A newly developed explainable deep residual multilayer perceptron framework detected cross-site scripting attacks with high accuracy while providing transparent justifications for each prediction, according to research published in the Journal of Transformative Technologies and Sustainable Development. The investigation directly addresses a growing vulnerability in web applications, where more than 8,500 cross-site scripting flaws were reported in 2025 alone, representing nearly 20 percent of all reported security vulnerabilities.
Although modern deep learning approaches show clear promise for identifying malicious scripts, existing detection models frequently struggle with high dimensionality, sparse inputs, severe class imbalance, and a widespread lack of interpretability. To resolve these operational challenges, researchers Aliazar Deneke Deferisha, Durga Prasad Sharma, Kibreab Adane, and Sudhir Kumar Mohapatra designed an architecture that couples automated feature selection with explainable artificial intelligence and structured data handling.
The research team evaluated their framework using a large-scale cybersecurity dataset containing 138,569 samples across 167 initial features. To streamline processing and remove noisy indicators, the authors applied a multi-stage feature selection pipeline that combined variance thresholding with an ANOVA F-test. This systematic screening step reduced the dataset to 50 informative input features before passing the data to the deep neural network.
The underlying deep residual multilayer perceptron incorporates batch normalization, skip-layer residual connections, and progressive dropout to maintain training stability across dense representations. Alongside this neural structure, the researchers systematically compared four distinct class balancing strategies: an unbalanced baseline configuration utilizing weighted loss classes, Synthetic Minority Over-sampling Technique, known as SMOTE, and two hybrid variants, SMOTE-Tomek and SMOTE-ENN.
Counter-intuitively, the unbalanced baseline with class weights achieved superior performance on key safety metrics, delivering the highest minority recall of 0.99, an overall accuracy of 0.9900, an area under the receiver operating characteristic curve of 0.9994, and only 58 false negatives. This 1.0 percent false negative rate outperformed all evaluated synthetic oversampling strategies in capturing rare malicious payloads.
Alternative balancing strategies demonstrated complementary strengths across specific operational indicators. SMOTE-Tomek recorded the highest overall accuracy at 0.9920 and the best precision for malicious payloads at 0.9938, while SMOTE-ENN produced the highest recall among the synthetic oversampling methods at 0.9811. These measured comparisons provide security practitioners with explicit trade-offs between precision and recall depending on deployment requirements.
To deliver model interpretability, the framework incorporated SHAP analysis, which identified statistical entropy, special character frequency, and the presence of JavaScript keywords as the most decisive predictive features. By generating local explanations alongside classifications, the system supplies security analysts with verifiable justifications to support active threat containment, regulatory compliance verification, and forensic investigations.
Subject of Research: Explainable deep learning framework for cross-site scripting attack detection
Article Title: XAI-Enabled Deep Residual MLP Model for Detecting Cross-Site Scripting (XSS) Attacks
Article References: Deferisha, A. D., Sharma, D. P., Adane, K., & Mohapatra, S. K. (2026). XAI-Enabled Deep Residual MLP Model for Detecting Cross-Site Scripting (XSS) Attacks. Journal of Transformative Technologies and Sustainable Development. https://doi.org/10.1007/s41314-026-00101-z
Image Credits: AI Generated
DOI: 10.1007/s41314-026-00101-z
Keywords: cross-site scripting, XSS detection, deep residual MLP, explainable artificial intelligence, SHAP analysis, class imbalance handling, SMOTE-Tomek, SMOTE-ENN, variance thresholding, ANOVA F-test, feature selection, statistical entropy, batch normalization, residual connections, progressive dropout, web application security, malicious payload detection
Cite Scienmag News
Heidi Tyler. (November 17, 2026). Class-Weighted Deep Residual Network Detects Cross-Site Scripting Web Attacks Accurately. Scienmag. https://scienmag.com/class-weighted-deep-residual-network-detects-cross-site-scripting-web-attacks-accurately/
Heidi Tyler. "Class-Weighted Deep Residual Network Detects Cross-Site Scripting Web Attacks Accurately." Scienmag, 17 November 2026, https://scienmag.com/class-weighted-deep-residual-network-detects-cross-site-scripting-web-attacks-accurately/. Accessed 17 November 2026.
Heidi Tyler. "Class-Weighted Deep Residual Network Detects Cross-Site Scripting Web Attacks Accurately." Scienmag. November 17, 2026. https://scienmag.com/class-weighted-deep-residual-network-detects-cross-site-scripting-web-attacks-accurately/








