Detecting Zero-Day Web Attacks with an Ensemble of LSTM, GRU, and Stacked Autoencoders

The increasing sophistication of web-based services has intensified the risk of zero-day attacks, exposing critical vulnerabilities in user information security. Traditional detection systems often rely on labeled attack data and struggle to identify novel threats without prior knowledge. This paper introduces a novel one-class ensemble method for detecting zero-day web attacks, combining the strengths of Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and stacked autoencoders through latent representation concatenation and compression. Additionally, a structured tokenization strategy based on character-level analysis is employed to enhance input consistency and reduce feature dimensionality. The proposed method was evaluated using the CSIC 2012 dataset, achieving 97.58% accuracy, 97.52% recall, 99.76% specificity, and 99.99% precision, with a false positive rate of just 0.2%. Compared to conventional ensemble techniques like majority voting, our approach demonstrates superior anomaly detection performance by fusing diverse feature representations at the latent level rather than the output level. These results highlight the model’s effectiveness in accurately detecting unknown web attacks with low false positives, addressing major limitations of existing detection frameworks.