Integration of Zero-Knowledge proofs (ZK) and Machine Learning to enhance Federated Learning Privacy and Security

Abstract

One revolutionary way to tackle privacy and security issues in federated learning (FL) is to include blockchain technology and zero-knowledge proofs (ZK) into machine learning frameworks. To strengthen FL's defences against threats such as model poisoning attacks, this work investigates the use of ZK proofs. This study presents a new technique that uses secure multi-party computation (MPC) to efficiently detect poisoned models, addressing the shortcomings of previous ZK systems. Data anonymization, encryption of sensitive information, and encoding of categorical data all contribute to the proposed model's privacy-preserving features. Adding a privacy-protecting layer is an integral part of ML model integration. ZK circuits employ ZK-SNARKs or Bulletproofs to generate proofs that the ML model may use to predict without disclosing the data. ZK-SNARKs are trusted, and request validation and data access rules control proof access.