A comprehensive review on deep learning algorithms: Security and privacy issues

Machine Learning (ML) algorithms are used to train the machines to perform various complicated tasks that begin to modify and improve with experiences. It has become widely used for automated decisions. In particular, the applications which have a profound impact on society that rely on Deep Learning (DL) for autonomous decisions, such as Patient Health Record (PHR), Unmanned Aerial Vehicles (UAVs), etc. Such impacts have a vital concern about the potential vulnerabilities introduced by DL. Traditional attackers have powerful motives that can alter and modify DL algorithms to subvert the outcomes. In poisoning attacks, an attacker can consciously change training dataset, which is used to operate the outcomes of decision-based model. While in privacy and evasion attacks, an adversary can also misclassify new datasets to infer private information. Therefore, in this paper, we have provided a review of security and privacy issues of DL algorithms and analyzed their applications and challenges based on state-of-the-art literature. We have classified attacks, devised a taxonomy, and comprehensive analysis of defense techniques for the most common attacks such as poisoning, evasion, model extraction, and model inversion. We have also presented various privacy preserving techniques to ensure the privacy of dataset. We have proposed a secure cryptographic framework for dataset based on hash functions and Homomorphic Encryption (HE) scheme. Finally, we have provided recent research challenges and future studies concerning security and privacy issues. We believed that the highlighted limitations and weaknesses provide possible research questions and open matters for designing efficient future DL algorithms.