Privacy-preserving deep learning model with integer quantization and secure multi-party computation

This paper introduces QSFL (Quantization and Secure Federated Learning), a novel framework designed to enhance the privacy and security of federated learning (FL). While FL enables collaborative training without exposing raw data, it remains vulnerable to white-box inversion and inference attacks. QSFL tackles these vulnerabilities by combining integer quantization with secure multi-party computation (SMC) protocols. Integer quantization transforms floating-point model parameters into integers, thereby reducing storage demands and enhancing computational efficiency. When integrated with SMC and leveraging Elliptic Curve Cryptography (ECC), QSFL ensures robust security, even in environments where up to \(n-2\) $n-2$ of n participants may collude. The effectiveness of QSFL is demonstrated through extensive testing on datasets such as MNIST, CSIC2010, and COVID-19 chest X-rays, showing superior training speed and accuracy compared to existing SMC and differential privacy techniques. This represents a significant advancement in secure and efficient FL. The proposed framework offers a versatile solution to safeguard privacy during the training of deep learning models, making it applicable across various domains including medical, financial, recommendation systems, and decision support in operations research.