Data-driven Verification of DNNs for Object Recognition

TL;DR

This paper introduces a gradient-free optimization method for testing DNNs by identifying perturbations that reveal weaknesses, demonstrated on railway track detection under various weather conditions.

Contribution

It presents a novel gradient-free testing approach that uncovers DNN vulnerabilities to specific perturbation combinations, surpassing traditional grid-based methods.

Findings

Successfully identified DNN weaknesses with perturbation chains.

Effective in detecting vulnerabilities under weather-related image distortions.

Demonstrated on railway track detection in diverse conditions.

Abstract

The paper proposes a new testing approach for Deep Neural Networks (DNN) using gradient-free optimization to find perturbation chains that successfully falsify the tested DNN, going beyond existing grid-based or combinatorial testing. Applying it to an image segmentation task of detecting railway tracks in images, we demonstrate that the approach can successfully identify weaknesses of the tested DNN regarding particular combinations of common perturbations (e.g., rain, fog, blur, noise) on specific clusters of test images.