Practical Differentially Private Online Advertising

Abstract Powered by machine learning technology, online advertising achieves accurate advertisement delivery to potential customers according to online user profiles. However, it raises serious privacy concerns since the learning process may reveal sensitive information in the profiles. It is highly desirable to provide high-quality advertisement recommendations while respecting an individual’s privacy. To address the privacy issues, we propose a practical privacy-preserving system that predicts advertisement Click Through Rate (CTR) accurately without revealing any sensitive information of individuals. Our system combines both offline and online training to achieve the goal. In the offline training phase, we develop two differentially private algorithms built upon Gradient Boosting Decision Tree (GBDT) and Field-aware Factorization Machine (FFM) algorithms, respectively, to transform features and train a privacy-preserving offline model. In the online training phase, we propose a privacy-preserving online learning algorithm to compensate for the decline of offline performance when online feature changes occur. We perform extensive experiments to evaluate the performance of our system based on three real-world datasets. The results demonstrate that our system can protect user privacy without compromising the accuracy of CTR prediction, with reasonable computation overhead.