A Comparative Assessment of Adaptive Differential Privacy Preserving Mechanism for Secure Federated Mental Health Classification

Abstract

Mental health applications consist of sensitive and confidential information; maintaining privacy protection is a significant concern. Federated Learning (FL) is an effective privacy-preserving approach that helps to train models without exposing raw data. During processing of model updates, sensitive information may potentially be revealed; therefore, additional security protection measures are necessary. Differential Privacy (DP) is one of the privacy-preserving techniques that inject noise to the sensitive data, and it helps to reduce such risks, but there are still some disadvantages to the adaptive mechanisms. This paper thoroughly evaluated a combination of various adaptive methods that enabled DP with FL. It processes five combinations of adaptive methods, such as gradient-norm-based clipping with time-decay adaptive noise, percentile-based clipping with gradient-norm-based adaptive noise, Moving average adaptive clipping with round-based adaptive noise, layer-wise clipping with layer-wise adaptive noise, and median clipping with epoch-based noise scheduling. Using the DASS-21 mental health patient dataset, it assesses the privacy-utility trade-off through Accuracy, AUC, Precision, Recall, F1 score, Clipping behavior, and Privacy budget analysis. The experimental findings demonstrate that adaptive privacy strategies have a substantial impact on both model performance and privacy preservation; moving-average adaptive clipping and round-based adaptive noise combination achieve a more balanced result than others. These findings provide a strong foundation for developing advanced privacy-aware federated learning frameworks for highly delicate healthcare applications.