EdgeSecureDP: Strengthening IoHTs Differential Privacy Through Graphvariate Skellam

The Internet of Health Things (IoHTs) has transformed healthcare systems, facilitating remote patient monitoring and personalized treatment. Federated Learning (FL) has emerged as a promising solution, enabling decentralized devices to collaboratively train machine learning models while ensuring privacy and security in healthcare applications. Differential Privacy (DP) is used to enhance privacy in FL frameworks by injecting controlled noise into data or model updates, preventing attackers from extracting specific information. However, existing DP mechanisms, such as the Gaussian mechanism and Univariate Skellam struggle to balance privacy-utility trade-off for graph-based data like drug-drug interactions. These mechanisms treat data points independently, failing to account for the complex interconnections between nodes (drugs) and edges (interactions), leaves the network vulnerable to structural attacks that can reverse-engineer relationships, thus limiting the security of collaborative drug discovery. To address these limitations, this work proposes Graphvariate Skellam, a novel DP approach that leverages graph structure information in FL settings, referred to as EdgeSecureDP. By exploiting the structural information encoded in graph edges, this method offers enhanced privacy protection. Experimental results and theoretical analysis demonstrate that Graphvariate Skellam effectively preserves privacy ( $15 \lt \epsilon \leq 20$ ) while achieving 78% accuracy in IoHT environments, making it a robust solution for privacy-preserving healthcare applications.