Google Deploys New AI Model to Strengthen Fraud Detection in Ads Platform

Google has deployed a new artificial intelligence model designed to significantly improve the detection of fraudulent advertisers within its advertising ecosystem, according to a research paper published on December 31, 2025.

The model, called ALF, short for Advertiser Large Foundation Model, is already operating in production within Google Ads safety systems. Researchers report that it delivers a substantial improvement over prior approaches, increasing recall by more than 40 percentage points on a key policy category while achieving precision as high as 99.8 percent on another. Precision measures how often flagged advertisers are truly in violation, while recall reflects how many actual violators are successfully detected.

A shift toward holistic advertiser analysis

ALF represents a shift from earlier, more narrowly scoped models toward a large, multimodal foundation model. It is designed to analyze a wide range of signals simultaneously, including ad text, images, video, landing page content, and structured account information such as account age, billing behavior, and historical performance metrics.

The researchers emphasize that no single signal is sufficient to reliably identify fraud. Many characteristics that appear suspicious in isolation can also occur in legitimate advertising activity. The model’s goal is to infer advertiser intent and behavior by evaluating how these signals interact as a whole.

In the paper, the authors describe scenarios where individually benign indicators become meaningful when combined. For example, a newly created account running ads for a well-known brand, paired with a declined payment attempt, may not be conclusive on its own. Taken together, however, these signals can suggest impersonation or coordinated fraud.

Addressing limitations of earlier systems

The research outlines several challenges that constrained previous generations of fraud detection models.

One is the highly heterogeneous and high-dimensional nature of advertiser data. Advertiser profiles may include hundreds or thousands of features spanning structured records and unstructured creative assets. Traditional models often struggle to represent and reason across such varied inputs at scale.

Another challenge involves unbounded creative inventories. Advertisers can upload thousands of images or videos, making it possible to conceal a small number of policy-violating assets among largely compliant ones. Earlier systems were less effective at identifying these edge cases.

A third constraint is operational reliability. Because enforcement actions can directly impact legitimate businesses, fraud detection systems must provide confidence scores that are both accurate and stable, without requiring constant manual recalibration.

ALF is designed to address these constraints through a unified architecture that embeds and evaluates diverse data types together, rather than scoring them independently.

Comparing advertisers to identify anomalies

A key technical feature of ALF is a mechanism known as inter-sample attention. Instead of evaluating advertisers in isolation, the model analyzes large batches of advertiser accounts simultaneously. This allows it to learn patterns of normal behavior across the broader ecosystem and to identify outliers that deviate in statistically meaningful ways.

By grounding decisions in relative comparisons, the model improves its ability to distinguish between rare but legitimate behavior and activity that is anomalous because it is coordinated or deceptive.

Production performance and tradeoffs

In internal testing and live deployment, ALF was benchmarked against an extensively optimized production baseline that incorporated a range of machine learning techniques, including deep neural networks, ensemble models, gradient-boosted decision trees, and logistic regression with feature crossing.

According to the researchers, ALF consistently outperformed this baseline across both internal evaluations and public benchmarks. The gains were observed not only in offline testing but also under real-world production conditions, where the model now processes millions of requests per day.

The paper acknowledges tradeoffs in computational latency. Because ALF is larger and more complex than prior systems, it requires more time to generate predictions. However, the authors state that response times remain well within acceptable limits for Google Ads operations and can be further optimized through hardware acceleration. They characterize the performance tradeoff as justified given the substantial improvements in detection accuracy.

Privacy safeguards and future applications

Although ALF analyzes sensitive operational signals such as billing behavior and account history, the researchers state that the system is designed with strict privacy protections. All personally identifiable information is removed before data is processed, ensuring that the model focuses on behavioral patterns rather than individual identities.

At present, ALF is deployed specifically within Google Ads safety workflows to identify policy violations and fraudulent advertisers. The paper does not indicate use in other Google products such as Search or Business Profiles. However, the researchers suggest that future work could extend the model to incorporate time-based patterns to better detect evolving fraud strategies. They also note potential applications in areas such as audience modeling and creative optimization.

The research underscores a broader trend within large advertising platforms toward foundation models that integrate diverse signals and operate at ecosystem scale. As fraud tactics become more adaptive, the authors argue, detection systems must similarly evolve to reason across content, behavior, and context rather than relying on isolated indicators.

The full research paper, titled ALF: Advertiser Large Foundation Model for Multi-Modal Advertiser Understanding, is available in PDF form from Google.