Minimum Adversarial Examples.

Area of Science:

• Information Security
• Deep Learning
• Computer Vision

Background:

• Deep neural networks are vulnerable to adversarial examples (AEs).
• Existing AE generation methods face challenges in defining minimum perturbations and their reachability.
• Previous optimization models lack complete solutions for these fundamental AE problems.

Purpose of the Study:

• To define minimum adversarial examples (AEs) and establish their theoretical lower bound.
• To address the computational inaccessibility of generating minimum AEs by proposing a new optimization model.
• To develop methods for generating controllable AEs under various constraints.

Main Methods:

• Proposed a definition for minimum AEs and derived their theoretical lower bound.
• Proved that generating minimum AEs is an NPC problem.
• Established a new, generalizable optimization model adaptable to any constraint.
• Devised specific methods for controllable AE generation using Lp and SSIM constraints.

Main Results:

• The new model effectively limits AE amplitude and reduces search costs, enhancing efficiency.
• AEs generated by the new model are closer to the minimum adversarial boundary, improving attack success rates.
• The model demonstrates superior attack capabilities, achieving approximately 10% higher success rates compared to baseline methods.
• Controllable AE generation is achieved under different constraints, suiting diverse applications.

Conclusions:

• The proposed optimization model offers a robust solution for generating minimum adversarial examples.
• The new approach overcomes limitations of existing methods, providing more effective and controllable adversarial attacks.
• This research significantly advances the understanding and generation of adversarial examples in information security.