Adversarial Machine Learning

This lecture delves into the concept of Adversarial Machine Learning, exploring how neural networks can be easily tricked by small perturbations in input data, leading to incorrect predictions. The instructor explains the importance of robustness in classification tasks, highlighting the distinction between human and neural network behavior. The lecture covers the theoretical foundations of adversarial examples, the risk associated with different norms, and the trade-off between standard and robust accuracy. Practical methods like Projected Gradient Descent (PGD) attacks and adversarial training are discussed, showcasing how training models with adversarial examples can improve robustness. The instructor also demonstrates a simple classification problem to illustrate how robust features can enhance prediction accuracy despite adversarial perturbations.