{"type":"ScholarlyArticle","title":"Exploiting Cultural Biases via Homoglyphs inText-to-Image Synthesis (Abstract Reprint)","authors":[{"name":"Lukas Struppek","slug":"lukas-struppek","given":"Lukas","family":"Struppek","affiliation":"Technical University of Darmstadt"},{"name":"Dominik Hintersdorf","slug":"dominik-hintersdorf","given":"Dominik","family":"Hintersdorf","affiliation":"Technical University of Darmstadt"},{"name":"Felix Friedrich","slug":"felix-friedrich","given":"Felix","family":"Friedrich","affiliation":"Technical University of Darmstadt; Hessian Center for AI (hessian.AI"},{"name":"Manuel Brack","slug":"manuel-brack","given":"Manuel","family":"Brack","affiliation":"German Center for Artificial Intelligence (DFKI); Technical University of Darmstadt"},{"name":"Patrick Schramowski","slug":"patrick-schramowski","given":"Patrick","family":"Schramowski","affiliation":"German Center for Artificial Intelligence (DFKI); Technical University of Darmstadt; Hessian Center for AI (hessian.AI"},{"name":"Kristian Kersting","slug":"kristian-kersting","given":"Kristian","family":"Kersting","affiliation":"Technical University of Darmstadt; Centre for Cognitive Science of Darmstadt {struppek ,; Hessian Center for AI (hessian.AI; German Center for Artificial Intelligence (DFKI)"}],"abstract":"Models for text-to-image synthesis, such as DALL-E 2 and Stable Diffusion, have recently drawn a lot of interest from academia and the general public. These models are capable of producing high-quality images that depict a variety of concepts and styles when conditioned on textual descriptions. However, these models adopt cultural characteristics associated with specific Unicode scripts from their vast amount of training data, which may not be immediately apparent. We show that by simply inserting single non-Latin characters in the textual description, common models reflect cultural biases in their generated images. We analyze this behavior both qualitatively and quantitatively and identify a model’s text encoder as the root cause of the phenomenon. Such behavior can be interpreted as a model feature, offering users a simple way to customize the image generation and reflect their own cultural background. Yet, malicious users or service providers may also try to intentionally bias the image generation. One goal might be to create racist stereotypes by replacing Latin characters with similarly-looking characters from non-Latin scripts, so-called homoglyphs. To mitigate such unnoticed script attacks, we propose a novel homoglyph unlearning method to fine-tune a text encoder, making it robust against homoglyph manipulations.","doi":"10.24963/ijcai.2024/958","url":"https://ijcai.org/proceedings/2024/958","pdf":"https://ijcai.org/proceedings/2024/0958.pdf","markdown":"https://ijcai.org/proceedings/2024/958.md","bibtex":"https://ijcai.org/proceedings/2024/bibtex/958","volume":{"slug":"2024","title":"Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence","short_title":"IJCAI 2024","url":"https://ijcai.org/proceedings/2024/","year":2024,"location":"Jeju, Korea","dates":"3-9 August 2024","editors":"Kate Larson","isbn":"978-1-956792-04-1","publisher":"International Joint Conferences on Artificial Intelligence Organization","doi_prefix":"10.24963/ijcai.2024"},"section":"Journal Track","track":"Journal Track","number":958,"pages":{"first":8486,"last":8486},"keywords":["Journal Track: Journal Track"],"videos":[],"references":[{"position":1,"raw":"[Struppek et al., 2023] Lukas Struppek, Dominik Hin- tersdorf, Felix Friedrich, Manuel Brack, Patrick Schramowski, and Kristian Kersting. Exploiting cul- tural biases via homoglyphs in text-to-image synthesis. J. Artif. Intell. Res., 78:1017-1068, 2023.","title":"Exploiting cultural biases via homoglyphs in text-to-image synthesis","authors":"Lukas Struppek, Dominik Hintersdorf, Felix Friedrich, Manuel Brack, Patrick Schramowski, Kristian Kersting","year":"2023","venue":"J. Artif. Intell. Res","doi":null,"url":null,"internal_url":null}],"license_note":"Copyright International Joint Conferences on Artificial Intelligence."}