Abstract
Online advertising networks utilize prediction models to determine the likelihood of an advertisement delivered to a user meeting a particular objective, e.g., making a purchase from the advertiser. Such models are trained based on historical information about users that clicked the ad logged by the advertising network and purchase data obtained from the advertiser. This mechanism is at odds with privacy since the advertising network receives user data from the advertiser. This disclosure describes the use of multi-party computation techniques to train a predictor. An MPC exchange is performed in which labels from the advertiser and a feature vector of user features from the ad network are utilized to train a prediction model. Partial models are obtained for each advertiser and are combined to obtain a trained predictor. User privacy is thus preserved since no user data from the advertiser is revealed to the advertising network and vice versa.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Anonymous, "Multi-party Computation to Train a Prediction Model for Advertisement Selection", Technical Disclosure Commons, (June 12, 2020)
https://www.tdcommons.org/dpubs_series/3319