Perception Deception - Physical Adversarial Attack Challenges

Explore the challenges and tactics of physical adversarial attacks on deep neural network-based object detection systems in this Black Hat conference talk. Dive into the fundamentals of YOLOv3, a state-of-the-art object detection method, and examine the progression from digital to physical attacks considering environmental constraints. Learn about car safety, perception while driving, and the current status of adversarial examples. Investigate the impact of L2 norm perturbations on DNNs and delve into physical white box attacks against YOLOv3. Discover techniques for object fabrication and vanishing, color management, and optimization strategies to overcome various challenges in implementing successful physical attacks. Gain valuable insights into the vulnerabilities of AI-powered object detection systems and potential countermeasures.

Intro
Car Safety - Unintended Acceleration
Car Perception While Driving
Behind Perception: End2End Object Detection
State-of-the-Art Vision-based Object Detection
Current Status of Adversarial Example
Adversarial Examples & L2 Norm Perturbations Impact to DNN
Explore Chances of Physical White Box Attack against YOLOV3
Deep Dive into YOLOv3
Training Dataset - MS COCO Dataset
YOLOv3 Prediction
Threat Model : Physical Image Patch Attack
Our Physical Attack Approach & Objectives
Differentiable Input Patch Construction
Attack Objective 1 - Object Fabrication
Attack Objective 2 - Object Vanishing
Challenges to the Success of Physical Attack
Tactics to the Challenges
Color Management with Non-Printability Loss
RT + TV for Various Distances & Angles
Put Everything Together: An Iterative Optimization
Conclusion & Takeaway