Privacy Preserved Modified Consensus Mechanism for Blockchain Empowered Decentralized Federated Learning Framework in Wireless Networks

Abstract

Advancements in communication, computing and sensor technology are creating new alternative solutions to the traditional approaches of meeting the ever-growing influx amount of data point demand of 21st century and beyond. To learn and improve on these data points for model training at edge devices, federated learning, as a potential distributed machine learning approach has recently surfaced to leverage multiple nodes to take up some parts of a large training process and aggregate it on a central server. This, however, poses a serious risk towards the privacy of the exchanged data points between the server and the nodes, creating a single point of failure. Most existing work focuses on how to improve Machine Learning performance without considering security. In this work we focus on to protect the data security without much compromising on Machine Learning performance. To achieve this idea, we will first investigate the consensus mechanisms of Blockchain, followed by an overview of related work and then comparison of Blockchain based Federated Learning framework. If possible, we may also end up proposing a new consensus mechanism to achieve the privacy preserving Federated Learning framework. 1. Explain why we need to integrate the two technologies, blockchain and FL. (For this we need to investigate the benefits and use cases of blockchain in ML/AI.) 2. Illustrate the network/system architecture which achieves the blockchain-enabled FL framework. 3. Compare some existing blockchain consensus mechanisms and give their pros and cons when applying to the above blockchain-enabled FL network. 4. Discuss how the blockchain-enabled FL network can be affected under different types of attacks (e.g., participant in each learning round sends a wrong update information or send the previous value without doing any learning work but getting the benefits (i.e., learning results) from other learning agents, etc.). Conclude which consensus can perform better under different attacks. 5. Conduct simulations to compare the performance of several consensus. We should consider several types of attacks when doing simulations.