Toward a Completely Blind Attacker for No-Reference Image Quality Assessment Models.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• No-reference image quality assessment (NR-IQA) models are essential for evaluating image fidelity without pristine references.
• These models are vulnerable to adversarial attacks, which can compromise downstream vision systems.
• Existing attack methods are computationally expensive, require Mean Opinion Score (MOS) data, and lack cross-model transferability.

Purpose of the Study:

• To develop a novel, efficient, and versatile black-box attack framework for NR-IQA models.
• To overcome the limitations of existing adversarial attack methods, specifically the need for MOS labels and surrogate models.
• To enhance the adversarial robustness assessment of NR-IQA models in real-world applications.

Main Methods:

• Proposed Degrade-to-OverReconstruct (DOR), a prior knowledge-driven attack framework operating in a "completely blind" manner.
• Generated universal adversarial examples by applying mild degradation followed by aggressive over-reconstruction using a Residual Denoising Diffusion Model (RDDM).
• Disrupted intrinsic Natural Scene Statistics (NSS) adaptively, a common foundation for NR-IQA models, without relying on MOS or surrogate models.

Main Results:

• Demonstrated strong attack performance and significant prediction bias induction solely based on distortion statistics.
• Achieved superior transferability of adversarial examples across diverse NR-IQA models with varying deep neural network architectures.
• Validated effectiveness on both synthetic (LIVE, TID2013) and authentic (CLIVE) image datasets.

Conclusions:

• DOR offers a practical, MOS-free solution for adversarial robustness assessment of NR-IQA models.
• Pioneered a diffusion model-based "completely blind" attack paradigm for NR-IQA security.
• The framework enhances the security and reliability of NR-IQA models in real-world deployments.