Zero-Knowledge Privacy-Preserving Federated Learning for Cross-Institutional Medical Imaging Diagnostics

Abstract

With increasing dependence on AI for medical imaging diagnostics, privacy concerns and strict regulations continue to restrict data sharing across healthcare institutions. To address this, we propose a novel framework that enables cross-institutional collaboration without compromising sensitive patient information. Our system integrates federated learning with advanced privacy-preserving techniques, including homomorphic encryption, secure aggregation, differential privacy, and zero-knowledge proofs. Hospitals retain their data locally and contribute encrypted, noise-added model updates, ensuring that raw data never leaves the premises. Secure aggregation and encryption prevent any entity, including the central server, from accessing individual contributions. Differential privacy introduces mathematically bounded noise to mitigate risks from inversion and membership attacks. Meanwhile, zero-knowledge proofs allow clients to verify the legitimacy of their training process and updates without revealing internal computations or data. This layered privacy defense effectively counters gradient inversion, model poisoning, and membership inference attacks, all while maintaining strong diagnostic performance. Evaluated on real-world medical imaging datasets, our method balances accuracy with compliance to privacy laws like HIPAA and GDPR. The proposed architecture offers a scalable and trustworthy approach to enable AI-driven diagnostics across hospitals, ensuring patient confidentiality is never compromised.