Decentralized federated learning for intrusion detection in 5G networks: an asynchronous consensus driven multi-agent approach

5G networks expand the wireless infrastructure attack surface while concurrently limiting data streams through in-your-face privacy regulations. Centralized traffic aggregation-based traditional intrusion-detection pipelines, thus, are plagued with bottlenecks, sole points of failure, and regulatory insurrection. This paper uses PA-CoL, a Parallel Asynchronous Consensus-based Learning framework in which multi-agent base stations train a shared deep neural model without ever exporting raw traffic records. Each agent performs local mini-batches on 5G-NIDD flow features and intermittently averages parameters with a randomly chosen neighbor; no parameter server or global synchrony is required. Experiments across three overlay graphs (Complete, Ring, Small-World), two data Types (IID and non-IID), and three federation sizes (5, 8, 10 agents) show that the newly developed scheme reaches F1 ≥ 0.99 under IID data and F1 ≥ 0.93 under severe non-IID skew. These results indicate that a lightweight, peer-to-peer consensus can deliver carrier-grade intrusion detection for privacy-sensitive 5G edge clouds and can pave the way toward 6G self-defending networks