A Norm Regularization Training Strategy for Robust Image Quality Assessment Models

Image Quality Assessment (IQA) models predict the quality score of input images. They can be categorized into Full-Reference (FR-) and No-Reference (NR-) IQA models based on the availability of reference images. These models are essential for performance evaluation and optimization guidance in the media industry. However, researchers have observed that introducing imperceptible perturbations to input images can notably influence the predicted scores of both FR- and NR-IQA models, resulting in inaccurate assessments of image quality. This phenomenon is known as adversarial attacks. In this paper, we initially define attacks targeted at both FR-IQA and NR-IQA models. Subsequently, we introduce a defense approach applicable to both types of models, aimed at enhancing the stability of predicted scores and boosting the adversarial robustness of IQA models. To be specific, we present theoretical evidence showing that the magnitude of score changes is related to the \(\ell _1\) norm of the model’s gradient with respect to the input image. Building upon this theoretical foundation, we propose a norm regularization training strategy aimed at reducing the \(\ell _1\) norm of the gradient, thereby boosting the robustness of IQA models. Experiments conducted on three FR-IQA and four NR-IQA models demonstrate the effectiveness of our strategy in reducing score changes in the presence of adversarial attacks. To the best of our knowledge, this work marks the first attempt to defend against adversarial attacks on both FR- and NR-IQA models. Our study offers valuable insights into the adversarial robustness of IQA models and provides a foundation for future research in this area.