PSAT-GAN: Efficient Adversarial Attacks Against Holistic Scene Understanding.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Deep neural networks (DNNs) enable holistic scene understanding (HSU) with parallel tasks.
• Existing methods for multi-task HSU are task-specific and less effective.
• Generating one-to-many adversarial examples for HSU is challenging due to complex requirements.

Purpose of the Study:

• To propose PSAT-GAN, an end-to-end framework for generating parallel, heterogeneous adversarial examples for HSU.
• To address the limitations of current methods in attacking multiple HSU tasks simultaneously.
• To develop an effective one-to-many mapping for adversarial examples in HSU.

Main Methods:

• PSAT-GAN utilizes a mixture of generative models and an adversarial classifier with partial weight sharing.
• The framework learns a one-to-many mapping for parallel adversarial examples tailored to specific HSU tasks.
• Novel adversarial and soft-constraint losses are employed to generate effective perturbations.

Main Results:

• PSAT-GAN successfully generates effective adversarial examples for various HSU tasks.
• The method demonstrates efficiency in creating both universal and image-dependent adversarial examples.
• Attacks are effective under both targeted and non-targeted settings for HSU.

Conclusions:

• PSAT-GAN provides an efficient end-to-end solution for attacking multiple HSU tasks.
• The proposed framework overcomes limitations of existing adversarial attack methods for HSU.
• PSAT-GAN enhances the security analysis of deep learning models in complex visual understanding scenarios.