Home / Case Studies / AdWords Fraud Detection
Production AI Engineering
Project 06

AdWords Fraud Detection

Graph-based detection infrastructure and reviewer tooling that protected the ad exchange ecosystem at global scale.

Company Google
Year 2018–2019
Role ML Engineer & Platform Lead
200+ Fraud reviewers empowered by the platform
Millions Abusive accounts identified across the ecosystem
VP Award Recognized for ecosystem integrity impact
01 — Challenge

The Problem

Vikram built detection infrastructure that empowered 200+ fraud analysts to identify millions of abusive accounts — converting individual review into coordinated, ring-level intelligence.

The ad exchange ecosystem faced coordinated abuse at massive scale — fraud networks operating across millions of accounts with behavioral patterns that individual account review could never catch. Manual review processes were reactive and slow, unable to surface the ring-and-chain relationships that distinguished coordinated fraud from isolated violations.

The detection challenge was compounded by the scale of data involved: processing massive volumes of account behavioral data in near-real time while maintaining accuracy high enough to avoid penalizing legitimate advertisers. Continuously evolving fraud patterns required the system to be architecturally adaptable, not just accurate at point-in-time.

"The goal wasn't to catch more fraud — it was to give 200 human reviewers the intelligence to catch fraud that no single person could ever see alone."
02 — Architecture

System Design

Component 01
Anomaly Detection Engine
Feature extraction across large volumes of account behavioral data. Suspicious pattern detection and abuse identification at ad exchange scale. Continuously updated detection models adapting to evolving fraud patterns.
Component 02
Account Review Platform
Data aggregation and relationship/similarity scoring per account. Ring and chain visualization of fraud networks for intuitive reviewer decision-making. Bulk suspension enablement multiplying reviewer throughput at scale. Graph-based relationship scoring turning complex fraud network signals into actionable visual intelligence.

Key Innovations

03 — Impact

Results

200+ reviewers Empowered to protect millions of accounts
★ VP Award Winner

The platform identified millions of abusive accounts, strengthening the integrity of the ad exchange advertising ecosystem at global scale. By empowering 200+ reviewers with intelligent tooling — graph visualization, similarity scoring, and bulk actions — the system multiplied human reviewer capacity without adding headcount. The project earned a VP Award, reflecting its significant financial and operational impact on one of the company's most revenue-critical platforms.

04 — Contribution

My Role

Explainer Video

Project Overview

05 — Stack

Technology

Backend
Java
Frontend
AngularJS
Data
Flume Apache Beam
ML
Anomaly Detection Models Graph-based Relationship Scoring
06 — Capabilities

Related Expertise