Projected Gradient Descent Attack

Projected Gradient Descent (PGD) is an iterative optimization algorithm used to generate adversarial examples—slightly perturbed inputs that cause machine learning models to misclassify or produce erroneous outputs. Current research focuses on improving PGD's effectiveness, exploring variations like raw gradient descent and incorporating techniques such as certified radii guidance to target specific model vulnerabilities, particularly in image segmentation and time series forecasting. This work is crucial for evaluating the robustness of deep learning models across diverse applications, from autonomous driving to medical diagnosis, and for developing more resilient and trustworthy AI systems.