Research question
Can adversarial perturbations survive real-world conditions such as distance and angle?
Key findings
- Printed stickers on a stop sign caused misclassification as a speed-limit sign in 100% of lab tests and 84% of drive-by tests
Robust Physical-World Attacks on Deep Learning Visual Classification
The stop-sign sticker attack demonstrated that adversarial perturbations can be physically realised and robust to viewing conditions, the canonical example for autonomous-vehicle security.
Contribute
Spot an error, or know a record we should add? Members can propose additions and submit corrections.