Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

An Art-centric perspective on AI-based content moderation of nudity

Piera Riccio, Georgina Curto, Thomas Hofmann, Nuria Oliver

TL;DR

A multi-modal zero-shot classification approach that improves artistic nudity classification is proposed, and empirically uncover the existence of a gender and a stylistic bias, as well as evident technical limitations, especially when only considering visual information.

Abstract

At a time when the influence of generative Artificial Intelligence on visual arts is a highly debated topic, we raise the attention towards a more subtle phenomenon: the algorithmic censorship of artistic nudity online. We analyze the performance of three "Not-Safe-For-Work'' image classifiers on artistic nudity, and empirically uncover the existence of a gender and a stylistic bias, as well as evident technical limitations, especially when only considering visual information. Hence, we propose a multi-modal zero-shot classification approach that improves artistic nudity classification. From our research, we draw several implications that we hope will inform future research on this topic.

An Art-centric perspective on AI-based content moderation of nudity

TL;DR

A multi-modal zero-shot classification approach that improves artistic nudity classification is proposed, and empirically uncover the existence of a gender and a stylistic bias, as well as evident technical limitations, especially when only considering visual information.

Abstract

At a time when the influence of generative Artificial Intelligence on visual arts is a highly debated topic, we raise the attention towards a more subtle phenomenon: the algorithmic censorship of artistic nudity online. We analyze the performance of three "Not-Safe-For-Work'' image classifiers on artistic nudity, and empirically uncover the existence of a gender and a stylistic bias, as well as evident technical limitations, especially when only considering visual information. Hence, we propose a multi-modal zero-shot classification approach that improves artistic nudity classification. From our research, we draw several implications that we hope will inform future research on this topic.
Paper Structure (20 sections, 3 figures, 4 tables)

This paper contains 20 sections, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Image classification algorithms for content moderation
  4. Artistic vs pornographic nudity
  5. Models and Data
  6. 1. NSFW classifiers
  7. 2. Datasets
  8. Content moderation on artistic nudity
  9. Analysis of Biases
  10. Sensitivity to gender and time period
  11. Inter-algorithm analysis
  12. Sensitivity to artistic style
  13. Fine-tuning
  14. Implementation
  15. Results
  16. ...and 5 more sections

Figures (3)

  • Figure 1: Exemplary images in D01 that are considered to be unsafe (top) or safe (bottom) by the three NSFW classifiers.
  • Figure 2: Exemplary images in D02 that are considered to be unsafe (first line) or safe (second line) by all the three models.
  • Figure 3: Left: Recall gain/loss (in percentage points) on each of the three test sets after fine-tuning each of the three NSFW classifiers. The results are shown as boxplots with the mean (white dot) and the standard deviation (bars) of the recall gain/loss over the 5 considered folds. Right: t-SNE projection of the CLIP textual embeddings of the considered terms in $S_{porn}$ and $S_{art}$ with PCA initialization. The existence of two clusters is confirmed via k-means.