Privacy-Preserving Machine Learning Pipelines For Publisher Data Collaboration: A Framework For Federated Inference In Post-Cookie Advertising Ecosystems

Abstract

Third-party cookie deprecation severs the data pipeline through which publishers have historically collaborated on machine learning model quality. This article describes a federated inference framework that enables competing publishers to contribute differentially private gradient updates to a shared model without exposing raw audience data. We propose a round-based coordination technique with gradient clipping, inclusion of Gaussian noise, and cryptographically safe aggregation that hides the individual publisher contributions while enabling the coordinator to compute an accurate aggregate. Formal privacy analysis establishes (ε, δ)-differential privacy per publisher per training round under the honest-but-curious threat model, with cumulative privacy spend tracked by a global privacy accountant coordinating the budget across all participants. Inference accuracy trade-offs are characterized as a function of privacy budget, dataset size, and participation breadth, using benchmark results from Abadi et al. (2016) to ground the analysis. The framework addresses the specific cross-publisher collaboration setting — competing commercial entities with heterogeneous datasets and no trusted arbitrator — that existing federated learning literature does not fully cover. Operational lessons from publisher cloud infrastructure are documented, including gradient weighting for large-publisher dominance, protocol exit-tolerance as a production necessity, and feature ontology negotiation as an underestimated integration cost.