Abstract
Federated learning enables multiple Internet service providers to collaboratively train traffic classification models without sharing raw traffic traces, which is attractive for privacy and business confidentiality in edge--cloud cluster deployments. However, incentive mismatches can motivate strategic behaviors such as low-effort training, selective participation, or subtle update bias, which degrade classification accuracy, fairness, and system stability. This paper proposes a co-evolutionary mechanism design framework that jointly optimizes coordinator policies and adaptive strategic Internet service provider behaviors. A population of coordinator mechanisms, including participation control, reward and penalty rules, auditing policies, and aggregation settings, is evolved via multi-objective search to balance classification utility with communication and computational overhead. In parallel, a population of strategic behaviors is evolved and expanded using a large language model that generates realistic tactics from observed system feedback. Experiments on federated traffic classification show that, under \((p_{\mathrm{str}}=0.4)\), the proposed approach achieves a macro-F1 of \((0.876)\) and improves the worst-ISP accuracy to 82.1%, outperforming fixed-rule and robustness-only baselines while maintaining practical coordination overhead.

This publication has 18 references indexed in Scilit: