Blockchain Empowered Reliable Federated Learning by Worker Selection: A Trustworthy Reputation Evaluation Method

Federated learning is a distributed machine learning framework that enables distributed model training with local datasets, which can effectively protect the data privacy of workers (i.e., intelligent edge nodes). The majority of federated learning algorithms assume that the workers are trusted and voluntarily participate in the cooperative model training process. However, the situation in practical application is not consistent with this. There are many challenges such as worker selection schemes for participating workers, which hamper the widespread adoption of federated learning. The existing research about worker selection scheme focused on multi-weight subjective logic model to calculate reputation value and adopted contract theory to motivate workers, which may exist subjective judgmental factors and unfair profit distribution. To address above challenges, we calculate the reputation value by model quality parameters to evaluate the reliability of workers. Blockchain is designed to store historical reputation value that realized tamperresistance and non-repudiation. Numerical results indicate that the worker selection scheme can improve the accuracy of the model and accelerate the model convergence.

I. Introduction

With the rise of machine learning and edge devices, recent years have witnessed rapid development of Intelligent Edge Computing (IEC), where a large number of novel mobile applications integrated into our daily life, such as autonomous driving[1], intelligent diagnosis[2], smart cities[3] and so on. However, it is still faced with the dilemma of lacking enough data sources in the current practice of artificial intelligence. In this context, distributed machine learning aggregates user raw data into a parameter server for model training, but it easily leads to data privacy leakage[4] and causes excessive storage overhead. Federated learning[5] as a collaborative machine learning framework has been emerging to meets the data usage compliance, and solves the problem of data island. The distributed model training is executed by workers with local datasets, which usually adopts the gradient descent optimization algorithm[6]. In the traditional federated learning, a centralized server is required to perform the model aggregation algorithm namely parameter server. The parameter server aggregate the encrypted model parameters insteading of uploading the workers’ raw data. Federated learning is an effective way to protect data privacy and reduces privacy disclosure risk in data transmission, which is deeply integrated with the emerging technologies such as cloud computing, blockchain and intelligent edge computing.