Oberseminar "Mathematik des Maschinellen Lernens und Angewandte Analysis" - Lennart Siethoff

In this talk, I will present results from my Master’s thesis on adversarial training problems in classification. Focusing on total-variation-regularized formulations, we build on previous work to establish the existence of optimal classifiers for a broader class of loss functions. Using Γ-convergence, we then derive an asymptotic result describing the behavior of the adversarial training problem as the adversarial budget tends to zero.