On the robustness of skeleton detection against adversarial attacks

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Human perception of object skeletons is robust to shape changes, crucial for object recognition.
• Deep learning models for skeleton detection exist, but their vulnerability to adversarial attacks is unknown.

Purpose of the Study:

• This study is the first to investigate the robustness of deep learning skeleton detection against adversarial attacks.
• To analyze the vulnerability of skeleton detection models to imperceptible data perturbations.

Main Methods:

• Systematic analysis using comprehensive adversarial attacking experiments.
• Proposal of a novel Frequency attack exploiting regular, interpretable perturbations.
• The Frequency attack uses excitatory-inhibition waveforms to confuse edge-sensitive filters.

Main Results:

• Skeleton detection models are demonstrably vulnerable to adversarial attacks.
• The proposed Frequency attack effectively disrupts model performance.
• Adversarial perturbations, even imperceptible ones, significantly impact skeleton detection.

Conclusions:

• Current deep learning-based skeleton detection models are susceptible to adversarial manipulation.
• Findings necessitate the development of more robust models incorporating explicit skeleton features.
• This research opens avenues for creating more resilient computer vision systems.