Researchers at CWI, together with the United Nations, have developed an AI method that filters sensitive information from the Humanitarian Data Exchange platform. This way, humanitarian data is better protected against misuse.
The United Nations’ Humanitarian Data Exchange (HDX) platform contains more than 19,000 datasets from 254 crisis areas worldwide. Local governments share data on conflicts, natural disasters, and humanitarian emergencies. This information is crucial for coordinating relief efforts but can be dangerous if sensitive data falls into the wrong hands.
CWI researcher Madelon Hulsebos and AI master’s student Liang Telkamp therefore developed a new method to detect and filter sensitive data. Their research introduces the concept of ‘contextually sensitive data.’ This does not only concern personal data but also information that, depending on time, place, and situation, can cause harm. For example, the coordinates of a hospital in the Netherlands are not immediately sensitive, whereas those of a hospital in a war zone are.
Previously, the UN relied on Google DLP, but this tool proved far less accurate and generated many false positives. The new AI approach uses UN guidelines and Large Language Models (such as GPT-4 and Qwen) to automatically assess datasets. This resulted in clear improvements: while Google DLP detected only 63 percent of sensitive personal data, the new method achieved up to 94 percent. In addition, the number of false positives was cut in half, significantly reducing the workload of Quality Assurance Officers.
The research also highlights the importance of explainability. The AI models were not only able to identify sensitive data but also to explain why certain information was labeled as sensitive. This increases the reliability and usability of the results.
The United Nations has now decided to integrate the new mechanisms into the HDX platform. In October, Hulsebos will present the results at a UN meeting in Barcelona. The applications extend beyond the UN itself: public cloud platforms and open data portals can also benefit. This makes the method a valuable step toward safer use of data at a time when datasets are also used as training material for AI models.
Read the full article on the website of CWI.
Read also
