Ensemblefedshield-Pruda: A Privacy-Preserving Adaptive Federated Learning Framework with Relevance- and Uncertainty-Guided Dynamic Update Attenuation for Mental Health Risk Prediction

Abstract

Federated Learning (FL) is becoming more popular for mental health risk prediction because it enables collaborative model training while keeping sensitive patient data localised and avoiding direct data exchange. Yet the distribution of model updates can still lead to the leakage of private information via gradient leakage, membership inference, and model inversion attacks. On the other hand, traditional Differential Privacy (DP) approaches generally result in a considerable loss of util-ity because of the noise they add and the use of fixed clipping methods. In order to overcome these disadvantages, A Privacy-Preserving Adaptive Federated Learning Framework with Relevance and Uncertainty-Guided Dynamic Update Attenuation (ENSEMBLEFEDSHIELD-PRUDA) is a frame-work that has been developed for safe and secure mental health risk prediction and offers protection against various threats. This framework manages the training of a Multi-Layer Perceptron (MLP) by means of moving-average adaptive clipping, gradient-based feature-relevance scoring, Monte Carlo dropout for uncertainty estimation, uncertainty-guided dynamic attenuation, and round-based adaptive differential privacy. These mechanisms ensure that the updates are protected before the secure global aggregation takes place, preserve useful model parameters, reduce unreliable client updates, and stabilise local optimisation. The effective performance of the proposed framework is further verified by extensive comparative experiments with five baseline models, namely Central-ised MLP, Standard Federated Learning (FedAvg), and Federated Learning with Fixed Differential Privacy, Federated Learning with Adaptive Clipping and Fixed DP, FedAvg with Uncertainty Guided DP, all of which show consistent improvements in predictive performance while still lacking strong privacy guarantees. As a result, the proposed framework attains a better balance between privacy, utility and enhances its resistance to information leakage. Finally, the experimental results indicate that the proposed framework (ENSEMBLEFEDSHIELD-PRUDA) offers robust privacy preservation and reliable mental health risk prediction, which means it is suitable for secure collab-orative healthcare analytics.