Design and analysis of explainable methodology to improve the performance of artificial intelligence model

Abstract

The increasing sophistication of cyber-attacks has underscored the critical need for newlineIntrusion Detection Systems (IDS) that are accurate but also transparent and reliable newlinein their decision-making processes. One of the key challenges in modern IDS is the newlineblack-box nature of many machine-learning models used for detection. These systems newlineoften operate without clear explanations, making it difficult for security professionals newlineto understand why a decision was made, mainly when the system flags an attack newlineor a benign activity. This lack of transparency undermines trust in the system and newlinecomplicates verifying its fairness and reliability. The proposed Hybrid-XIDS offers newlinelocal and global explanations and diverse counterfactual scenarios to enhance model newlinetransparency and facilitate 99% accurate security decisions. newlineCyber attackers frequently exploit vulnerabilities in IDS models, deploying sophisticated newlineadversarial attacks to deceive or mislead these systems. These attacks newlinecan subtly alter input data, causing the IDS to misclassify an attack as normal traffic newlineor vice versa, rendering the system ineffective. To address this, IDS must be designed newlineto withstand adversarial perturbations and continue to operate effectively even when newlinefaced with such manipulations. Integrating adversarial robustness into Hybrid-XIDS newlineensures that these systems remain resilient under real-world conditions and can accurately newlinedetect threats despite adversarial attempts to bypass detection. The Proposed newlineOptimized CandW attack consistently outperforms other methods, achieving a 48% newlineaverage accuracy reduction, up to 97% attack success rate, smaller perturbations newline(0.23-0.26 in L2 norm), faster runtime, and 92% transferability from Neural Networks newlineto Decision Trees, showcasing its robustness and efficiency. In an adversarial environment, newlinethe Hybrid-XIDS model demonstrated a 99% accuracy rate while maintaining newlinetransparency in its decision-making process. newlineWhile transparency and adversarial robustness are essential, standard metrics to newlineevaluate IDS pe