Federated hyperdimensional ensembles for mobile malware detection

Abstract

Mobile devices face an ever-increasing threat from cyber attacks, with malware constantly evolving to circumvent traditional detection systems. Adaptive techniques attempt to address this problem by retraining models through updated data, but this approach requires sharing sensitive user information, raising significant privacy concerns. This work introduces a novel federated learning architecture for mobile malware detection to address these challenges. The proposed approach integrates the efficiency of Hyperdimensional Computing (HDC) to create an ensemble of lightweight models. The federated architecture enables collaborative model training across multiple devices without requiring users to share their private data, thus preserving privacy. The architecture supports heterogeneous clients, allowing devices with different computational capabilities to participate effectively in the training process. Experimental evaluations demonstrate the system’s effectiveness in maintaining high performance in a privacy-preserving manner, even in dynamic, resource-constrained contexts.