A paper got accepted!

We are thrilled to announce that our journal paper entitled “Not All Languages Are Equally Safe: Evaluating Cross-Lingual Data Poisoning Attacks Against Text-to-Image Models” has been accepted for publication in IEICE Transactions on Information and Systems, Special Section on Information and Communication System Security. This paper is an extended version of our ICPR 2026 paper. Congratulations Kakebayashi-kun!

In this work, we systematically evaluate the vulnerability of text-to-image (T2I) models to data poisoning attacks across 10 languages. We show that poisoned data constructed in English transfers to non-English prompts, and that several languages, such as Indonesian, Japanese, and Turkish, are notably more vulnerable than English, on both Stable Diffusion v2 and SDXL. Our analysis indicates that syntactic and genetic distance from English, as well as weaker cross-lingual alignment of the text encoder, are associated with higher attack success rates. Finally, we show that fine-tuning the text encoder with non-English data to improve cross-lingual alignment can reduce attack success rates.

Ryohei Kakebayashi and Tatsuya Mori. "Not All Languages Are Equally Safe: Evaluating Cross-Lingual Data Poisoning Attacks Against Text-to-Image Models." IEICE Transactions on Information and Systems, Special Section on Information and Communication System Security, to appear (Jun 2027).