New attack corrupts attention mechanisms in detection transformers

Researchers have developed a novel adversarial attack targeting the attention mechanisms within detection transformers, a class of models crucial for object detection in safety-critical applications. This new method, termed "Corrupting Attention," directly manipulates the encoder's attention objective, leading to a significant degradation in detection performance. Unlike previous attacks that focused on the final detection output, this approach disrupts the model's spatial reasoning by corrupting its attention, resulting in a substantial drop in mean Average Precision (mAP) on datasets like COCO and demonstrating effectiveness across different attention formulations. AI

IMPACT This research highlights a new vulnerability in object detection transformers, potentially impacting the safety and reliability of AI systems in critical applications.

RANK_REASON Academic paper detailing a new adversarial attack method on AI models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• COCO
• DETR-R50
• DINO-Swin-L
• Ridma Bashith Jayasundara Jayasundara Walpola Kankanamalage

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →