Investigating the Impact of Adversarial Evasion Attacks on Model Explanation in IoT-based DDoS Detection Systems

Abstract

Explainable Artificial Intelligence (XAI) is increasingly being used by intrusion detection systems (IDS) to enhance transparency and enable human-centered cybersecurity decision-making. However, adversarial evasion attacks present a dual threat, as they can mislead both the models themselves and their interpretability outputs intended to explain IDS model predictions. This study investigates how vulnerable DDoS detection in IoT networks is to such adversarial manipulations using the CICIoT2023 dataset. Machine learning and deep learning models, including Random Forest and LSTM, were analyzed using LIME and SHAP frameworks to generate understandable explanations after being trained to detect DDoS traffic. Adversarial perturbed examples were introduced to test the resilience of the explanation, integrity, and accuracy of prediction. The findings indicate that small perturbations can significantly reduce the accuracy of detecting attacks, resulting in false feature attributions. In adversarial settings, LSTM false negatives went up as much as 24,876 to 27,993, decreasing the accuracy from 87.13% to 86.86%, whereas random forest misclassifications went up by 363 to 12,195; however, the accuracy drop was from 99.96% to 98.95%. The overlap in top-5 feature rankings for LSTM was 50%, and that of RF was 45%, respectively. Also, the SHAP cosine similarity declined to 35% for LSTM and 55% for RF, indicating important differences in interpretability. The results of this study highlight limitations in existing explainable IDS approaches and the necessity of adversarial-resistant XAI methods to ensure reliable, trustworthy, and understandable cybersecurity analytics when using the Internet of Things.