Towards Deep Learning for Efficient Image Encryption
Kirtee Panwar, Sonal Kukreja, Akansha Singh, Krishna Kant Singh
Abstract
Deep learning has become a topic of great concern in many domains and also in end-to-end image encryption. Traditional image encryption techniques employ rounds of diffusion and confusion to obtain an explicit trade-off between security and efficiency. With deep learning approach adequate solutions are obtained for current challenges in image encryption schemes such as efficiency, cryptographic strength, etc. In this paper, the latest trends in end-to-end encryption schemes based on deep learning are summarized. Firstly, the existing deep learning-based encryption systems are categorized into three categories: encryption with style transfer, Style transfer with enhanced diffusion properties, and Combining Deep Neural networks with chaotic systems. Each of these methodologies is discussed and relevant conclusions are made. Secondly, these methods are compared and analyzed for their achievements and drawbacks in terms of cryptographic properties of generated cipher images and quality of the recovered images. Third, the possibility of new cryptographic attacks are discussed such as Hidden Factors Leakage and Network Architecture Leakage as consequences of combination of deep learning approaches with an encryption system. Finally, conclusions are made based on comparison and analysis of deep learning approach in end-to-end encryption/decryption systems, providing basis for further research.