Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Stochastic BIQA: Median Randomized Smoothing for Certified Blind Image Quality Assessment

Ekaterina Shumitskaya, Mikhail Pautov, Dmitriy Vatolin, Anastasia Antsiferova

TL;DR

This work focuses on developing a provably robust no-reference IQA metric, based on Median Smoothing combined with an additional convolution denoiser with ranking loss to improve the SROCC and PLCC scores of the defended IQA metric.

Abstract

Most modern No-Reference Image-Quality Assessment (NR-IQA) metrics are based on neural networks vulnerable to adversarial attacks. Attacks on such metrics lead to incorrect image/video quality predictions, which poses significant risks, especially in public benchmarks. Developers of image processing algorithms may unfairly increase the score of a target IQA metric without improving the actual quality of the adversarial image. Although some empirical defenses for IQA metrics were proposed, they do not provide theoretical guarantees and may be vulnerable to adaptive attacks. This work focuses on developing a provably robust no-reference IQA metric. Our method is based on Median Smoothing (MS) combined with an additional convolution denoiser with ranking loss to improve the SROCC and PLCC scores of the defended IQA metric. Compared with two prior methods on three datasets, our method exhibited superior SROCC and PLCC scores while maintaining comparable certified guarantees.

Stochastic BIQA: Median Randomized Smoothing for Certified Blind Image Quality Assessment

TL;DR

This work focuses on developing a provably robust no-reference IQA metric, based on Median Smoothing combined with an additional convolution denoiser with ranking loss to improve the SROCC and PLCC scores of the defended IQA metric.

Abstract

Most modern No-Reference Image-Quality Assessment (NR-IQA) metrics are based on neural networks vulnerable to adversarial attacks. Attacks on such metrics lead to incorrect image/video quality predictions, which poses significant risks, especially in public benchmarks. Developers of image processing algorithms may unfairly increase the score of a target IQA metric without improving the actual quality of the adversarial image. Although some empirical defenses for IQA metrics were proposed, they do not provide theoretical guarantees and may be vulnerable to adaptive attacks. This work focuses on developing a provably robust no-reference IQA metric. Our method is based on Median Smoothing (MS) combined with an additional convolution denoiser with ranking loss to improve the SROCC and PLCC scores of the defended IQA metric. Compared with two prior methods on three datasets, our method exhibited superior SROCC and PLCC scores while maintaining comparable certified guarantees.

Paper Structure

This paper contains 31 sections, 2 theorems, 20 equations, 2 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Attacks and defenses on image quality metrics
  4. Randomized Smoothing for Image Classification
  5. Median Smoothing for the Regression Models
  6. Proposed method
  7. Problem statement
  8. Method
  9. Median Randomized Smoothing for IQA
  10. Auxiliary image denoiser.
  11. Mean squared error.
  12. Ranking loss.
  13. Target loss.
  14. Denoiser architecture.
  15. Bounded Deviation of Rank Correlation
  16. ...and 16 more sections

Key Result

Theorem 1

Given $G(M(x))$ in the form from eq:e2, for all $\|u\|_2 \le \varepsilon,$ Here, $\underline{p} = \Phi(- \frac{\varepsilon}{\sigma})$, $\overline{p} = \Phi(\frac{\varepsilon}{\sigma})$ and $\Phi(\cdot)$ is the Gaussian cumulative density function. $\underline{H}_{\underline{p}}(M(x))$ and $\overline{H}_{\overline{p}}(M(x))$ are defined as the percentiles of the smoothed fu

Figures (2)

  • Figure 1: a) The overview of the proposed DMS-IQA method. b) Training scheme of the auxiliary denoiser.
  • Figure 2: Comparison of the proposed DMS-IQA method with MS, DMS and undefended metric on adversarial images for CLIP-IQA+ and DBCNN NR IQA metrics. Green dots represent the certified guarantees. Results on the plot are averaged across images.

Theorems & Definitions (5)

  • Definition 3.1: $\tau-$closeness
  • Remark 1
  • Theorem 1
  • Remark 2
  • Lemma 1