Ad Fraud Detection and Risk Control Practices at Alibaba Mama

1. Ad Risk Control Process

Describes the interaction flow between advertisers, risk control team, and downstream business, where ad creatives pass content risk review before display, and risk control filters invalid traffic during the display phase.

2. Invalid Traffic

Defines invalid traffic as low‑quality (e.g., repeated clicks, frequency control) and cheating traffic (zero conversion probability). Enumerates sub‑types such as competitor consumption, rank boosting, natural order brushing, non‑malicious waste, and Taobao affiliate (Taoke) cheating.

2.1 Competitor Consumption

Attackers generate fake traffic to exhaust competitors' budgets, causing premature ad removal.

2.2 Rank Boosting

Black‑market actors inflate ad quality scores to obtain higher placement.

2.3 Natural Order Brushing

Fake orders and reviews are used to improve shop metrics, sometimes unintentionally affecting ads.

2.4 Non‑Malicious Waste

Pre‑loading links, crawler traffic, and browser hijacking produce wasteful clicks that should not be billed.

2.5 Taoke Transaction Cheating

Includes traffic hijacking and black‑market affiliate recruitment, as well as fabricated delivery addresses for CPA campaigns.

3. Algorithmic Practice

Highlights the need for high‑precision, real‑time detection of sophisticated, distributed attacks and introduces a closed‑loop system comprising perception, insight, automatic disposal, and objective evaluation.

3.1 Perception

Early detection of anomalous traffic by comparing observed clicks against predicted normal levels, capturing both victim‑visible and platform‑visible anomalies.

3.2 Insight

Human analysts refine perceived anomalies, using high‑dimensional visual analytics to discover new risk patterns and distinguish simple from advanced fraud.

3.3 Disposal

Real‑time streaming filters combine unsupervised, semi‑supervised features with supervised ensemble detectors; an hourly batch filter handles more complex cases. Taoke cheating results in commission freezing and tiered penalties.

3.4 Evaluation

Four metrics—online supervised precision/recall and offline unsupervised precision/recall—are measured using an offline unlabeled sample pool and a graded sample library to differentiate simple and advanced fraud.

4. Summary

Research focuses on high‑dimensional anomaly detection, large‑scale graph learning, model interpretability, and data visualization, emphasizing that ad risk control demands extreme algorithmic robustness, precision, and industrial‑scale deployment.