Data-Driven Suitability Analysis to Enable Machine Learning Explainability and Security

Abstract

This work posits that suitability analyses that pair machine learning best practices with domain knowledge mapped to statistically significant metrics can determine boundaries for responsible data usage in machine learning models. A suitability analysis was described and tested for a malware analysis model called @DisCo and was tested across three datasets. This analysis predicted @DisCo's ability to correctly dissect data and come to reasonable conclusions. The suitability analysis correctly identified the acceptability of each dataset based on the structure of the data alongside knowledge of how @DisCo was trained and underlying domain knowledge. This process is repeatable and automate-able such that data that is not fit for @DisCo can be blocked and inaccurate results would be replaced with an explanation of why data is not fit for the model.1