An Efficient Approach for Security and Privacy Preserving based on Machine Learning and Federated Learning (FL): Analysis and Performance Optimization for Secure Multiparty Computing

Abstract

Federated Learning (FL) is an approach that allows numerous users to train a single machine learning model with the oversight of a central server, and with their training data stored locally on their devices. The approach is relevant in alleviating the risks associated with violations in data privacy. It is a process by which a pool of clients collaborates towards solving machine learning problems, with a central coordinator being the one who coordinates the entire process. The paper will review the latest advances in privacy-preserving federated learning and discuss it in the context of machine learning. It assesses privacy-related solutions, which are already in existence, such as secure aggregation, meta-learning, blockchain technology, decentralized training, searchable encryption, and data privacy mechanisms and zero-knowledge proofs. Federated learning (FL) is an emerging technology that can be used in the realm of the intelligence of the Internet of Things. However, the information that is model-related can be shared in FL and reveal the sensitive data of the participants. In this regard, we propose a new privacy-preserving FL framework, which is founded on a new chained secure multiparty computing technique, which we call chain-PPFL. The scheme we are proposing is based mostly on two mechanisms: 1) a single-masking mechanism, which protects the information that is exchanged between participants in a serial chain frame and 2) a chained-communication mechanism, which allows the masked information to be communicated between participants in a serial chain frame. We run large-scale experiments with respect to simulation by comparing the training accuracy and the leak defence to other state-of-the-art schemes with two publicly available data sets (MNIST and CIFAR-100). We established data sample distributions (IID and NonIID), and training models (CNN, MLP and L-BFGS) in our experiments. The experiment results show that the chain-PPFL scheme can offer a realistic privacy preservation (which is the same as the various privacy with ϵ to near zero) to FL at the cost of communication, and without compromising the accuracy and convergence rate of the training model.