Privacy in AI

Recent research on privacy in AI is increasingly focused on enhancing the security of large vision-language models (LVLMs) and large language models (LLMs) against potential exploitation by malicious actors. The introduction of methods like neuron-level gradient gating aims to bolster privacy safeguards without sacrificing model performance, addressing critical vulnerabilities where sensitive information could be extracted. Additionally, studies reveal that many models struggle to respect contextual integrity, particularly in geolocation tasks, leading to over-disclosure of sensitive information. This highlights the need for models to incorporate nuanced reasoning about privacy expectations in real-world scenarios. Furthermore, new attacks on privacy-preserving systems have exposed weaknesses in existing protections, underscoring the inadequacy of heuristic approaches. The field is shifting towards more robust strategies that decouple generalizability from privacy risks, aiming to enhance both user safety and model utility in practical applications across sectors like healthcare and finance.