This repository contains the source code, datasets, generated content, and scripts for the paper "From Meme to Threat: On the Hateful Meme Understanding and Induced Hateful Content Generation in Open-Source Vision Language Models."
In this paper, we present an in-depth evaluation of VLMs' ability to interpret hateful memes by curating a dataset of 39 hateful memes and over 12,000 responses from seven representative VLMs using carefully designed prompts. We also assess how malicious users could exploit VLMs and hateful memes to generate hateful content systematically. We generate hateful content (hate speech, jokes, and slogans) based on hateful memes and calculate the hatefulness of the generated content.
We have published the annotated dataset on Hugging Face (https://huggingface.co/datasets/TrustAIRLab/Hateful_Memes_in_VLM). Due to the inclusion of sensitive information in the source code and dataset, the detailed project data is hosted on Zenodo (https://zenodo.org/records/14752660) and is available upon request.