Balancing Privacy and Performance: Exploring Encryption and Quantization in Content-Based Image Retrieval Systems

This paper presents three significant contributions to the field of privacy-preserving Content-Based Image Retrieval (CBIR) systems for medical imaging. First, we introduce a novel framework that integrates VGG-16 Convolutional Neural Network with a multi-tiered encryption scheme specifically designed for medical image security. Second, we propose an innovative approach to model optimization through three distinct quantization methods (max, 99% percentile, and KL divergence), which significantly reduces computational overhead while maintaining retrieval accuracy. Third, we provide comprehensive empirical evidence demonstrating the framework's effectiveness across multiple medical imaging modalities, achieving 94.6% accuracy with 99% percentile quantization while maintaining privacy through encryption. Our experimental results, conducted on a dataset of 1,200 medical images across three anatomical categories (lung, brain, and bone), show that our approach successfully balances the competing demands of privacy preservation, computational efficiency, and retrieval accuracy. This work represents a significant advancement in making secure CBIR systems practically deployable in resource-constrained healthcare environments.