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

Guest Speaker : Liu Lu, Huawei Federated Learning Expert (Beijing University Master’s graduate).

Overview : The talk focuses on the application of federated learning in the advertising industry.

1. Advertising Business Background

The advertising ecosystem consists of three main parties: advertisers, ad platforms, and users. Advertisers place tasks on platforms, which use big‑data and deep models to recommend suitable ads to users, leading to conversions and commercial value.

Ad formats : Banner, native, rewarded video, interstitial, and splash ads, which can be displayed across various media.

Media resources : Huawei’s own media (Video, Browser, Cloud Space, Reading) and third‑party media (games, video, news). Huawei’s high‑quality user base exceeds 740 million.

Conversion tracking : Advertisers’ conversion events (install, activation, payment, etc.) are reported to the platform via APIs or SDKs, enabling data‑driven optimization.

Why conversion tracking : It bridges the data gap between platforms and advertisers, allowing accurate measurement of conversions and supporting models such as oCPC.

Potential issues : Missing OAID, advertiser willingness to return data, privacy compliance, and data transmission reliability.

2. Conversion Process from Different Perspectives

User view (APP ad) : User sees the ad, clicks to download, installs, and activates the app.

Client view : Handles request, participates in bidding, records impressions, clicks, download start/completion, installation, and activation events.

Server view : Receives and logs exposure, click, download start/completion, installation, and activation events, completing the conversion flow.

3. Algorithm Techniques in the Ad Chain

The ad transaction flow involves SSP, ADX, DSP, and media, typically completing within ~100 ms. Data is stored in a DMP for user segmentation.

Key models: Click‑Through Rate (CTR) prediction and Conversion‑Rate (CVR) prediction.

Bid price (ECPM) = price × pCTR × pCVR × 1000; the highest ECPM wins the auction.

Features used include user, ad, context, high‑level (generalizable) and low‑level (interpretable) features.

4. Vertical Federated Learning for Advertising

Vertical FL solves label scarcity and feature sparsity by jointly training models between advertisers and DSPs, using TICS (Trusted Intelligent Computing Service) for secure multi‑party computation and blockchain‑based auditability.

TICS provides alliance management, data fusion, secure computation, and containerized deployment, supporting both cloud (CCE) and edge nodes.

Training steps: join alliance → select job type (federated data analysis or model training) → PSI for user intersection → sample alignment → feature selection → model training and evaluation (accuracy, AUC, KS, F1, recall).

Two product modes: (1) Advertiser provides only label Y; (2) Advertiser provides both label Y and feature X for joint training.

Performance considerations: offline vs. online inference, with latency constraints around 100 ms.

5. Horizontal Federated Learning for Advertising

Horizontal FL, originally proposed by Google, leverages client devices and ad servers using MindSpore’s FL Scheduler and FL Server.

Advantages: privacy‑preserving secure aggregation, distributed federated aggregation, efficiency improvements via model and communication compression, and ease of switching between single‑node and FL modes.

Compression pipeline on the client: weight difference → sparse encoding → quantization before uploading; server performs inverse operations before aggregation.

Example: ALBERT model reduces parameters from 99,221 to 7,937 after sparse encoding, with negligible accuracy loss.

6. Security and Defense in Federated Learning

Attacks can occur before, during, or after training, including data poisoning, backdoor attacks, and gradient‑inference attacks.

Defenses are categorized into: (1) Model‑stability enhancements (additional trainable layers, dropout, auto‑encoders); (2) Gradient perturbation (differential privacy, clipping, noise injection, sparsification); (3) Hardware‑based Trusted Execution Environments (TEE); (4) Homomorphic encryption combined with TEE.

7. Q&A Highlights

No publicly available vertical FL advertising datasets due to privacy and commercial value.

TICS currently supports logistic regression and XGBoost; deep models are forthcoming.

Federated learning complements, not replaces, conversion tracking (RTA) for fallback scenarios.

Huawei has deployed end‑cloud FL cases using MindSpore for overseas ad recommendation.

Conclusion : While homomorphic encryption research is advancing, practical training remains challenging; shallow networks incur 2‑3× performance loss; various optimizations (parallelism, hardware acceleration) can improve efficiency.