On Adversarial Robustness of Deep Image Deblurring

Abstract

Recent approaches employ deep learning-based solutions for the recovery of a sharp image from its blurry observation. This paper introduces adversarial attacks against deep learningbased image deblurring methods and evaluates the robustness of these neural networks to untargeted and targeted attacks. We demonstrate that imperceptible distortion can significantly degrade the performance of state-of-the-art deblurring networks, even producing drastically different content in the output, indicating the strong need to include adversarially robust training not only in classification but also for image recovery.

‘Pre-training+Fine-tuning’ Model-based 1/2 Folded Image Restoration

“…The latter mainly uses input-output model queries to generate adversarial samples, e.g., ZOO [36], SimBA [22], GeoDA [37], and Square Attack [23]. Attackers are not only targeting computer vision tasks [38]- [40], but are also trying to harm other DNN-based tasks, such as natural language processing [41] and image generation [42].…”

“…Recent research has explored the topic of adversarial attacks in low-level image restoration tasks, such as derain [52], superresolution [53]- [55], dehaze [40], deblur [38], and reflection removal [56]. Yu et-al.…”

“…Similar to adversarial attacks on other image restoration tasks [38], [56], the critical element in UDC adversarial attacks is the creation of visually imperceptible perturbations on the input images. The perturbations must be specifically tailored to the restoration models, ensuring they do not substantially degrade the visual quality.…”

Conditional Generative Models are Provably Robust: Pointwise Guarantees for Bayesian Inverse Problems

“…Antun et al [6] observed that deep learning for inverse problems comes with instabilities in the sense that ,,tiny, almost undetectable perturbations, both in the image and sampling domain, may result in severe artifacts in the reconstruction", while Genzel et al [16] attested in their comprehensive tests that ,, deep-learning-based methods are at least as robust as TV minimization with respect to adversarial noise". The authors of [15] showed experimentally the sensitivity of NN to perturbations for the inverse problem of image deblurring.…”

May denoising remove structures? How to reconstruct invariances of CT denoising algorithms

“…17,18 settings. Recently, adversarial attacks have also been investigated for CT imaging 19,20. While our work is not concerned with fooling DNNs via adversarial examples, we will see in Sec.…”