Advances in AI Privacy and Safety

The field of AI research is moving towards a greater emphasis on privacy and safety, with a focus on developing more robust and nuanced models that can understand and enforce privacy principles. Recent studies have highlighted the importance of situating privacy preference elicitation within real-world data flows and have introduced new approaches for evaluating the harmfulness of content generated by large language models. Noteworthy papers in this area include:

  • Falcon, which introduces a large-scale vision-language safety dataset and a specialized evaluator for identifying harmful content in complex and safety-critical multimodal dialogue scenarios.
  • LLaVAShield, which presents a systematic definition and study of multimodal multi-turn dialogue safety and introduces a powerful tool for detecting and assessing risk in user inputs and assistant responses.

Sources

Not My Agent, Not My Boundary? Elicitation of Personal Privacy Boundaries in AI-Delegated Information Sharing

A First Look at Privacy Risks of Android Task-executable Voice Assistant Applications

Falcon: A Cross-Modal Evaluation Dataset for Comprehensive Safety Perception

Assessing Visual Privacy Risks in Multimodal AI: A Novel Taxonomy-Grounded Evaluation of Vision-Language Models

Defeating Cerberus: Concept-Guided Privacy-Leakage Mitigation in Multimodal Language Models

LLaVAShield: Safeguarding Multimodal Multi-Turn Dialogues in Vision-Language Models

Judging by Appearances? Auditing and Intervening Vision-Language Models for Bail Prediction

Built with on top of