Introduction to ALF
Google has developed a new AI model called ALF (Advertiser Large Foundation Model) to detect fraud in the Google Ads system. This model has shown a significant improvement over the previous system, with a 40% increase in detection rate and 99.8% precision on specific policies. The research paper, published on December 31, 2025, provides details on how ALF works and its effectiveness in detecting fraudulent activities.
What is ALF?
ALF is a multimodal large foundation model that analyzes various factors such as text, images, and videos, along with account age, billing details, and historical performance metrics. This approach allows for a better understanding of advertiser behavior and intent. The researchers explain that many of these factors alone may not raise any red flags, but when combined, they can indicate potentially problematic activity.
Challenges Overcome by ALF
The previous systems faced several challenges, including:
Heterogeneous and High-Dimensional Data
ALF can handle multiple data formats, including structured and unstructured data, and process hundreds or thousands of data points associated with each advertiser.
Unbounded Sets of Creative Assets
The model can analyze thousands of creative assets, such as images, and identify malicious ones among innocent assets.
Real-World Reliability and Trustworthiness
ALF generates trustworthy confidence scores to identify malicious intent without affecting innocent advertisers.
Privacy and Safety
The researchers emphasize that ALF is designed with strict privacy safeguards. All personally identifiable information (PII) is removed before the AI processes any data, ensuring that the model identifies risk based on behavioral patterns rather than sensitive personal data.
The Secret Sauce: Inter-Sample Attention
ALF uses a technique called "Inter-Sample Attention" to improve its detection skills. By analyzing large advertiser batches, the model learns what normal activity looks like across the entire ecosystem and becomes more accurate in spotting suspicious outliers.
ALF Outperforms Production Benchmarks
The researchers’ experiments show that ALF significantly outperforms a heavily tuned production baseline, boosting recall by over 40 percentage points on one critical policy while increasing precision to 99.8% on another. Although ALF’s latency is higher due to its larger model size, it remains within the acceptable range for production use.
Improved Fraud Detection
ALF is now deployed to the Google Ads Safety system to identify advertisers violating Google Ads policies. Future work may focus on time-based factors ("temporal dynamics") to catch evolving patterns and audience modeling and creative optimization.
Conclusion
The development of ALF marks a significant improvement in detecting fraud in the Google Ads system. Its ability to analyze multiple factors, handle large datasets, and generate trustworthy confidence scores makes it an effective tool in identifying malicious intent. As ALF continues to operate at scale, it is likely to have a positive impact on the online advertising ecosystem, providing a safer and more reliable experience for users. The research paper on ALF can be found at ALF: Advertiser Large Foundation Model for Multi-Modal Advertiser Understanding.