Detection of Attacks with an Adversarial Machine Learning Approach

The increasing sophistication of cyber threats necessitates the development of robust and intelligent intrusion detection systems (IDS) capable of detecting evolving attack patterns. Traditional IDS rely on signature-based and heuristic methods, which struggle to adapt to novel threats. Recent advancements in deep learning (DL) have enhanced IDS capabilities; however, these models remain vulnerable to adversarial attacks, particularly Trojan backdoors, which can compromise detection accuracy. A significant research gap exists in evaluating IDS robustness when subjected to adversarial manipulation. This work aims to address this gap by embedding a Trojan backdoor into an IDS model to assess its vulnerabilities and resilience. We propose a novel attack detection framework leveraging Deep Neural Networks (DNN) and Long Short-Term Memory (LSTM) networks to classify network traffic anomalies. Additionally, we employ Generative Adversarial Networks (GAN) to generate attack packets for evaluating IDS robustness. Our experimental results demonstrate that our enhanced IDS model achieves 99% accuracy in multi-class classification and 97% accuracy in binary classification, representing a 52.18% and 3% improvement compared to baseline models.