Safelog: Supporting Web Search and Mining by Differentially-Private Query Logs

Query logs can be very useful for advancing web search and web mining research. Since these web query logs contain private, possibly sensitive data, they need to be effectively anonymized before they can be released for research use. Anonymization of query logs differs from that of structured data since they are generated based on natural language and the vocabulary (domain) is infinite. This unstructured, unbounded data set poses different challenges for producing privacy-preserving query logs. To mitigate these challenges, we propose using a differential privacy framework called Safelog to generate anonymized query logs that contain sufficient contextual information to allow existing web search and web mining algorithms to use the data and attain meaningful results. The key to achieving high privacy guarantees is the introduction of a query pool for augmenting the query log during the sanitation process. We empirically validate the effectiveness of our framework for generating usable, privacy preserving logs for web search, and demonstrate that it is possible to maintain high utility for this task while guaranteeing sufficient levels of privacy. We conclude with a discussion of other web mining tasks that can be supported by these anonymized logs and show some preliminary results for the task of website clustering.