Instance attack: an explanation-based vulnerability analysis framework against DNNs for malware detection.

Area of Science:

• Cybersecurity
• Machine Learning
• Software Engineering

Background:

• Deep neural networks (DNNs) are crucial for malware detection but their robustness is a concern.
• Generating adversarial examples for DNNs typically requires model knowledge or extensive data, which are often unavailable.
• Existing methods face limitations in real-world black-box scenarios.

Purpose of the Study:

• To introduce an interpretable, instance-based attack method for DNN malware detection in black-box environments.
• To develop a novel function-preserving transformation algorithm for data subsections.
• To enhance the robustness of malware detection systems.

Main Methods:

• Developed an instance-based attack using a specific binary and malware classifier.
• Employed data augmentation to train a simple, interpretable model.
• Introduced a function-preserving transformation algorithm targeting data subsections.
• Utilized binary diversification to neutralize influential sections for adversarial example generation.

Main Results:

• The instance-based attack successfully generated adversarial examples, fooling DNNs with nearly 100% success in some cases.
• Identified data subsections as having a significant impact on malware detection.
• Demonstrated superior performance compared to state-of-the-art adversarial attack methods.
• Validated the technique's effectiveness in a black-box setting.

Conclusions:

• The proposed instance-based attack is effective and interpretable for DNN malware detection.
• The novel transformation algorithm and binary diversification enhance adversarial example generation.
• This approach offers a practical solution for improving malware detector robustness in real-world conditions.