3. PromptFake
[This project is a part of my tenure at University of South Carolina as an affiliate Researcher]
How did I generate this image? What's the back story?
My conversations in the research group at USC and the ongoing work on explainable fact verification led to exploring stable diffusion and it was during that time I noticed Elon Musk had fired the entire team responsible for fake news verification on Twitter.
I was experimenting with stable diffusion which is a text-to-image generation model. My prompt was Elon Musk got married to Kamla Harris and it generated the image you see.
This further led to a research proposal of generating a benchmark multimodal fact verification dataset and baseline model for verifying the claims.
Problem
Combating disinformation is one of the burning societal crises - about 67% of the American population believes that disinformation produces a lot of uncertainty, and 10% of them knowingly propagate disinformation.
Disinformation can manipulate democratic processes and public opinion
It can disrupt the share market, cause panic and anxiety in society, and even death during crises.
Despite progress in automatic text-based fact verification (e.g., FEVER, LIAR), there is no significant effort on multimodal fact verification
Proposed Solution: A benchmark 1mn multimodal fake news dataset
Prompt-based data Augmentation: In the proposed solution, the use of prompts from fact verification datasets such as FEVER is fed into a stable diffusion model to synthesize images corresponding to the prompts
Document Generation of text-based prompts: Large language models such as BLOOM or GPT3 generate documents of text based on the same prompts to create fake news article.
5W Explanation: To evaluate the generated document based on the prompt, a novel 5W architecture is proposed
My Contribution
Worked on data augmentation and experimented with 5W explanation
Results
Work published at EMNLP 2023 [Paper Link]