Rethinking the optimization objective for transferable adversarial examples from a fuzzy perspective.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Cybersecurity

Background:

• Transfer-based attacks generate adversarial examples effective against unfamiliar models.
• Current methods face challenges in transferring minor perturbations to diverse victim models.
• Untransferable examples are characterized by a defined 'fuzzy domain'.

Purpose of the Study:

• To develop novel methods for enhancing the transferability of adversarial examples.
• To address the limitations of existing transfer-based attacks in breaching unfamiliar models.
• To improve the success rate of adversarial attacks on unseen deep learning models.

Main Methods:

• Introduced the concept of a 'fuzzy domain' to characterize untransferable adversarial examples.
• Proposed Fuzzy Optimization-based Transferable Attack (FOTA) to maximize cross-entropy loss and membership functions.
• Developed Adaptive FOTA (Ada-FOTA) for dynamic adversarial example updates to maximize transferability.

Main Results:

• FOTA improved adversarial example transferability by 5.4% on ImageNet against naturally-trained models.
• Ada-FOTA further boosted transferability by an additional 13.8% compared to existing methods.
• The proposed methods demonstrated significant improvements in attacking unseen victim models.

Conclusions:

• FOTA and Ada-FOTA effectively enhance the transferability of adversarial examples.
• These methods offer a robust solution for attacking unfamiliar models with minor perturbations.
• The fuzzy domain concept provides a new perspective for understanding and improving adversarial attacks.