PRESERVING DATA PRIVACY AND PREVENTING MALICIOUS ATTACKS FOR FEDERATED LEARNING USING BLOCKCHAIN

Abstract

Federated Learning Federated Learning is currently used in many fields that require high data security such as healthcare. FL allows models to be trained directly on personal devices such as mobile phones, tablets, or IoT devices. Instead of sending all raw data to the server, only the model parameters are sent to the central server for synthesis. This helps protect the data privacy of participating machines. However, in FL, there are still many dangers when there are external attackers who want to access the original data through the sent model, or the appearance of malicious clients to reduce the performance of the model. This thesis has solved the above two problems by combining Smart Contract and RSA and AES encryption algorithms. In addition, the proposed FL system also has a mechanism to calculate contribution points and reputation points to eliminate malicious clients using the Multi-Krum algorithm. To demonstrate the effectiveness of the method, the system has installed a few malicious clients to attack the global model. After the experiment, the accuracy of the model when trained in the proposed system is always stable between 92% and 98% while the conventional FL has a very low accuracy of less than 20%.