A Privacy-preserving and Communication-Efficient Federated Learning solution for Industrial Applications

There has been a lot of interest in privacy-preserving federated learning because of its potential to allow collaborative model training without compromising participants' privacy. When it comes to federated learning that respects users' privacy, this thesis examines a wide range of possible protection and attack tactics. First, I introduce the idea of privacy-protecting federated learning and discuss its structure, benefits, and drawbacks. Differential privacy, secure aggregation, and homomorphic encryption are only some of the defensive mechanisms I cover next to keep participants' information private. In addition, I look at the attack methods, such as membership inference and model inversion, that potentially jeopardize participants' privacy in privacy-preserving federated learning. I examine the result of model inversion attack and the measures taken to counter them. In this thesis, I consider three distinct industrial use cases from the DAIS project which will be used in real-world applications in a near future and implement a federated learning system for them while keeping in mind the need for privacy in federated learning environments. As a further step, I suggest a new client selection method based on each client's amount of data to improve the federated learning framework's accuracy and the efficacy of its communications. Also, I propose an innovative method, Parameter Randomization, to enhance the privacy and communication efficiency of federated learning systems. By introducing these two approaches, this thesis gives a thorough explanation of the field of privacy-preserving and communication-efficient federated learning and emphasizes the need for robust defense and mitigation mechanisms to protect participant privacy against attacks while keeping the accuracy of the models as high as possible.