Localized Query Attack Toward Transformer-Based Visible Object Detectors.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Transformer-based detectors excel in visible-object detection.
• Adversarial patches, a type of adversarial example, can disrupt these detectors by introducing perturbations.
• Existing methods for adversarial patches are suboptimal due to feature discrepancies and overlook cross-attention mechanisms.

Purpose of the Study:

• To introduce a novel attack method, Localized Query Attack (LQA), to effectively disrupt transformer-based object detectors.
• To address limitations of traditional methods by targeting both encoder self-attention and decoder cross-attention.
• To improve the performance of adversarial patch attacks on object detection models.

Main Methods:

• LQA targets object features by enhancing self-attention between adversarial patches and foreground regions.
• It computes a joint attention matrix in the decoder to diminish encoder outputs and residual components.
• This process amplifies the adversarial patch's relative importance, intensifying the attack.

Main Results:

• LQA demonstrated an approximate 20% improvement in transfer attack performance compared to the second-best method.
• The method showed effectiveness across various transformer-based detectors.
• Real-world scenario validations confirmed the practical efficacy of LQA.

Conclusions:

• LQA offers a more effective approach to adversarial patch attacks on transformer detectors.
• The method's ability to interfere with both self-attention and cross-attention mechanisms is key to its success.
• LQA represents a significant advancement in understanding and mitigating adversarial vulnerabilities in object detection.