Adversarial_Malware_Detection

Modern cybersecurity relies increasingly on automated machine learning classifiers to safeguard systems against malicious software. Traditional signature-based detection mechanisms fail to identify zero-day and obfuscated threats, prompting a shift toward deep learning architectures. Visualizing raw binary executables as two-dimensional grayscale images allows Convolutional Neural Networks (CNNs) to extract rich spatial texture patterns without executing untrusted code. However, neural network models exhibit critical vulnerabilities when subjected to targeted adversarial perturbations. This paper presents an end-to-end framework for binary malware image classification and systematically evaluates model robustness against white-box gradient-based attacks. We construct a CNN classifier trained on converted binary image representations and integrate the Fast Gradient Sign Method (FGSM) using the Adversarial Robustness Toolbox (ART). Empirical evaluation demonstrates that while the proposed CNN achieves an optimal baseline classification accuracy of 96.2% on unperturbed data, its performance degrades down to 18.7% under an FGSM perturbation factor of ε = 0.30. These findings highlight the necessity of incorporating adversarial evaluation and defensive training mechanisms into production-level machine learning security infrastructure.