Robustness Verification of k-Nearest Neighbor Classifiers by Abstract Interpretation

Abstract interpretation is an established mathematical framework introduced by Cousot and Cousot in 1977 and ubiquitously used in static program analysis. In recent years, many noteworthy works have shown how abstract interpretation can be successfully applied to formally verify robustness properties of some major machine learning techniques like (deep) neural networks, decision trees and support vector machines. This research work aims to pursue this line of research by proposing a novel abstract interpretation-based framework for designing a sound abstract version of the k-Nearest Neighbors (kNN) algorithm, a well-known non-parametric supervised learning method widely used for classification and regression tasks, which is then instantiated to the standard interval domain approximating the range of numerical features, to verify its robustness and stability properties. This verification approach has been fully implemented and evaluated on several datasets, including standard benchmark datasets for individual fairness verification, and then compared with some related works finding adversarial examples on kNNs. The experimental results turned out to be very promising and showed high percentages of provable robustness and stability in most of the reference datasets, thus making a step forward in the current state-of-the-art of formal verification of machine learning models.