In this study, we present a novel approach to network traffic anomaly detection by integrating two key innovations into the Seq2Seq model: the use of m-sequence pseudo-random codes for word embedding and the optimization of the attention mechanism through the Boruta feature selection algorithm. Traditional methods for sequence modeling often struggle with capturing the complexity of network traffic logs, especially when dealing with variable-length sequences and diverse data types. To address these challenges, we first enhance the model’s input representation by encoding network traffic data using m-sequence codes, which improves feature embedding and model convergence. Secondly, we optimize the attention mechanism using Boruta, ensuring that the model dynamically focuses on the most relevant features for anomaly detection. Experimental results on publicly available datasets demonstrate that our proposed model outperforms baseline models, including ARIMA, GRU, and traditional Seq2Seq models, in terms of accuracy, precision, recall, and F1-score. Ablation studies further confirm the significant contributions of both innovations to the overall performance. This approach not only advances the state-of-the-art in anomaly detection but also offers practical implications for real-time network security monitoring.