Evading Deep Neural Network and Random Forest Classifiers by Generating Adversarial Samples

Abstract

With recent advancements in computing technology, machine learning and neural networks are becoming more wide-spread in different applications or software, such as intrusion detection applications, antivirus software and so on. Therefore, data safety and privacy protection are increasingly reliant on these models. Deep Neural Networks (DNN) and Random Forests (RF) are two of the most widely-used, accurate classifiers which have been applied to malware detection. Although their effectiveness has been promising, the recent adversarial machine learning research raises the concerns on their robustness and resilience against being attacked or poisoned by adversarial samples. In this particular research, we evaluate the performance of two adversarial sample generation algorithms - Jacobian-based Saliency Map Attack (JSMA) and Fast Gradient Sign Method (FGSM) on poisoning the deep neural networks and random forests models for function call graph based malware detection. The returned results show that FGSM and JSMA gained high success rates by modifying the samples to pass through the trained DNN and RF models.