Deconstructing Binary Classifiers in Computer Vision

Abstract

This paper further develops the novel notion of deconstructive learning and proposes a practical model for deconstructing a broad class of binary classifiers commonly used in vision applications. Specifically, the problem studied in this paper is: Given an image-based binary classifier ${\mathbf {C}}$ as a black-box oracle, how much can we learn of its internal working by simply querying it? To formulate and answer this question computationally, we propose a novel framework that explicitly identifies and delineates the computer vision and machine learning components, and we propose an effective deconstruction algorithm for deconstructing binary classifiers with the typical two-component design that employ support vector machine or cascade of linear classifiers as their internal feature classifiers. The deconstruction algorithm simultaneously searches over a collection of candidate feature spaces by probing the spaces for the decision boundaries, using the labels provided by the given classifier. In particular, we demonstrate that it is possible to ascertain the type of kernel function used by the classifier and the number of support vectors (and the subspace spanned by the support vectors) using only image queries and ascertain the unknown feature space too. Furthermore, again using only simple image queries, we are able to completely deconstruct OpenCV’s pedestrian detector, ascertain the exact feature used, the type of classifier employed and recover the (almost) exact linear classifier.