Privacy-Preserving Stroke Prediction under Non-IID Federated Learning: Adaptive Differential Privacy and Data-Skew-Aware Aggregation

Abstract

Classification of stroke using federated healthcare data poses challenges because of privacy issues and variation in local distribution of data. FL offers collaboration of models training without the exchange of raw patient data but non-IID client data can lead to variation in sample size and class distribution. DP offers an additional privacy measure, but comes with the drawback of utility-privacy trade-off. In this research, a Data Skew-aware Adaptive Differential Privacy and Aggregation scheme is proposed for federated stroke classification. A composite Data Skew Index (DSI) that takes into account both the quantity skew and label-distribution skew is employed to derive the Gaussian DP noise and influence factor in the process of aggregation. Rather than discarding the clients with deviating updates, our approach uses soft influence adjustment, keeping all clients intact. Experiments are performed using the stroke classification data set with 12 non-IID clients and Gaussian Byzantine perturbation. Comparison is made between the proposed scheme and FedAvg, Krum, Bulyan, and Adaptive Differential Privacy through ablation study, attack sensitivity analysis, five seed study, and privacy-utility analysis. With five different seeds, the proposed scheme getsomnis iste natus error sit voluptatem accusantium doloremque laudantium