The rapid growth of digital technologies has intensified the complexity and frequency of cyber threats. Hence, the traditional intrusion detection systems, that uses static rules, faces challenges to detect and mitigate these attacks. To address these challenges, this paper presents an Autoencoder-based classifier and Bidirectional Generative Adversarial Network (BiGAN) for resilient intrusion detection. The novel approach used in this paper combines an Autoencoder with stacked LSTM layers and attention mechanisms to learn patterns in network traffic. It also uses skip connections to capture time-based dependencies and custom loss function with regularization to make the model more robust. The BiGAN enables unsupervised anomaly detection through bidirectional data-latent space mapping, demonstrating its resilience to adversarial attacks. The experimental evaluations demonstrate that both models deliver competitive detection accuracy, thereby reducing reliance on labelled data. The obtained results from the benchmark datasets highlight hybrid deep learning’s potential for intrusion detection.