Highly Accurate Adaptive Federated Forests Based on Resistance to Adversarial Attacks in Wireless Traffic Prediction.

Abstract

Current 5G communication services have limitations, prompting the development of the Beyond 5G (B5G) network. B5G aims to extend the scope of communication to encompass land, sea, air, and space while enhancing communication intelligence and evolving into an omnipresent converged information network. This expansion demands higher standards for communication rates and intelligent processing across multiple devices. Furthermore, traffic prediction is crucial for the intelligent and efficient planning and management of communication networks, optimizing resource allocation, and enhancing network performance and communication speeds and is an important part of B5G's performance. Federated learning addresses privacy and transmission cost issues in model training, making it widely applicable in traffic prediction. However, traditional federated learning models are susceptible to adversarial attacks that can compromise model outcomes. To safeguard traffic prediction from such attacks and ensure the reliability of the prediction system, this paper introduces the Adaptive Threshold Modified Federated Forest (ATMFF). ATMFF employs adaptive threshold modification, utilizing a confusion matrix rate-based screening-weighted aggregation of weak classifiers to adjust the decision threshold. This approach enhances the accuracy of recognizing adversarial samples, thereby ensuring the reliability of the traffic prediction model. Our experiments, based on real 5G traffic data, demonstrate that ATMFF's adversarial sample recognition accuracy surpasses that of traditional multiboost models and models without adaptive threshold modified. This improvement bolsters the security and reliability of intelligent traffic classification services.