Secure and Robust Deep Learning Using Homomorphic Encryption and Adversarial Approach for Image Classification

Abstract

The field of deep learning is facing some complex challenges when it comes to balancing sensitive data with privacy and security. With the emergence of quantum computers, encryption vulnerabilities have become a major concern. However, there is a promising solution in the form of fully homomorphic encryption (FHE) that enables encryption without decryption, creating a secure environment. To further enhance the security of deep learning models, we can employ techniques like conditional GANs. We are excited to present a novel PPDL approach for image classification that integrates FHE with adversarial learning to improve resilience. However, it is essential to note such an approach comes with a high computational cost and longer runtime. Nonetheless, it is a small price to pay for the extra layer of security it provides. . Our Research combined fully homomorphic encryption and adversarial machine learning to develop a reliable and accurate model. We protected sensitive information with CKKS encryption. The custom dataset, created with Conditional GANS, showed a 94% accuracy rate when tested with a CNN model. However, when we encrypted the model and dataset using CKKS, the accuracy dropped slightly to 92%. Our findings hold promise for future research and we are excited to share them with you.