Daedalus: Breaking Nonmaximum Suppression in Object Detection via Adversarial Examples.

Area of Science:

• Computer Vision
• Artificial Intelligence Security
• Machine Learning Robustness

Background:

• Nonmaximum suppression (NMS) is a critical post-processing step in object detection (OD) to eliminate redundant bounding boxes.
• The security and robustness of NMS against adversarial attacks have not been thoroughly investigated.
• Vulnerabilities in OD systems can have severe consequences in safety-critical applications.

Purpose of the Study:

• To demonstrate that NMS, a standard component in OD, is vulnerable to adversarial attacks.
• To introduce a novel adversarial attack, named Daedalus, designed to disrupt NMS functionality.
• To evaluate the effectiveness and generalizability of the Daedalus attack across various OD models and applications.

Main Methods:

• The Daedalus attack manipulates detection box dimensions to bypass NMS filtering.
• Adversarial examples are crafted using an ensemble of popular OD models to enhance transferability.
• The attack is tested against real-world OD systems, including physical demonstrations with printed posters.

Main Results:

• The Daedalus attack successfully evades NMS, leading to an extreme increase in false positives (up to 99.9%).
• Mean Average Precision (mAP) scores are reduced to 0, indicating a complete failure of detection accuracy.
• The attack maintains a low distortion cost on input images and is effective even when victim model parameters are unknown.

Conclusions:

• Standard NMS is not secure against adversarial manipulation, posing a significant threat to OD systems.
• The Daedalus attack provides a practical and effective method for disrupting OD systems by exploiting NMS vulnerabilities.
• Robust adversarial example crafting strategies are essential for securing OD applications against such attacks.