Application of Differential Privacy-Enhanced FedAvg Algorithm in Medical Sensor Data Processing

Abstract

Federated learning is increasingly being used in computer-driven medical analysis to facilitate collaborative model training without sharing sensitive patient data. This paper discusses the privacy protection issues in distributed medical sensor networks and proposes a federated averaging algorithm that integrates adaptive differential privacy. The privacy budget can be dynamically allocated, distributing privacy resources most reasonably across rounds and devices based on the gradient norm and client participation. Experiments were conducted on the Intensive Care Unit (ICU) and benchmark Electrocardiogram (ECG) sensor datasets, simulating over a hundred heterogeneous and non-independent and identically distributed (non-IID) client scenarios. The above results indicate that under moderate privacy constraints, the proposed framework achieved a global accuracy of 89.7% on ECG data and 87.2% on ICU data, with macro F1 scores exceeding 0.92 and 0.90, respectively. Compared to fixed allocation, adaptive allocation of privacy budgets reduced the fairness gap per client by over 60%. Convergence remains stable, with only a 5% increase in computational cost. In the case of a privacy budget of 0.5, the attack success rate is close to the level of random guessing. The above results indicate that in federated learning, well-tuned differential privacy can provide robust privacy protection while maintaining high model utility. Therefore, this can be implemented in secure medical analysis across various regions of society.