FLB2: Layer 2 Blockchain Implementation Scheme on Federated Learning Technique

Abstract

The usage of the federated learning (FL) concept in the artificial intelligence (AI) field has increased. The main concept of FL is to tackle the centralized-based approach, which requires the model to update training data to the cloud server by creating a decentralized deep learning (DL) model. However, the current FL model is still not completely decentralized, as each client needs to upload the training data to a centralized aggregator. Thus, this paper proposed an implementation of the FL scheme by using blockchain to tackle this problem. The proposed system uses the blockchain as the place to exchange training data instead of sending the training data immediately to the aggregator. In addition, this paper also tried to implement the layer 2 blockchain to minimize the time needed to exchange training information between each client and aggregator. The simulation result of this paper shows that we are able to implement the layer 2 blockchain in the FL system successfully. Also, it is shown that by using the layer 2 blockchain, training data exchange time is able to be reduced by around 50% compared to the layer 1 blockchain. In addition, this paper shows that the implementation of the layer 2 blockchain does not affect the performance of the FL model in terms of accuracy.