conclusions
The optimized differential privacy defense scheme can enhance the security of federated learning-based emotion generation models. It avoids psychological data distortion caused by network attacks, guarantees accurate perception of users’ psychological pressure and real mental states, stabilizes the reliability of intelligent mental health analysis systems, and promotes the safe application of AI technology in digital mental health services.
Acknowledgement
Dezhou Key Laboratory of Intelligent Network Security (PT2025KJT005).
Corresponding Author
Ameerah Muhsinah Binti Jamil, Faculty of Computing, Universiti Malaysia Pahang Al-Sultan Abdullah, 26600-Pekan, Pahang, Malaysia.