Key points are not available for this paper at this time.
Privacy-Preserving Machine Learning is a method for preventing data leakage in machine learning algorithms. This method provides different techniques to train ML model collaboratively without revealing private information. It involves protecting against malicious attacks aimed at obtaining confidential information and causing breaches in data. Unlike regular ML, FL trains locally and does not send data to a server. This review paper examines the privacy risk associated with both traditional ML and FL. We categorize the attacks based on data, model and communication vectors, analyze the effects in privacy sensitive domains such as healthcare. Additionally, it also evaluates advanced defence strategies including Homomorphic Encryption, Differential Privacy, Secure Multi-Party Computation, and anonymization, discussing their effectiveness, scalability and trade offs between privacy and model utility. Future research directions include integrating adaptive privacy methods, using explainable AI (XAI) to improve model transparency, and creating strong privacy frameworks for edge computing and IoT applications. Finally, this paper gives an in-depth look at how privacy problems might be addressed in ML and FL, assuring the ethical use of these technologies in real-world applications.
Sasirekha et al. (Mon,) studied this question.