CITP: Cross-instance targeted perturbations

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Universal Adversarial Perturbations (UAPs) deceive models with a single perturbation across diverse instances.
• Existing research on UAPs focuses on non-targeted attacks, lacking transferability for specific class classifications.
• Targeted UAPs are crucial for understanding model vulnerabilities in specific classification tasks.

Purpose of the Study:

• To propose CITP, a generative adversarial framework for creating cross-instance targeted perturbations.
• To enhance the transferability of UAPs for classifying images into a specific target class.
• To extend targeted adversarial attacks beyond image data to video data.

Main Methods:

• CITP leverages shared features within target classes to generate transferable perturbations.
• The framework employs a discriminator to differentiate adversarial samples from target class images, learning the class distribution.
• A mid-level feature discriminator is integrated to boost perturbation transferability across various model architectures.

Main Results:

• CITP demonstrates exceptional transferability in cross-instance targeted attacks.
• The proposed method shows strong performance against four different defense mechanisms.
• CITP successfully enables precise targeted attacks on video data, extending its applicability.

Conclusions:

• CITP effectively generates transferable targeted UAPs, overcoming limitations of previous methods.
• The framework offers a robust approach for adversarial attacks on image and video data.
• CITP advances the understanding of model robustness and adversarial vulnerability in deep learning.