UAV network traffic data are typically stored in a distributed manner, which poses challenges for collaborative intrusion detection under privacy constraints. To address this challenge, we propose a federated intrusion detection method that integrates a Transformer with Row-momentum normalized preconditioning (RMNP). During local training at each client, the method employs a Transformer to model network traffic features and uses self-attention to capture global correlations among different features. RMNP is further applied to optimize two-dimensional weight matrices, improving the convergence of local model training. At the server, an adaptive aggregation strategy determines the aggregation weights based on client sample sizes, label-distribution balance, and local training loss, reducing the influence of variations in local model quality on global model training. Experimental results on the UAVIDS-2025 dataset show that the proposed method achieves an accuracy of 94.78% and a Macro-F1 score of 95.07%.

intrusion detection,federated learning,Transformer,RMNP optimizer