Generating adversarial examples without specifying a target model.

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Adversarial examples pose a significant security threat to deep learning models.
• Existing adversarial example generation methods often require extensive querying of the target model, increasing detectability, especially in black-box scenarios.

Purpose of the Study:

• To propose a novel method for generating adversarial examples that does not require querying the target model.
• To address the detectability issue associated with traditional adversarial attack methods.

Main Methods:

• Developed Attack Without a Target Model (AWTM), a Generative Adversarial Network (GAN)-based algorithm.
• AWTM generates adversarial examples without specifying or querying a target model.

Main Results:

• Achieved a maximum attack success rate of 81.78% on the MNIST dataset.
• Achieved a maximum attack success rate of 87.99% on the CIFAR-10 dataset.
• Demonstrated a low time cost due to its GAN-based architecture.

Conclusions:

• AWTM offers an effective and efficient solution for generating adversarial examples.
• The method enhances the security of deep learning models by circumventing the need for target model queries, thus reducing detectability.