Adversarial Robustness of Neural Networks

Neural networks have shown impressive performance across various tasks such as classification, segmentation, etc. They consist of interconnected sets of neurons mimicking the human brain. Neural networks are trained using a gradient descent algorithm that updates the parameters of the network. However, neural networks are brittle and are prone to adversarial attacks.

In this project, we’ll learn techniques to perform adversarial attacks on already trained neural networks. We will also visualize the input images with and without adversarial noise. We’ll train networks to identify adversarial examples. Finally, we’ll understand mitigation methods such as adversarial fine-tuning. We’ll use the PyTorch library to implement the logic for training neural networks and classifiers. We will also use Adversarial Robustness Toolbox (ART), a Python library, to perform adversarial attacks. The input images will be visualized using the Matplotlib library.