Data ID Extraction Networks for Unsupervised Class- and Classifier-Free Detection of Adversarial Examples

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep neural networks (DNNs) are powerful but vulnerable to adversarial examples.
• Adversarial attacks pose significant threats to the reliability of DNNs.
• Existing detection methods often require labeled data or knowledge of the attack.

Purpose of the Study:

• To propose an unsupervised, class- and classifier-free adversarial detection method.
• To develop a robust detector that does not require prior knowledge of adversarial examples, classes, or the original classifier.
• To enhance the security and trustworthiness of deep learning models against adversarial manipulations.

Main Methods:

• Developed an adversarial detector leveraging sample structural information, capturing residual information and variable-wise structural relationships.
• Introduced a novel attribute, data identity (ID), combining extracted residual and structural information for adversarial sample identification.
• Trained the detector using only unlabeled clean data, making it broadly applicable.

Main Results:

• The proposed method achieved state-of-the-art performance in detecting adversarial attacks on CIFAR-10 and ImageNet datasets.
• Demonstrated superior detection capabilities compared to existing adversarial detection techniques.
• Visualization experiments confirmed the effectiveness of structural information in identifying adversarial examples.

Conclusions:

• The unsupervised, class- and classifier-free approach offers a highly effective strategy for mitigating adversarial threats.
• The data identity attribute successfully distinguishes adversarial examples by analyzing their structural properties.
• This method provides a promising direction for building more resilient deep learning systems.