Core Concepts
Adversarial attacks can significantly degrade the performance of no-reference image quality assessment metrics. This study investigates the effectiveness of various adversarial purification methods in defending against such attacks and restoring the original metric scores.
Abstract
This study focuses on improving the robustness of no-reference image quality assessment (IQA) metrics against adversarial attacks. The authors first create a dataset of adversarial images by applying 10 different attack methods to the Linearity, MetaIQA, and SPAQ no-reference IQA metrics. They then evaluate the performance of 16 adversarial purification techniques in defending against these attacks.
The key highlights and insights from the study are:
Simple transformations like image flipping and rotation can effectively neutralize the effects of adversarial attacks, but they may degrade the visual quality of the purified images.
More complex purification methods like DiffPure combined with unsharp masking provide the best balance between restoring the original metric scores and preserving the visual quality of the images.
The proposed FCN filter defense is particularly effective against the unrestricted AdvCF color attack, outperforming other methods in terms of output quality, attack neutralization, and metric score stability.
The authors provide a comprehensive analysis of the trade-offs between different evaluation metrics (quality score, gain score, and SROCC) for the tested purification techniques, offering insights into their strengths and weaknesses.
The study highlights the importance of developing provable defenses for IQA metrics, as the current work focuses on empirical attacks and defenses, which can lead to an endless cycle of attack and defense development.
Stats
Linearity metric shows high performance (correlation with subjective quality and speed) and medium robustness to adversarial attacks.
The NIPS 2017: Adversarial Learning Development Set was used as the reference dataset, with 10 attacks applied to each image.
Quotes
"Adversarial robustness has started to develop. This area is not as well-studied as the robustness of image classification or detection methods."
"Robust metrics are essential for developing contemporary image processing and compression methods. Such metrics will lead to the development of trusted benchmarks and allow researchers to use metrics as an optimization component to train processing methods and reduce costly subjective tests."