A semi-synchronous federated learning framework with chaos-based encryption for enhanced security in medical image sharing
Animesh Roy, Deva Raj Mahanta, Lipi B. Mahanta
Abstract
Protecting sensitive health information and promoting clinical research depend on medical data security. This paper suggests an innovative framework that integrates healthcare engineering, chaotic encryption, and artificial intelligence (AI) to address the privacy issue of medical data. A novel semi-synchronous, decentralized, privacy-enhancing Federated Learning (FL) model built on Convolutional Neural Networks (CNNs) is put forth. The approach integrates federated learning with chaos-based encryption, utilizing the Henon Logistic Crossed Couple Map (HLCML) to strengthen the security of hospital images stored on cloud servers. With its foundation in chaos-based approaches, the encryption algorithm is non-interactive, uses weighted parameters in each aggregation phase, and offers strong privacy protection using semi-synchronous and differential privacy techniques. Extensive simulations demonstrate the algorithm's resilience to various threats, achieving over 85% convergence in privacy-enhanced FL rounds within 100 communication rounds and delivering strong privacy protection with a noise multiplier of ϵ = 0.25 . Using MobileNetV2 CNN, the framework achieves an average accuracy of 94.3% on non-i.i.d. medical datasets. The HLCML-based encryption protects weight parameters and stops possible data leaks while lowering the computational cost to 0.0143 seconds each round. Theoretical and empirical results confirm the model's capability to enhance privacy for medical institutions and deliver strong performance in non-i.i.d. environments, marking a significant advancement in medical data security. • Proposed Semi-Synchronous model secures hospital data in Medical Image Sharing. • Federated Learning ensures safe image and weight training without data centralization. • Differential privacy in FL protects hospital data and boosts model convergence speed. • Chaos-based HLCML encryption prevents server access to model values, securing weights. • Evaluations show superior privacy and efficiency on non-i.i.d. datasets over alternatives.