PART 6 · MODULE 4Hiding from AI
Adversarial Object Detection
Evading object detection systems
Loading module...
Evading object detection systems
Object detection models like YOLO and Faster R-CNN scan images to find and label every object. They power self-driving cars, security cameras, and warehouse robots. But adversarial attacks can make these systems blind to specific objects.
Instead of modifying the entire image, attackers can place a small adversarial patch— a specially designed pattern — on or near an object. The patch confuses the detector so much that it simply does not see the object at all.
Object Detection with Adversarial Patch
These attacks are not limited to digital images. Researchers have demonstrated attacks that work in the real, physical world.
Printing specific patterns on T-shirts or jackets that make the wearer invisible to person-detection cameras. Walk past a security system completely undetected.
Small stickers placed on stop signs that cause self-driving car vision systems to misread them as speed limit signs or ignore them entirely.
Specially printed eyeglass frames that fool facial recognition systems into misidentifying the wearer as a different person.
When object detection fails in an autonomous system, the consequences are not just wrong labels on a screen. A self-driving car that cannot see a pedestrian, a drone that cannot detect an obstacle, or a security camera that misses an intruder — these are life-or-death scenarios.
The physical-world viability of these attacks means that any system relying on computer vision for safety-critical decisions must be hardened against adversarial inputs.
If AI cannot reliably detect objects, autonomous systems become dangerous. Adversarial patches that work in the physical world mean this is not just a theoretical concern — it is a practical threat to any system that relies on computer vision.
For hands-on practice, check out the AI Security course on GitHub.