Tencent join forces with WPP to prevent isolated data silos for advertisers

Using the Cloud Security Privacy Computing (CSPC) Platform, WPP and Tencent Cloud Big Data, have already run campaigns of multi-party federated learning with brands like Pernod Ricard.

Federated Learning is a distributed machine-learning framework that breaks down data silos to unleash the full potential of AI. Founded on the core principle of never disclosing underlying data, different parties can complete joint modeling by exchanging encrypted intermediate results and carrying out data enrichment legally and compliantly

“At the beginning of this year, we upgraded Shen Dun Federated Computing to Tencent Cloud Secure Privacy Computing, which is based on the Tencent Angel PowerFL privacy computing framework, with the protection technology of private data such as Federated Learning (FL), Secure Multiparty Computation (MPC), and Trusted Execution Environment (TEE),” said Naruto Guo, head of Tencent Cloud’s security privacy computing product.

“Customized privacy protection transformation is carried out for algorithms such as machine learning and data analysis to create privacy computing products with full links of data applications. The upgraded product ensures fast completion of privacy computing tasks with underlying data staying local. While ensuring data security, it can also maximize the value of data and solve the problem of data silos faced by enterprises.”

He adds: “Cooperation with WPP is a very good starting point in the field of advertising and marketing. The excellent security and performance of the platform play an important role in this advertising cooperation, ensuring data security and maximizing the value of enterprise data. The follow-up promotion and application also require joint efforts of all partners in the industry, and we look forward to creating more outstanding cases in the future.”

How will federated learning help advertisers?

• In the case of Pernod Ricard, for the implementation process, multiple parties input their owned audience profiles and device IDs in their respective modeling environments, which are based on the high-quality seed audiences of its brands Martell and the Glenlivet.

• After completing the model training, all parties perform federated scoring and reasoning for devices based on the platform’s hidden query function.

• As the core principle of federated learning is to never disclose the underlying data, all parties performed joint modeling by exchanging encrypted intermediate results and finally generated the accurate model expected by Martell and The Glenlivet through feature engineering and algorithm tuning.

• The audience group scored and selected by this model is then applied in the activation process of the mainstream publishers in the market.

• The results showed that for brands with large volumes of CRM historical sample data like Martell, the delivery performance generated by federated learning for relatively long-duration campaigns is significantly improved compared with that in the control group.

• For relatively newer brands such as The Glenlivet, the accurate model generated by federated learning using limited CRM data has obtained high-quality exposure, twice higher than that of the control group in short-duration campaigns.