New benchmarks and tuning data improve VLM privacy awareness

Researchers have developed new methods to enhance the privacy awareness of Visual Language Models (VLMs). They introduced two benchmarks, PrivBench and PrivBench-H, designed to evaluate VLMs' understanding of visual privacy in line with GDPR. Additionally, a curated instruction-tuning dataset called PrivTune was created to improve privacy sensitivity. Fine-tuning existing VLMs with a small amount of PrivTune data significantly boosted their performance on these privacy benchmarks, even surpassing GPT-4, while maintaining their general task capabilities. AI

IMPACT Enhances privacy safeguards in VLMs, potentially leading to safer integration into user-facing applications.

RANK_REASON The cluster contains an academic paper detailing new benchmarks and a fine-tuning dataset for improving the privacy awareness of Visual Language Models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• General Data Protection Regulation
• GPT-4
• PrivBench
• PrivBench-H
• PrivTune
• Visual Language Models

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →