Query-efficient decision-based adversarial attack with low query budget.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning Security

Background:

• Deep neural networks (DNNs) achieve high performance but are vulnerable to adversarial examples.
• Real-world model attacks are challenging due to limited access to model internals, making decision-based attacks crucial.
• Existing decision-based attacks require numerous queries, limiting their practical application.

Purpose of the Study:

• To develop a novel, efficient decision-based attack method for generating high-quality adversarial examples.
• To reduce the query budget required for successful adversarial attacks on DNNs.
• To enhance the security analysis of DNNs against adversarial manipulations.

Main Methods:

• Introduced the One Plane One Query Attack (OPOQA) method.
• Employed a strategy of generating candidate examples within a 2D plane sampled from the low-frequency discrete cosine transform (DCT) space.
• Utilized inverse DCT (iDCT) to transform samples back to pixel space, querying only one candidate per iteration to minimize query count.

Main Results:

• Achieved high attack success rates on mainstream models like Inception-v3, ResNet, and VGG on the ImageNet dataset.
• Demonstrated superior performance compared to the state-of-the-art EPCA algorithm, with attack success rates increasing by 4.5% and 11.5% at RMSE thresholds of 0.05 and 0.01, respectively.
• Successfully generated virtually imperceptible adversarial examples within a low query budget (1000 queries).

Conclusions:

• OPOQA is an effective and efficient decision-based attack for generating high-quality adversarial examples.
• The proposed method significantly reduces the query complexity for adversarial attacks, making them more practical.
• This research contributes to understanding and improving the security robustness of deep neural networks.