Detecting Adversarial Attacks

We'll be diving into adversarial attack techniques, specifically white-box attacks. In contrast to black-box attacks, white-box techniques require full access to the model (architecture, weights, gradients, etc). To further explore the differences between white-box and black-box attacks, make sure to check out the Pentesting Fundamentals room. Additionally, if you want to learn more about AI and machine learning (ML) attacks, make sure you check out the Introduction to AI threats room. Most adversarial attacks work by adding small, crafted noise to an input to trick the model into outputting the wrong prediction or result, but in a way that is nearly identical to the expected result. We'll be looking at three examples: FGSM, BIM and PGD.

Learning Prerequisites

• Python proficiency is not necessary to complete the room, but it will allow you to explore the practical more deeply.
• Familiarity with Jupyter Notebook. (opens in new tab)

Learning Objectives

• Understand adversarial attacks in AI and Machine Learning.
• Identify adversarial attacks, specifically the Fast Gradient Sign Method (FGSM).
• Learn examples of detection mechanisms for adversarial attacks.