Data Privacy and Security in Artificial Intelligence Models: Emerging Challenges and Defensive Approaches

Abstract

The rapid growth of artificial intelligence (AI) and deep learning techniques leads to their extensive use in fields like audio-visual recognition and natural language processing. But there have been recently found some security threats concerning the currently existing mainstream AI models, which might prevent further development of these technologies. In this connection, the problems of data security and privacy protection have become topical for studying. The issues of data leakage and privacy are examined mainly from two points of view: data leakage via model outputs and data leakage via model updates. As for data leakage via model outputs, the article considers principles and state-of-the-art of model theft attacks, model inversion attacks, and membership inference attacks. Data leakage via model updates is considered within the framework of the research of ways to steal private data while performing distributed training. Concerning data and privacy protection, three common ways of defending are considered: model structure defense, information obfuscation defense, and query control defense. In this article, the up-to-date research in the field of data security and privacy protection in AI deep learning models is reviewed. It concentrates on the theoretical background, research results, and practical applications of data theft and defense technologies in AI deep learning models.