VARF: Verifying and Analyzing Robustness of Random Forests

With the large-scale application of machine learning in various fields, the security of models has attracted great attention. Recent studies have shown that tree-based models are vulnerable to adversarial examples. This problem may cause serious security risks. It is important to verify the safety of models. In this paper, we study the robustness verification problem of Random Forests (RF) which is a fundamental machine learning technique. We reduce the verification problem of an RF model into a constraint solving problem solved by modern SMT solvers. Then we present a novel method based on the minimal unsatisfiable core to explain the robustness over a sample. Furthermore, we propose an algorithm for measuring Local Robustness Feature Importance (LRFI). The LRFI builds a link between the features and the robustness. It can identify which features are more important for providing robustness of the model. We have implemented these methods into a tool named VARF. We evaluate VARF on two public datasets, demonstrating its scalability and ability to verify large models.