Federated Learning in Advertising: Business Background, Conversion Flow, Algorithmic Techniques, Vertical & Horizontal FL, and Security

Huawei federated learning expert Liu Lu ("Avocado") presented the application of federated learning in the advertising sector, outlining six main topics: advertising business background, conversion flow from different perspectives, algorithmic techniques in the ad chain, vertical federated learning for ads, horizontal federated learning for ads, and attack‑defense technologies.

The advertising ecosystem consists of three primary parties—advertisers, ad platforms, and users. Huawei Ads supports multiple ad formats (banner, native, rewarded video, interstitial, splash) across various media resources (Huawei-owned and third‑party). Conversion tracking collects user actions (install, activation, payment) via APIs or SDKs, enabling platforms to link conversions to ad tasks and support models such as oCPC.

Conversion flow is examined from three viewpoints: (1) the user view, where a user sees an ad, clicks, downloads, and opens the app; (2) the client view, which records requests, bidding, impressions, clicks, downloads, installations, and activations with detailed event logging; (3) the server view, which aggregates exposure, click, download, install, and activation events for downstream analysis.

The ad pipeline involves SSP, ADX, and DSP components. Advertisers submit bids to DSPs; the highest‑bid ad wins the auction and is served. Models such as click‑through‑rate (CTR) and conversion‑rate (CVR) are trained using exposure and click data, with features spanning user, ad, context, high‑level, and low‑level attributes. The ECPM formula (price × pCTR × pCVR × 1000) determines winning bids.

Vertical federated learning addresses label scarcity and feature sparsity in conversion‑rate prediction. Using Huawei's Trusted Intelligent Computing Service (TICS), parties register data sources, define privacy policies, and perform PSI‑based sample alignment. After alignment, features and labels from advertisers and DSPs are combined to train models, with two product forms: label‑only sharing or label + feature sharing. TICS supports both containerized and edge deployments.

Horizontal federated learning leverages MindSpore’s FL Scheduler and FL Server. It offers privacy‑preserving aggregation, distributed scalability, efficiency improvements via model and communication compression (weight‑difference, sparsity, quantization), and simple API usage. An ALBERT example shows that compression can improve bandwidth without sacrificing accuracy.

Security considerations cover attacks before, during, and after training, including data poisoning, backdoor insertion, and gradient leakage. Defenses include model‑stability enhancements (additional trainable layers, dropout, auto‑encoders), gradient perturbation (noise, clipping, sparsification), hardware Trusted Execution Environments, and homomorphic encryption combined with TEE.

The Q&A session clarified that public advertising datasets for vertical FL are unavailable due to privacy and commercial constraints, TICS currently supports logistic regression and XGBoost with deeper models forthcoming, and federated learning complements rather than replaces existing conversion‑tracking mechanisms.