Robust Adversarial Example Detection Algorithm Based on High-Level Feature Differences.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Adversarial examples (AEs) pose a significant threat to deep learning models.
• Existing detection algorithms' accuracy is often compromised by attack characteristics and image preprocessing.
• The impact of preprocessing on detection robustness remains under-explored.

Purpose of the Study:

• To propose a novel adversarial example detection algorithm robust to both attacks and preprocessing.
• To enhance the reliability of deep learning systems against sophisticated adversarial threats.
• To address the overlooked influence of image preprocessing on detection performance.

Main Methods:

• Developed a novel detection algorithm based on high-level feature differences (HFDs).
• Employs an encoder to extract high-level features from test and counterpart training images.
• Classifies images as adversarial if feature similarity is low, exploiting semantic conflicts.

Main Results:

• The HFD-based method demonstrated significant improvements over FS, DF, and MD detection algorithms.
• Achieved detection accuracy comparable to the ESRM method.
• Exhibited superior robustness against preprocessing operations like downsampling and common corruptions.

Conclusions:

• The proposed HFD-based algorithm offers high detection accuracy across diverse attacks while maintaining resilience to preprocessing.
• The method is applicable to various target models, providing a valuable new perspective for adversarial example detection.
• This research highlights the importance of considering preprocessing in the design of robust defense mechanisms.