DENAS: automated rule generation by knowledge extraction from neural networks

Deep neural networks (DNNs) have been widely applied in the software development process to automatically learn patterns from massive data. However, many applications still make decisions based on rules that are manually crafted and veried by domain experts due to safety or security concerns. In this paper, we aim to close the gap between DNNs and rule-based systems by automating the rule generation process via extracting knowledge from welltrained DNNs. Existing techniques with similar purposes either rely on specic DNNs input instances or use inherently unstable random sampling of the input space. Therefore, these approaches either limit the exploration area to a local decision-space of the DNNs or fail to converge to a consistent set of rules. The resulting rules thus lack representativeness and stability.

In this paper, we address the two aforementioned shortcomings by discovering a global property of the DNNs and use it to remodel the DNNs decision-boundary. We name this property as the activation probability, and show that this property is stable. With this insight, we propose an approach named DENAS including a novel rule-generation algorithm. Our proposed algorithm approximates the non-linear decision boundary of DNNs by iteratively superimposing a linearized optimization function.

We evaluate the representativeness, stability and accuracy of DENAS against ve state-of-the-art techniques (LEMNA, Gradient, IG, DeepTaylor, and DTExtract) on three software engineering and security applications: Binary analysis, PDF malware detection, and Android malware detection. Our results show that DENAS can generate more representative rules consistently in a more stable manner over other approaches. We further oer case studies that demonstrate the applications of DENAS such as debugging faults in the DNNs and generating signatures that can detect zero-day malware.