Enhancing Evasive Malware Detection Using Supervised Contrastive Learning and Online Adversarial Training

Document Type

News Article

Publication Date

Fall 9-28-2026

Abstract

Machine learning-based malware detectors achieve high accuracy on static analysis of Windows Portable Executable (PE) files; however, they remain vulnerable to adversarial evasion attacks. This paper studies a white-box feature-space threat model where Projected Gradient Descent (PGD) perturbations are applied to malicious feature vectors from EMBER-2024 dataset. To enhance robustness, this work proposes a framework that integrates supervised contrastive learning (SupCon) with online adversarial training. In the proposed approach, adversarial examples are generated dynamically during training using the current model parameters and jointly optimized with clean samples via a composite loss combining clean binary crossentropy, adversarial cross-entropy, and supervised contrastive learning. Experiments on a balanced 30% subset of EMBER-2024 training set and the full test set demonstrate that the baseline detector is highly vulnerable to PGD attacks, with an evasion rate of 56.8% and robust recall of 43.2%. The proposed model reduces the evasion rate to 24.2% and increases the robust recall to 75.8%. The results demonstrate that robust malware detection requires both structured representation learning and adaptive adversarial exposure during training.

Link to Full Text

Share

COinS