SEGA: A Transferable Signed Ensemble Gaussian Black-Box Attack Against No-Reference Image Quality Assessment Models

Area of Science:

• Computer Vision
• Machine Learning
• Cybersecurity

Background:

• No-Reference Image Quality Assessment (NR-IQA) models are crucial for image analysis.
• Adversarial attacks reveal vulnerabilities but often lack transferability in black-box settings.
• Existing white-box attacks fail against inaccessible NR-IQA models.

Purpose of the Study:

• To address the challenge of low transferability in black-box attacks against NR-IQA models.
• To propose a novel transferable attack method for NR-IQA models.
• To improve the robustness and security analysis of NR-IQA systems.

Main Methods:

• Proposed Signed Ensemble Gaussian black-box Attack (SEGA).
• Approximated target model gradients using Gaussian smoothing and ensembling.
• Implemented a perturbation filter mask for imperceptibility.

Main Results:

• SEGA demonstrated superior transferability compared to existing methods.
• Successful black-box attacks were enabled against NR-IQA models.
• The proposed method effectively reveals NR-IQA model vulnerabilities.

Conclusions:

• SEGA is an effective solution for transferable black-box attacks on NR-IQA models.
• The findings guide the development of more robust NR-IQA systems.
• This work advances adversarial attack research in image quality assessment.