A Distributed Attack-Resistant Trust Model for Automatic Modulation Classification

Abstract

Recently, the performance of automatic modulation recognition (AMC) has been dramatically improved with the assistance of federated learning (FL). However, FL-based AMC still faces the issue of secure sharing of local model parameters, resulting in poor anti-attack capacity. Motivated by this, a Blockchain-federated learning (BFL) framework is proposed for AMC in this letter, where the AMC model is cooperatively trained by the sharing of local model parameters with Blockchain. In addition, a parameter validity evaluation method is designed therein for the aggregation process, which greatly weakens the influence of malicious nodes. On the basis of enriching training samples, the anti-attack ability of FL-based AMC schemes is significantly improved for proposed BFL framework. Simulation results show that the recognition accuracy of the proposed framework is increased by more than 10% when malicious nodes exist, on the premise of acceptable recognition accuracy.