Literature Characterization

Overview

To organize the included papers, we used the categorization scheme, which indexes the papers according to the following 3 dimensions. Literature categorization dimensions.
Problem Setup includes the assumptions, definitions, and the scope of the problem. For example, the data distributions that the paper focus on, the type of learning tasks, and the learning algorithm the paper studies. It also records how is robustness defined and evaluated in the paper.
Data Property records the data properties discussed in the paper. We identified 8 data properties in total, seven of them are application-agnostic and one includes the data properties that are specific to application domains. We organized the included papers by the data properties they discussed, which helps see the high-level trends regarding each data property and model adversarial robustness.
Practicality assesses the applicability, measurability, and explanability of the paper's conclusions. It evaluates if the data properties discussed are backed by quantitative metrics and whether the paper present techniques to modulate in order to improve robustness. Additionally, this dimension also documents the verification methods for the paper's conclusions, noting whether they were substantiated through formal proof or empirical validation. Where relevant, the setup utilized for empirical validations, such as the datasets employed for experimentation, is also recorded. Furthermore, this dimension also indicates whether the paper provides explanation for the correlation identified.

Detailed Categorization of the papers

Categorization of the properties of each papers is given in the tables below, you can also find the pdf version with the references included here: Download

We used the following abbreviation to denote the datasets discussed in the papers: M for MNIST, FM for FASHION-MNIST, S for SVHN, C-10 for CIFAR-10, C-100 for CIFAR-100, IN for IMAGENET, TI for TINY-IMAGENET, CA for CELEBA, HM for HALFMOON, M1V7 for MNIST 1V7, A for ABALONE, L for LSUN, CS for CITYSCAPES, TO for TASKONOMY, W for Wikipedia-31K, AC for AmazonCat-13K, MC for MINC, G for GTOS, F for Fundoscopy, CX for Chest X-Ray, D for Dermoscopy.

Table 5: Detailed paper categorization - part 1.

Table 5: Detailed paper categorization - part 2.

Table 5: Detailed paper categorization - part 3.

Table 5: Detailed paper categorization - part 4.