Robust machine learning for safety-critical systems
Practical robust detection of railway defets usig AI
Project description
In this project, we will apply, adapt, and potentially extend state-of-the-art techniques from the area of adversarial machine learning to assess the robustness of a computer vision model used to identify railway defects. Concretely, we will collaborate with SBB, whose team has already trained a computer vision model for finding railway defects, and Siemens Mobility, whose team has extensive experience in the domain of safety certification. In the considered use case, the machine learning model takes as input an image of a rail segment and identifies the presence of defects (if any). Inspecting railways for such defects is a safety-critical task which requires that the machine learning model reliably identifies defects even in the presence of pixel noise and geometric rotations caused by the camera. To this end, we will apply widely adopted robustness specifications for deep learning models and assess the model’s robustness using state-of-the-art methods for adversarial testing, which aim to find concrete input images where the model fails, as well as techniques for robustness verification, which aim to prove the absence of mispredictions for a given robustness specification. Further, we will evaluate the effectiveness of empirical (adversarial training) and provable defenses (both training with abstract interpretation and smoothing), to assess their impact and benefit to the practical use case of SBB.