AIIT: An adjustable integration adversarial attack based on image transformation.

Area of Science:

• Artificial Intelligence
• Machine Learning Security

Background:

• Deep neural networks (DNNs) are vulnerable to adversarial attacks, posing risks in critical applications like autonomous driving.
• Current black-box attacks using model integration often average surrogate model outputs, limiting adversarial example diversity and attack success.

Purpose of the Study:

• To propose a novel hybrid method, AIIT (Artificial Intelligence Integration and Transformation), to overcome limitations in existing black-box adversarial attacks.
• To enhance the diversity and transferability of adversarial examples against DNNs.

Main Methods:

• Developed AIIT, a hybrid approach combining model integration with input transformation techniques.
• Incorporated dynamic gradient adjustment for improved model integration.
• Introduced a gradient optimal algorithm to mitigate surrogate model overfitting.

Main Results:

• AIIT improved attack success rates by 13% to 35% compared to existing methods.
• Achieved an average attack success rate of 90% across various datasets.
• Demonstrated enhanced transferability of adversarial examples to different model architectures.

Conclusions:

• The proposed AIIT method effectively enhances black-box adversarial attacks on DNNs.
• AIIT addresses the limitations of gradient averaging and improves adversarial example diversity and transferability.
• This research contributes to understanding and improving adversarial attack methodologies in AI security.