To address the challenges of identifying unknown attack patterns and ensuring detection security in campus network anomaly traffic detection, this paper proposes FedParallel, a federated learning-based anomaly traffic detection algorithm. The algorithm employs Long Short-Term Memory networks to capture intrinsic temporal dependencies in traffic time-series data, while integrating a Parallel Transformer to model complex global interactions among data points, thereby significantly improving detection accuracy and efficiency. Furthermore, a federated learning framework is adopted to exchange anomaly scores and unclassified label information among models, enabling continuous optimization. Experimental results demonstrate the superior performance of the optimized model on the NSL-KDD and CICIDS2017 datasets: AUC-ROC scores reach 96.99% and 97.95%, while AUC-PR scores achieve 97.84% and 96.38%, respectively. Compared to three state-of-the-art models: VTT, 1DCNN-BiLSTM, and GRU-BWFA, the proposed model exhibits significant improvements across all metrics, validating its comprehensive performance advantages and its effectiveness in meeting real-world anomaly traffic detection requirements.