Achieving broad generalization in healthcare AI necessitates multi-institutional data, but strict data privacy laws keep clinical records confined within hospitals. Federated Learning (FL) enables collaboration without sharing raw data, offering a path to secure, scalable clinical prediction. Recent challenges include the growing exposure of electronic health records to adversarial manipulation, and the presence of non-IID client data with uneven sample sizes that degrade global model performance. Many current FL methods still use uniform or fixed aggregation rules that ignore client quality, making them sensitive to data imbalance and attack-induced noise. We present an attack-aware FL pipeline that starts with hospital side preprocessing and Deep Particle Swarm Optimization (DPSO) feature selection to obtain a compact, discriminative attribute set. Clients train a temporal Deep Convolutional Neural Network (DCNN) model round wise, logging per-round accuracy and effective sample count; robustness is probed with Projected Gradient Descent (PGD) adversarial perturbations. The server performs dynamic aggregation using a normalized (accuracy × samples) client score to weight each parameter tensor during averaging. Updated global weights are redistributed to hospitals and the cycle repeats to convergence, yielding a privacy preserving, quality aware global predictor. The method is evaluated on the UC Irvine Heart Disease dataset under two settings: the original clean data and an adversarially perturbed variant. It attains 94.44 % global accuracy on the UCL Heart Disease dataset and 97.36 % on the UCL Breast Canser dataset. The proposed approach facilitates a lightweight, edge-deployable intrusion detection system for healthcare, providing real-time and privacy-preserving protection through federated learning.
Naeem et al. (Sun,) studied this question.