New AI Model Helps Google Ads Identify Fraudulent Advertisers

SEO
1

image
image760×400 50.2 KB

Google Deploys ALF: A Next-Gen AI for Fraud Detection in Google Ads

On December 31, 2025, Google published a research paper introducing ALF (Advertiser Large Foundation Model), a new AI system for detecting fraud in Google Ads. According to the paper, ALF is already deployed and has shown substantial improvements over previous systems, achieving:

  • +40 percentage points increase in detection recall for key policies

  • 99.8% precision on certain policy violations

What is ALF?

ALF is a multimodal large foundation model designed to understand advertisers holistically. It analyzes:

  • Structured data: account age, billing details, historical performance

  • Unstructured creative assets: text, images, and videos

  • Landing page content

The model works by comparing multiple signals together, rather than evaluating them in isolation. For example:

“An advertiser might have a newly created account, ads featuring a well-known brand, and a single declined payment. Individually, these may seem innocuous, but together they strongly suggest fraudulent intent.”

Key Challenges ALF Overcomes

  1. Heterogeneous & High-Dimensional Data

    • Advertiser data comes in many formats (structured and unstructured) and contains thousands of features, which previous models struggled to process effectively.

  2. Unbounded Sets of Creative Assets

    • Malicious content can be hidden among thousands of innocent assets. Previous systems could not reliably detect these outliers.

  3. Real-World Reliability & Trustworthiness

    • ALF generates trustworthy confidence scores to minimize false positives, protecting legitimate advertisers while maintaining detection accuracy.

Privacy and Safety

  • ALF analyzes sensitive account signals, but all personally identifiable information (PII) is stripped before processing.

  • The model focuses on behavioral patterns, not individual identities.

The “Secret Sauce”: Inter-Sample Attention

  • ALF doesn’t evaluate advertisers in isolation.

  • Using large advertiser batches, it compares behaviors across the ecosystem.

  • This contextual comparison helps identify suspicious outliers more accurately than previous systems.

Performance Highlights

  • ALF outperforms production baselines and public benchmarks.

  • Gains are achieved across multiple evaluation metrics in real-world production conditions, not just offline testing.

  • While ALF has higher latency than simpler models, it remains well within acceptable production limits and scales to handle millions of requests daily.

Impact on Google Ads Safety

  • ALF is currently deployed for fraud detection and policy enforcement in Google Ads.

  • Future potential applications include:

    • Temporal modeling for evolving fraudulent patterns

    • Audience modeling

    • Creative optimization

Why It Matters

ALF represents a major leap in AI-driven advertiser monitoring:

  • Combines multimodal content understanding with structured account data

  • Delivers high precision and recall, improving ad ecosystem trustworthiness

  • Balances accuracy with scalable deployment at production scale

Original Research Paper:
ALF: Advertiser Large Foundation Model for Multi-Modal Advertiser Understanding