Tree-based Models' Robustness Against Evasion Attacks is Enhanced by Randomness

PROJECT TITLE :

Using Randomness to Improve Robustness of Tree-based Models Against Evasion Attacks

ABSTRACT:

Applications in the field of security have seen widespread adoption of Machine Learning models. On the other hand, it is common knowledge that adversaries can modify their attacks so as to avoid being discovered. Making Machine Learning models more resistant to these kinds of assaults has been the subject of some research and development. Randomization, on the other hand, which is a straightforward method that shows promise, is not sufficiently researched. In addition, the majority of the published works concentrate on models that include error functions that can be differentiated, whereas tree-based models, despite the fact that they lack such error functions, are quite popular because it is simple to understand them. In this paper, a novel randomization-based approach to improving the robustness of tree-based models against evasion attacks is presented as a possible solution. The method that has been suggested integrates randomization into both the time spent on model training and the time spent on model application (meaning when the model is used to detect attacks). We also apply this strategy to the random forest, which is a preexisting Machine Learning method that already incorporates randomness during the training phase but still frequently fails to produce robust models. A novel weighted-random-forest method was proposed as a means of generating more robust models, and a clustering method was suggested as a means of adding randomness at the time of model application. We also came up with a theoretical framework to provide a lower bound for the amount of work that our opponents put in. The robustness of the random-forest method is significantly improved thanks to our approach, as demonstrated by experiments on intrusion detection and spam filtering data.

Did you like this research project?

To get this research project Guidelines, Training and Code... Click Here