Zero-day attack detection with a Dynamic-Weighted Contractive Autoencoder and GAN-based evaluation

Abstract

Anomaly detection, which has faced quite a challenge in zero-day attacks whose nature is novel and unpredictable, shall be addressed here. This research proposes a novel method for zero-day attacks with an adaptive loss-based Dynamic-Weighted Contractive Autoencoder (DW-CAE). The proposed method differs from the traditional autoencoder approach because it balances reconstruction and Contractive penalty and pays particular attention to features that are difficult to reconstruct. The training of DW-CAE on normal data learns invariant feature representations that enable the efficient detection of anomalies based on high reconstruction errors. The dynamic weighting mechanism further enhances the adaptive balancing of reconstruction and Contractive penalty to increase the model’s sensitivity and robustness against unseen attacks. Furthermore, we have utilized GANs to generate novel synthetic zero-day attack data for rigorous evaluation of the model. CAE and dynamic weight coordination introduce an innovative and robust model for detecting zero-day attacks. Experimental results are shown on the CICIoT2023, CICDDoS2019, ToN-IoT, and synthetic datasets, validating the performance of the proposed approach. The proposed DW-CAE demonstrates a significant performance gain over the fixed-weight CAE, achieving a significant improvement across the three benchmark datasets, highlighting its effectiveness across diverse intrusion detection scenarios.