Adversarial Purification for No-Reference Image-Quality Metrics: Applicability Study and New Methods

Recently, the area of adversarial attacks on image quality metrics has begunto be explored, whereas the area of defences remains under-researched. In thisstudy, we aim to cover that case and check the transferability of adversarialpurification defences from image classifiers to IQA methods. In this paper, weapply several widespread attacks on IQA models and examine the success of thedefences against them. The purification methodologies covered differentpreprocessing techniques, including geometrical transformations, compression,denoising, and modern neural network-based methods. Also, we address thechallenge of assessing the efficacy of a defensive methodology by proposingways to estimate output visual quality and the success of neutralizing attacks.Defences were tested against attack on three IQA metrics – Linearity, MetaIQAand SPAQ. The code for attacks and defences is available at: (link is hiddenfor a blind review).