A Federated Learning Based Intrusion Detection Method with Multi-Scale Parallel Convolution and Adaptive Soft Prediction Clustering

To address the issue of reduced intrusion detection rates caused by non-independent and identically distributed (non-IID) data in real-world network environments, this paper proposes a federated-learning-based intrusion detection method integrating multi-scale parallel convolution and adaptive soft prediction clustering. First, we use the average soft prediction vector of each client as the representation, employ Euclidean distance to measure similarity between clients, and combine the silhouette coefficient to adaptively determine the optimal number of clusters. Based on this, we implement a client clustering mechanism using soft prediction representations. Subsequently, a two-stage hierarchical aggregation strategy is employed, enabling similar clients to first reach consensus within clusters before forming a global model through inter-cluster aggregation. To effectively capture features at different scales and enhance feature representation capabilities during this process, a multi-scale parallel convolution network is employed as the local model. The multi-scale parallel layer integrates features from multiple receptive fields, thereby enhancing both local feature extraction and global feature representation. Experiments conducted on two intrusion detection datasets demonstrate that the proposed method outperforms methods such as FedAvg and FedProx in the accuracy of detection across multiple heterogeneous scenarios, thereby enhancing the model’s detection performance and generalization capabilities.