Microsoft has deployed their first monotonic classifier to detect malware attacks since late 2018. The attacker cannot evade the model by adding redundant features. It is inspired by the paper from Incer et al. on robust malware detection using monotonic classifier, based on the following idea: If we select features…

Gmail has improved its malicious document classifier using deep learning recently. It can now detect many insertion-only attacks that generate PDF malware variants, including adding padding to the file and adding benign textbook pages. However, I found that by making trivial deletion within malicious PDFs, one can still evade Gmail’s…

I had many SMT formulas that cannot be solved by Z3. They just hanged. But after I reformulate them as 0–1 ILP, Gurobi can efficiently solve them. SAT problem is a particular case of 0–1 Integer Linear Program (ILP). It can be formulated as a 0-1 ILP feasibility problem, with…

XGBoost can train monotonicity for user-specified features, in the Gradient Boosted Decision Trees model. I have used linear programming to verify that the trained monotonicity is sound. This is how they enforce it. Preliminary: To learn a node in a decision tree of the ensemble, the training procedure enumerates all…