Enhancing AI-powered Malware Detection Against Adversarial Attacks Using Architecture-aware Robustness and Explainability Analysis

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Machine-learning based malware detectors perform well on clean data but are vulnerable to feature-space and PE-level changes.This work provides an architecture-aware evaluation of Win32 and Win64 using EMBER2024 and a separate PE-level protocol.PGD achieved Neural Network attack success rates of 0.7307 and 0.7073 at perturbation budget of 0.10.50 percent adversarial training reduced the respective attack success rates to means of 0.1974 and 0.1561 with standard deviations of 0.0280 and 0.0109.Heterogeneous surrogates exhibited architecturedependent transfer, and direct score-only attacks on frozen XGBoost resulted in 0.7877 and 0.9473.The PE-level protocol used 50,000 unique files and yielded 99,933 modified-benign variants.XGBoost achieved true positive rates of 0.9737 and 0.9855 for real-malware at benign false positive rates of 0.0105 and 0.0083, respectively.Hard-negative defense reduced modified-benign alert rates by 74.5% and 85.9% on Win32 and Win64, respectively.SHAP analysis revealed stable architecture-dependent rankings.Static PE validation did not show preservation of runtime functionality.