Abstract
Deep learning has emerged as a fundamental component in the advancement of contemporary Intrusion Detection Systems (IDS), facilitating the automatic recognition of intricate attack patterns within network data. Notwithstanding their superior detection capabilities, these models are becoming progressively vulnerable to hostile interference. This study assesses the susceptibility of a Deep Neural Network (DNN)-based Intrusion Detection System (IDS) to two prominent adversarial attacks: the Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD). The IDS was evaluated using the benchmark datasets NSL-KDD and CICIDS2017 under both pristine and adversarial situations to assess the decline in detection efficacy. The DNN attained impressive accuracy rates of 97.80 % and 98.10 % on the NSLKDD and CICIDS2017 datasets, respectively, under optimal input conditions. When exposed to FGSM and PGD attacks with ε=0.1, the accuracy decreased to 79.45 % (FGSM) and 75.30 % (PGD) on NSL-KDD, and to 82.65% (FGSM) and 78.40 % (PGD) on CICIDS2017. Attack success rates (ASR) attained 22.50 % for PGD and 18.35 % for FGSM on NSL-KDD, with marginally lower yet still notable ASRs recorded on CICIDS2017. Furthermore, as the perturbation budget escalated from ε=0.01 to ε=0.20, all performance metrics, including precision, recall, and F 1-score, exhibited a marked reduction, therefore substantiating a direct association between perturbation intensity and model susceptibility. These findings validate that even cutting-edge IDS models can be significantly compromised by meticulously designed hostile inputs. The results underscore the critical necessity of incorporating adversarial defense techniques into Intrusion Detection System (IDS) architecture to ensure resilience in practical cybersecurity contexts where such attacks are becoming more likely.