MalFuzz: Coverage-guided fuzzing on deep learning-based malware classification model.

With the continuous development of deep learning, more and more domains use deep learning technique to solve key problems. The security issues of deep learning models have also received more and more attention. Nowadays, malware has become a huge security threat in cyberspace. Traditional signature-based malware detection methods are not adaptable to the current large-scale malware detection. Thus many deep learning-based malware detection models are widely used in real malware detection scenarios. Therefore, we need to secure the deep learning-based malware detection models. However, model testing currently focuses on image and natural language processing models. There is no related work to test deep learning-based malware detection models specifically. Therefore, to fill this gap, we propose MalFuzz. MalFuzz uses the idea of coverage-guided fuzzing to test deep learning-based malware detection models. To solve the model state representation problem, MalFuzz uses the first and last layer neuron values to approximately represent the model state. To solve the new coverage calculation problem, MalFuzz uses the fast approximate nearest neighbor algorithm to compute the new coverage. The mutation strategy and seed selection strategy in image model or natural language processing model testing is not appropriate in deep learning-based malware detection model testing. Hence MalFuzz designs the seed selection strategy and seed mutation strategy for malware detection model testing. We performed extensive experiments to demonstrate the effectiveness of MalFuzz. Based on MalConv, Convnet, and CNN 2-d, we compared the modified TensorFuzz and MAB-malware with MalFuzz. Experiment results show that MalFuzz can detect more model classification errors. Likewise, the mutation operation of MalFuzz can retain the original functionality of malware with high probability. Moreover, the seed selection strategy of MalFuzz can help us explore the model state space quickly.

Abstract licence: CC BY

52 more references, whose wording the paper’s licence keeps private.

A person counts once a day per paper. We keep no cookie and no address: a reader is a hash that changes every day. Any month under five reads shows no number.