A recent publication of a comprehensive technical report marking the completion of Phase 1 of the Christchurch Call Initiative on Algorithmic Outcomes (CCIAO) was announced last month which highlights advancements in AI Auditing and safety, allowing external researchers to conduct privacy-preserving audits of production recommender algorithms at LinkedIn and Daily Motion using PySyft and OpenDP.
LinkedIn made impression data associated with their production recommender algorithm available for remote query (using PySyft APIs) by external researchers. The dataset consisted of ~70 million rows of data related to LinkedIn public post activity, with each row in the dataset representing the top‑ranked post in a user’s feed for a particular session. To ensure private information about users couldn’t leak into researchers’ queries, LinkedIn required that the researchers use OpenDP to implement differentially private queries, with each allocated a privacy budget set by the LinkedIn data managers. This enabled the researchers to answer meaningful research questions using the data, whilst ensuring user privacy was protected through the combined use of PySyft and OpenDP.