2022
DOI: 10.48550/arxiv.2209.08891
|View full text |Cite
Preprint
|
Sign up to set email alerts
|

Exploiting Cultural Biases via Homoglyphs in Text-to-Image Synthesis

Help me understand this report

Search citation statements

Order By: Relevance

Paper Sections

Select...
3
2

Citation Types

0
6
0

Year Published

2023
2023
2024
2024

Publication Types

Select...
3
1

Relationship

0
4

Authors

Journals

citations
Cited by 4 publications
(6 citation statements)
references
References 0 publications
0
6
0
Order By: Relevance
“…Recent approaches in the field of model vulnerability have mainly concentrated on evaluating the quality of generated images through untargeted attacks [9,10,36,37]. These studies have looked into the correlation of images with the attributes [37] and objects [9,10] of the prompt.…”
Section: Introductionmentioning
confidence: 99%
See 1 more Smart Citation
“…Recent approaches in the field of model vulnerability have mainly concentrated on evaluating the quality of generated images through untargeted attacks [9,10,36,37]. These studies have looked into the correlation of images with the attributes [37] and objects [9,10] of the prompt.…”
Section: Introductionmentioning
confidence: 99%
“…3) Difficulty in revealing model vulnerability mechanisms. Although we have a series of studies to expose the vulnerabilities of Stable Diffusion [9,36], there is limited information on revealing the reasons behind the success of adversarial attacks. Furthermore, the results of the existing vulnerability analysis [11,28] for small-scale models such as ResNet [13] cannot be applied directly to Stable Diffusion, which poses significant challenges in revealing the vulnerability of the model.…”
Section: Introductionmentioning
confidence: 99%
“…Moreover, it proposes Safe Latent Diffusion, which successfully removes and suppresses the inappropriate content with additional guidance. Another ethical issue, the fairness of social group, is studied in [89], [90]. Specifically, [89] finds that simple homoglyph replacements in the text descriptions can induce culture bias in models, i.e., generating images from different culture.…”
Section: Ethical Issues and Risksmentioning
confidence: 99%
“…Another ethical issue, the fairness of social group, is studied in [89], [90]. Specifically, [89] finds that simple homoglyph replacements in the text descriptions can induce culture bias in models, i.e., generating images from different culture. [90] introduce an Ethical NaTural Language Interventions in Text-to-Image GENeration (ENTIGEN) benchmark dataset, which can evaluate the change of generated images with ethical interventions by three axes: gender, skin color, and culture.…”
Section: Ethical Issues and Risksmentioning
confidence: 99%
See 1 more Smart Citation