Local imperceptible adversarial attacks against human pose estimation networks.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning Security

Background:

• Deep neural networks (DNNs) are susceptible to adversarial attacks.
• Research on adversarial attacks targeting human pose estimation (HPE), specifically body joint detection, remains limited.
• Adapting classification-based attacks to body joint regression is challenging, and attack effectiveness often conflicts with imperceptibility.

Purpose of the Study:

• To propose novel local imperceptible attack methods for HPE networks.
• To address the challenge of balancing attack effectiveness and visual imperceptibility in HPE.
• To develop a method for crafting perceptual adversarial attacks tailored for body joint regression tasks.

Main Methods:

• Reformulating imperceptible attacks on body joint regression as a constrained maximum allowable attack problem.
• Employing iterative gradient-based strength refinement and greedy-based pixel selection for approximate solutions.
• Developing a method that optimizes for both human perception and attack efficacy.

Main Results:

• Demonstrated the effectiveness of local imperceptible attacks against state-of-the-art HPE models like HigherHRNet, DEKR, and ViTPose.
• Achieved high attack effectiveness with a significant reduction in perturbed pixels, often requiring only around 4% of pixels.
• Showcased excellent imperceptibility of the crafted adversarial perturbations.

Conclusions:

• The proposed method successfully generates effective and imperceptible adversarial attacks for HPE.
• The approach offers a viable solution for securing HPE systems against adversarial manipulation.
• Minimal pixel perturbations are sufficient to compromise HPE models, highlighting the vulnerability of current systems.