Hiding Objects from Computer Vision by Exploiting Correlation Biases
summary
Automated composites exploit an implicit correlation bias in vision models trained on COCO 2017 to hide objects: round shapes near dogs become frisbees, stop signs pasted on fruit or low-correlating scenes drop detection. A pipeline using segmentation and ensembles (ResNet50, RetinaNet, YOLOv3, TinyYOLOv3) generates 1,000 adversarial images; commercial cloud APIs detect only ~10.2% and ~36.8% above 50% confidence. A printed stop sign is unseen by 4 of 5 networks, and adding 200 adversarial samples to training restores detection to ~98% with negligible MAP loss (57.8→57.7).
AI-written summary · about
chapters (5)
- 0:00Start
- 5:34Random Image Manipulation
- 10:12Automatic Adversarial Generation
- 20:42Testing Against Physical Objects
- 25:17Benefits of Correlation Based Attacks
related talks
23mBuilding high-quality computer vision models using only a few examples
25mComputer Vision in Melanoma Detection
12mCascade models in computer vision to boost accuracy and performance
55mPerception Deception: Physical Adversarial Attack Challenges
48mStay Ahead of Adversarial AI in OT/ICS Environments – Mitigating CWE-1039
36mAdversarial Machine Learning: How Secure Is ML?
38mUsing adversarial samples to robustify your Neural Network Models
44mComputers are Stupid: Protecting "AI" from Itself