Professor and Associate Dean for Research Dong Wang and students from his Social Sensing and Intelligence Lab will be presenting papers at the ACM Web Conference in Dubai from June 29 to July 3.
"Red-Teaming Privacy-Protective Perturbations: Blind Face Restoration as an Attack Strategy," a paper by Wang and Zelin Li, an Informatics PhD student, examines whether current approaches for protecting personal photos from AI misuse are effective. Many new tools allow people to add subtle changes to their photos before sharing them online, making it more difficult to use those images to create AI-generated content without permission.
The study reveals that an AI technology originally developed to restore old, damaged, or low-quality face photos can also weaken these protections. Through extensive experiments, the researchers show that face restoration models can remove protective modifications and bypass existing safeguards more effectively than current attack techniques. The findings raise important questions about the security of today’s photo-protection methods and highlight the need for stronger privacy safeguards.
PhD student Yaokun Liu and Wang are co-presenting the paper "Mind the Ambiguity: Aleatoric Uncertainty Quantification in LLMs for Safe Medical Question Answering."
This project tackles a key safety challenge in AI-powered medical assistants: people often ask health-related questions without providing enough information, such as their age, symptoms, medications, or medical history. Yet large language models (LLMs) still respond with a confident answer, even when these important details are missing.
The researchers introduce CV-MedBench, a new benchmark that evaluates how well AI models handle clear versus ambiguous medical questions. They find that models often detect uncertainty from their internal signals before generating a response. Building on this insight, the team develops AU-Probe, a lightweight approach that helps AI systems recognize when a question is too general to answer safely. Instead of making a guess, the system will ask users for additional information first. Experiments across four open-source LLMs show that this "clarify before answer" approach improves medical question responses and helps make AI tools safer, more accurate, and more trustworthy.
Launched in 1989, the conference brings together scholars, researchers, policymakers, practitioners, and internet users focused on the future of the web. The conference has seen the launch of groundbreaking innovations in internet technology, including search engines.