Data-driven Verification of DNNs for Object Recognition
This addresses the need for robust DNN testing in safety-critical domains like autonomous systems, but it is incremental as it builds on existing testing methods.
The paper tackles the problem of verifying deep neural networks (DNNs) for object recognition by proposing a gradient-free optimization approach to find perturbation chains that falsify DNNs, demonstrating its ability to identify weaknesses in a railway track segmentation task regarding specific combinations of perturbations like rain and fog.
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.