Abstract
The increasing prevalence of machine learning technology highlights the urgent need to delve into its insinuations for safety and confidentiality. While inquiry on the safety aspects of mechanism knowledge has garnered considerable attention, privacy considerations have often taken a backseat, although recent years have seen a significant upswing in privacy-focused research. In an effort to contribute to this growing field, we conducted an analysis encompassing more than 40 articles addressing privacy threats in the context of mechanism knowledge, published ended the historical seven centuries. We have contributed to this research by creating a thorough threat architecture and an assault taxonomy. These tools help in categorizing various attacks based on the assets they target and the knowledge adversaries possess. We also conducted an in-depth exploration of the different privacy threats posed by machine learning, shedding light on their mechanisms and implications. Furthermore, our research includes a preliminary investigation into the underlying reasons for privacy breaches in machine learning systems. This aspect delves into the root causes of privacy leaks, shedding light on the factors that make such incidents more likely to occur. In addition to identifying privacy threats and their causes, we have compiled a summary of the most commonly proposed defense mechanisms against these threats. These defences can serve as a resource for organizations and researchers seeking to bolster the privacy of their machine learning systems. Lastly, we recognize that the field of machine learning privacy still faces unanswered questions and developing difficulties. As such, we encourage further research and exploration of potential future areas of interest. By addressing these unresolved issues and embracing emerging technologies and methodologies, we can better safeguard the privacy of individuals in an increasingly data-driven and machine learning-driven world.