This paper focuses on the art and science of creating adversarial attacks on object detectors. While most previous research has concentrated on adversarial attacks on classifiers, this study explores the challenges of fooling object detectors, which localize objects within an image. To deceive an object detector, an adversarial example must mislead every potential bounding box (prior) in the image, making it more challenging than fooling a classifier’s single output. The researchers present a systematic investigation of adversarial attacks on state-of-the-art object detection frameworks, training patterns to suppress objectness scores produced by commonly used detectors. Their ultimate goal is to develop a wearable “invisibility” cloak that can render the wearer undetectable to object detectors. They describe their approach, which involves rendering patterns over detected persons in images and using a gradient descent algorithm to minimize objectness scores for each object prior. The paper concludes with acknowledgments to Facebook AI for their support.
| Signal | Change | 10y horizon | Driving force |
|---|---|---|---|
| Adversarial attacks on object detectors | From fooling classifiers to detectors | Improved techniques and defenses | Desire for privacy and evasion of detection |
| Wearable “invisibility” cloak | Rendering wearer imperceptible | Widespread use of cloaking technology | Desire for privacy and anonymity |
| Stylish pullover with adversarial patterns | Evading object detectors | Integration of adversarial patterns | Blending fashion with privacy |
| Delayed demo due to COVID | Temporarily postponed demo | Resumption of full-scale demos | Impact of external events on project |
| Loading images from COCO dataset | Dataset as input for detection | Advanced datasets for detection | Advancement in dataset creation |
| Gradient descent algorithm to minimize objectness scores | Optimization for object suppression | Enhanced optimization algorithms | Improvement in algorithm efficiency |
| Support from Facebook AI for the project | Collaboration and assistance | Increased collaboration in AI field | Collaborative efforts in research |