Adaptive network security defense method combining multi-level federated learning and adversarial training

In order to enhance the network security defense capability in distributed environments, this paper proposes an adaptive defense method that combines multi-level federated learning and adversarial training. This method has demonstrated excellent performance in multiple network attack scenarios. In DDoS attack scenarios, its accuracy reaches 96.8% and F1 score is 0.962, which is significantly better than traditional methods. When facing unknown types of attacks such as zero day attacks and new DDoS attacks, the detection rates are 83.7% and 87.6%, respectively. In addition, this method performs outstandingly in terms of communication efficiency, with a single round communication data volume of only 745.3MB for 100 participants, which is 26.1% of the traditional method. The experimental results show that this method effectively reduces communication overhead and system delay while ensuring defense accuracy, and has good robustness and scalability.