Deep learning models security: A systematic review

Deep learning models and the digital records they generate have remarkably increased their adoption of many practical applications. While the success of deep learning in multimedia applications, especially images, helps tackle some of the most challenging problems, one of its copyright violations, ownership conflict, poses a grave concern for many potential applications. Many works on intellectual property protection for such models have proposed to verify ownership. Therefore, it is necessary to conduct a comprehensive study on the security of deep learning models to evaluate their strong background, state-of-the-art solutions, possible attacks, current limitations and notable improvements. This survey attempts to systematically discuss and summarise the recent advanced security solutions for deep learning models through watermarking, encryption and fingerprinting. Our study explores the recent applications, possible attacks, current limitations and notable suggestions regarding deep learning. It also comprehensively evaluates the recent research gaps and opportunities in detail to empower researchers and practitioners to provide additional secure solutions for deep learning models. This extensive survey is the first to consider model security through several notable techniques.