Investigation of robustness of b-Tagging algorithms for the CMS Experiment

With great success, deep learning as one form of machine learning is utilized for various ap- plications and shows its benefits also in the field of high-energy physics, or more specifically, for jet flavour tagging. At the same time, studies on AI safety show the susceptibility of neural networks, even if the modifications of the inputs are only marginal. Those intriguing, yet con- cerning properties are analysed in the context of b-tagging, where the chosen architecture replicates the DeepCSV algorithm, which is utilized at the CMS experiment. Investigating the response of the tagger is motivated by the later usage of the outputs in physics analyses, as the values for simulation and data are deviating. The vulnerability of the neural network will be probed by employing adversarial attacks, which involves the addition of a random Gaussian noise term or the systematic application of the Fast Gradient Sign Method. Investigations of the input shapes as well as the ROC curves show how the aforementioned adjustments influence the features and the performance. The scan of the area under the ROC curve over epoch reveals the tradeoff between model per- formance and susceptibility. A defense strategy, namely adversarial training, improves the robustness of the classifier and represents a better tradeoff, compared to the basic training. This method constitutes a promising candidate to reduce discrepancies when evaluating the model on simulated events as well as on data.