Adversarial attacks dataset for low light image enhancement

Enhancement tasks for low-light images are an open problem whose solutions can impact downstream tasks in fields like surveillance, medicine, autonomous driving, or object detection. As a result, the necessity to create algorithms that can withstand harmful and subtle disturbances has garnered interest, as both humans and machines can significantly influence how programs perform. Enhancing images in low-light conditions is a field of research where deep convolutional neural networks have shown considerable effectiveness. Nevertheless, the reliability of these systems poses a significant challenge that needs examination and remedies against hacking attempts. This image repository dataset evaluates adversarial robustness by introducing 12 examples, each leveraging a known adversarial attack or noise perturbation. The repository contains a total of 4970 digital images through the application of adversarial examples to five different image databases (i.e., MEF, LIME, LOLv1, LOLv2-R, LOLv2-S) that include pictures of landscapes, objects, or locations with low, normal, or high illumination conditions. We incorporate original images from the five databases utilized alongside the adversarial examples to facilitate easy access and distribution of this dataset.