CKKS-zkSNARKs Enhanced Federated Learning for Medical Data

Abstract

The widespread adoption of big data and AI technologies has accelerated the advancement of intelligent medical diagnostics. However, the sensitivity of medical data poses a dual challenge of privacy leakage and computational inefficiency in cross-institutional collaboration. Traditional federated learning (FL) schemes struggle to balance privacy protection, model accuracy, and communication costs, particularly for real-time processing of high-resolution medical images. To address this, we propose CZ-FLMed, a privacy-preserving FL framework integrating CKKS fully homomorphic encryption (FHE) and zkSNARKs zero-knowledge proofs. The framework employs a customized Convolutional Neural Network (CNN) for medical image training, the CKKS segmented encryption strategy for reducing communication overhead, and the lightweight Groth16 protocol for secure identity verification. This enables efficient encrypted model aggregation and authentication. The experiments results conducted in this paper on Chest X-Ray pneumonia dataset and MNIST handwritten digits demonstrate that CZ-FLMed achieves 83.05 % test accuracy in pneumonia classification. Compared to Paillier encryption, it reduces communication costs by 92.84 % and improves encryption efficiency by 404 times. Thus, the framework balances model accuracy, computational efficiency, and privacy preservation, offering a practical solution for multicenter medical collaboration.