Attack and Defense of Dynamic Analysis-Based, Adversarial Neural Malware Detection Models

Abstract

Recently researchers have proposed using deep learning-based systems for malware detection. Unfortunately, all deep learning classification systems are vulnerable to adversarial attacks where miscreants can avoid detection by the classification algorithm with very few perturbations of the input data. Previous work has studied adversarial attacks against static analysisbased malware classifiers which only classify the content of the unknown file without execution. However, since the majority of malware is either packed or encrypted, malware classification based on static analysis often fails to detect these types of files. To overcome this limitation, anti-malware companies typically perform dynamic analysis by emulating each file in the antimalware engine or performing in-depth scanning in a virtual machine. These strategies allow the analysis of the malware after unpacking or decryption. In this work, we study different strategies of crafting adversarial samples for dynamic analysis. These strategies operate on sparse, binary inputs in contrast to continuous inputs such as pixels in images. We then study the effects of two, previously proposed defensive mechanisms against crafted adversarial samples including the distillation and ensemble defenses. We also propose and evaluate the weight decay defense. Experiments show that with these three defensive strategies, the number of successfully crafted adversarial samples is reduced compared to a standard baseline system without any defenses. In particular, the ensemble defense is the most resilient to adversarial attacks. Importantly, none of the defenses significantly reduce the classification accuracy for detecting malware. Finally, we demonstrate that while adding additional hidden layers to neural models does not significantly improve the malware classification accuracy, it does significantly increase the classifier's robustness to adversarial attacks.

On Effectiveness of Adversarial Examples and Defenses for Malware Classification

“…The only defense that provided significant robustness is adversarial training. We use the saddle point formulation presented in [24] and the results look promising. It improves robustness against all attacks that adhere to our threat model.…”

“…Stokes et al [24] propose different variants for creating adversarial examples, relaxing the constraints on allowed modifications. Instead of only adding features, which preserves the functionality of the malware, the authors also allow for the removal of features.…”

“…A modification of the ensemble method is proposed by Tramèr [24] looked at ensemble learning to defend malware detectors. In our experiments we train three different models and combine them into an ensemble.…”

Malytics: A Malware Detection Scheme

“…This family sits silently, collecting information and sending it to a remote location. Its variants have a wide range of actions 10 . Calleja [48] has particular analysed on Plankton.…”

“…Although deep learning models can be this neural network, it is computationally very expensive to use backpropagation to learn a very large feature space. A common solution for this situation is to use random projection techniques [10]. The projected feature space is then fed to the deep neural network.…”