Introduction to Adversarial Attacks in Machine Learning - Lecture 1

Explore the fundamentals of adversarial attacks in machine learning through this introductory lecture from the University of Central Florida's CAP6412 course. Delve into real-world examples of attacks on face recognition, semantic segmentation, object detection, and 3D-printed objects. Learn essential terminology, vector operations, and norms before diving into various attack methods such as Fast Gradient Sign Method (FGSM), Momentum Iterative FGSM, Projected Gradient Descent, and Carlini and Wagner (C&W). Gain insights into DeepFool algorithms for binary and multi-class classifiers, and understand the potential vulnerabilities in AI systems across different domains.

Intro
Attacks in the Real World
Fooling Face Recognition (Impersonation)
Adversarial Attack on Semantic Segmentation
Semantic Segmentation and Object Detection
Changing facial attributes and Gender
Adversarial attack on mobile phone cameras
Attack on a 3D-printed turtle
Attack on 3D Object Detection
Project Description
Terminology
Vector operations
Norms (Unit Ball)
Fast Gradient Sign Method (FGSM)
Momentum Iterative FGSM (MI-FGSM)
Projected Gradient Descent PGD
L-BFGS (Limited memory BFGS: Broyden-Fletcher-Goldfarb-Shanno algorithm)
Carlini and Wagner (C&W)
DeepFool (Binary Affine Classifier)
DeepFool (Binary Classifier)
DeepFool (Multi-Class Classifier)
Last Two Topics
Slides Credits