Building Resilient Malware Detection System against Poisoning Attacks: Effectiveness of Deep Learning Techniques

Le résumé fourni par la source

The crux of antimalware activities is malware detection, which has mostly relied on machine learning algorithms due to their good performance. Resulting from the antimalware sophistication, attackers have employed poisoning attacks for training data manipulation to prevent malware detection. Existing strategies against poisoning attacks have explored the classical machine learning algorithms, neglecting deep learning techniques, which are the state-of-the-arts. This paper focused on developing a resilient malware detection system by investigating the deep learning models’ effectiveness against grey-box poisoning attacks. The resilience of deep learning algorithms (traditional and transfer learning) against data poisoning attacks was investigated. The University of Arizona Malware Analysis logs of Windows call sequence were evaluated. Simulating various degrees of label modification under a grey-box poisoning attack scenario, the resilience of the deep learning models such as Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Bidirectional Long Short-Term Memory (Bi-LSTM), in both traditional deep learning and transfer learning modelling was evaluated using F1-score, sensitivity, specificity, recall, precision, and accuracy. Comparing the models’ performance for minimal-level, average-level and extreme-level label modification attacks, transfer learning recorded best accuracy of 0.791, 0.757 and 0.661, respectively compared to the traditional deep learning’s best accuracy of 0.519, 0.518 and 0.730. Specifically, CNN consistently outperformed LSTM and Bi-LSTM in both traditional deep learning and transfer learning evaluation. Bi-LSTM, however, recorded the highest specificity of 0.944 for the malware class in minimal-level and average-level attack models. The deep learning models also showed more resilient against poisoning attacks involving higher level of label modifications than state-of-the-art classical machine learning approach. Generally, the deep learning techniques’ performance was best when there was no data poisoning and declined as the poisoned data points increased.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.