Towards Adversarial Robustness in Blind Image Quality Assessment with Soft Thresholding Norm

In this study, we address the issue of adversarial robustness within the context of Blind Image Quality Assessment (BIQA), an area of heightened importance due to the inherent susceptibility of Deep Neural Networks (DNNs) to adversarial assaults. Current approaches primarily rely on adversarial training, which, despite its efficacy, imposes a significant computational burden. Our research proposes an alternative strategy known as the Soft Thresholding Norm (ST Norm). This approach counters the ‘feature shift’ phenomenon, identified by a substantial Euclidean Distance Statistics (EDS) between original and adversarial features, through the imposition of sparse constraints on potential features following batch normalization. This novel method offers several advantages: it reduces the Lipschitz constant yielding smoother models, seamlessly integrates with existing models, and boasts inherent denoising capabilities, thereby effectively mitigating the impact of adversarial perturbations. Results suggest that our approach achieves robustness comparable to adversarial training but with significantly less computational overhead. Moreover, it consistently outperforms other adversarial defense strategies on BIQA datasets, highlighting its practical effectiveness in enhancing adversarial robustness. This study underscores the potential of the Soft Thresholding Norm within the realm of IQA tasks, positioning it as a resource-efficient alternative to traditional adversarial training methodologies.