Google published a research paper about a new AI model for detecting fraud in the Google Ads system that’s a strong improvement over what they were previously using. What’s interesting is that the research paper, dated December 31, 2025, says that the new AI is deployed, resulting in an improvement in the detection rate of over 40 percentage points and achieving 99.8% precision on specific policies.
ALF: Advertiser Large Foundation Model
The new AI is called ALF (Advertiser Large Foundation Model), the details of which were published on December 31, 2025. ALF is a multimodal large foundation model that analyzes text, images, and video, together with factors like account age, billing details, and historical performance metrics.
The researchers explain that many of these factors in isolation won’t flag an account as potentially problematic, but that comparing all of these factors together provides a better understanding of advertiser behavior and intent.
They write:
“A core challenge in this ecosystem is to accurately and efficiently understand advertiser intent and behavior. This understanding is critical for several key applications, including matching users with ads and identifying fraud and policy violations.
Addressing this challenge requires a holistic approach, processing diverse data types including structured account information (e.g., account age, billing details), multi-modal ad creative assets (text, images, videos), and landing page content.
For example, an advertiser might have a recently created account, have text and image ads for a well known large brand, and have had a credit card payment declined once. Although each element could exist innocently in isolation, the combination strongly suggests a fraudulent operation.”
The researchers address three challenges that previous systems were unable to overcome:
1. Heterogeneous and High-Dimensional Data
Heterogeneous data refers to the fact that advertiser data comes in multiple formats, not just one type. This includes structured data like account age and billing type and unstructured data like creative assets such as images, text, and video. High-dimensional data refers to the hundreds or thousands of data points associated with each advertiser, causing the mathematical representation of each one to become high-dimensional, which presents challenges for conventional models.
2. Unbounded Sets of Creative Assets
Advertisers could have thousands of creative assets, such as images, and hide one or two malicious ones among thousands of innocent assets. This scenario overwhelmed the previous system.
3. Real-World Reliability and Trustworthiness
The system needs to be able to generate trustworthy confidence scores that a business has malicious intent because a false positive would otherwise affect an innocent advertiser. The system must be expected to work without having to constantly retune it to catch mistakes.
Privacy and Safety
Although ALF analyzes sensitive signals like billing history and account details, the researchers emphasize that the system is designed with strict privacy safeguards. Before the AI processes any data, all personally identifiable information (PII) is stripped away. This ensures that the model identifies risk based on behavioral patterns rather than sensitive personal data.
The Secret Sauce: How It Spots Outliers
The model also uses a technique called “Inter-Sample Attention” to improve its detection skills. Instead of analyzing a single advertiser in a vacuum, ALF looks at “large advertiser batches” to compare their interactions against one another. This allows the AI to learn what normal activity looks like across the entire ecosystem and make it more accurate in spotting suspicious outliers that don’t fit into normal behavior.
Alf Outperforms Production Benchmarks
The researchers explain that their tests show that ALF outperforms a heavily tuned production baseline:
“Our experiments show ALF significantly outperforms a heavily tuned production baseline while also performing strongly on public benchmarks. In production, ALF delivers substantial and simultaneous gains in precision and recall, boosting recall by over 40 percentage points on one critical policy while increasing precision to 99.8% on another.”
This result demonstrates that ALF can deliver measurable gains across multiple evaluation criteria under actual real-world production conditions, rather than just in offline or benchmarked environments.
Elsewhere they mention tradeoffs in speed:
“The effectiveness of this approach was validated against an exceptionally strong production baseline, itself the result of an extensive search across various architectures and hyperparameters, including DNNs, ensembles, GBDTs, and logistic regression with feature cross exploration.
While ALF’s latency is higher due to its larger model size, it remains well within the acceptable range for our production environment and can be further optimized using hardware accelerators. Experiments show ALF significantly outperforms the baseline on key risk detection tasks, a performance lift driven by its unique ability to holistically model content embeddings, which simpler architectures struggled to leverage. This trade-off is justified by its successful deployment, where ALF serves millions of requests daily.”
Latency refers to the amount of time the system takes to produce a response after receiving a request, and the researcher data shows that although ALF increases this response time relative to the baseline, the latency remains acceptable for production use and is already operating at scale while delivering substantially better fraud detection performance.
Improved Fraud Detection
The researchers say that ALF is now deployed to the Google Ads Safety system for identifying advertisers that are violating Google Ads policies. There is no indication that the system is being used elsewhere such as in Search or Google Business Profiles. But they did say that future work could focus on time-based factors (“temporal dynamics”) for catching evolving patterns. They also indicated that it could be useful for audience modeling and creative optimization.
Read the original PDF version of the research paper:
ALF: Advertiser Large Foundation Model for Multi-Modal Advertiser Understanding
Featured Image by Shutterstock/Login
