White Box Adversarial Attacks on Malware Detection Neural Networks: A Targeted Approach

This paper delves into the realm of adversarial attacks on neural networks used in malware detection, with a specific focus on white box attacks. Unlike conventional domains, modifying executable files demands careful consideration due to the sensitivity of binary values. The proposed approach identifies editable bytes in a malicious binary, allowing for adversarial modifications without altering functionality. The study demonstrates a nearly 99% success rate in misclassifying adversarial examples. While acknowledging challenges in black box attacks, the paper suggests potential causes for non-transferability. It explores optimization using explainable machine learning techniques, reducing edited bytes while maintaining a 93% success rate.