Perception Deception: Physical Adversarial Attack Challenges
summary
Demonstrates adapting digital adversarial methods to create printable physical patches that fool YOLOv3 under white‑box conditions. Uses differentiable resize and perspective transforms, Adam optimization with total variation and a non-printability score, and random vertex perturbations to tolerate placement error. Targets are object fabrication (spawn detections) and object vanishing (remove detections). Builds printed patches and monitor-based demos that make cars or stop signs disappear at varied angles/distances in controlled lighting; lighting robustness remains limited.
AI-written summary · about
chapters (22)
- 0:00Intro
- 5:31Car Safety - Unintended Acceleration
- 8:59Car Perception While Driving
- 10:35Behind Perception: End2End Object Detection
- 11:21State-of-the-Art Vision-based Object Detection
- 15:34Current Status of Adversarial Example
- 17:38Adversarial Examples & L2 Norm Perturbations Impact to DNN
- 20:36Explore Chances of Physical White Box Attack against YOLOV3
- 21:05Deep Dive into YOLOv3
- 22:03Training Dataset - MS COCO Dataset
- 22:45YOLOv3 Prediction
- 27:28Threat Model : Physical Image Patch Attack
- 28:37Our Physical Attack Approach & Objectives
- 28:51Differentiable Input Patch Construction
- 30:45Attack Objective 1 - Object Fabrication
- 33:02Attack Objective 2 - Object Vanishing
- 34:58Challenges to the Success of Physical Attack
- 36:04Tactics to the Challenges
- 37:05Color Management with Non-Printability Loss
- 40:08RT + TV for Various Distances & Angles
- 43:09Put Everything Together: An Iterative Optimization
- 47:00Conclusion & Takeaway
related talks
38mUsing adversarial samples to robustify your Neural Network Models
36mAdversarial Attacks on Deep Learning
16mPerspective-Shift Attacks Against Optical Perception Sensors: A Novel Attack...
21mAdversarial Attacks
34mHacking the Blackbox
36mAdversarial Machine Learning: How Secure Is ML?
06mWhen Collaboration Betrays: The Evolving Attack Surface of...
45mWinning by deception: tactics, techniques, and procedures of adversarial A.I.