Explainable AI and Random Forest based reliable intrusion detection system

Abstract

Emerging cyber threats — particularly adversarial attacks on machine learning-based Intrusion Detection Systems (IDS) — pose critical risks to network security by exploiting model vulnerabilities and training blind spots. These attacks, often carried out under black-box threat models, involve crafting perturbations that force misclassification without direct access to model parameters, making them especially dangerous in real-world deployments. Traditional IDS models remain ill-equipped to handle such scenarios, relying heavily on adversarial retraining, which is computationally expensive and limited to known attack patterns. To address these challenges, we propose a novel IDS framework that enhances adversarial resilience without retraining by integrating Explainable AI (XAI)-driven credibility assessment with a dual-layered defense pipeline. At its core is a Credibility Assessment Module (CAM) that leverages SHAP (Shapley Additive Explanations) to identify inconsistencies between local and global feature attributions, flagging suspicious predictions for reassessment. The secondary pipeline employs Transformer-based semantic payload inspection alongside behavioral classifiers operating on contextual features, ensuring modal and architectural separation to prevent adversarial transferability. These capabilities enable the system to counter a wide spectrum of threats, ranging from traditional attacks to advanced black-box adversarial techniques such as HopSkipJump and ZOO, which craft minimal perturbations to evade detection. The proposed system is evaluated on two comprehensive and diverse datasets: CSE-CIC IDS 2018, which captures modern attack vectors such as SSH brute force, DoS, and DDoS; and CIC-IoT 23, which focuses on IoT-specific traffic and threats. These datasets were chosen for their realism, broad protocol coverage, and relevance to both conventional and emerging network environments. Our framework outperforms state-of-the-art adversarial defenses and multimodal IDS models, maintaining high accuracy under clean conditions while significantly improving resilience against black-box adversarial attacks. This work introduces a new paradigm in trustworthy IDS design, where explainability and processing diversity form the backbone of proactive, resilient cybersecurity.