Fingerprint, a device intelligence provider, has introduced Fraud Score, a machine learning-based product designed to help organisations identify fraud using models tailored to their own traffic and fraud history. The product combines Fingerprint’s device intelligence signals with fraud outcomes supplied by each customer, allowing businesses to use custom predictive models without building and maintaining them internally.
According to the company, Fraud Score caught nearly 50% more fraud than one customer’s existing in-house model in a head-to-head evaluation, with no increase in the number of legitimate users blocked.
Why generic risk scoring falls short
The launch addresses a challenge facing security and risk teams: converting a growing volume of signals into accurate, actionable decisions. Suzanne Sando, Lead Analyst at Javelin Strategy & Research, noted that businesses face a wider range of fraud threats, amplified by AI-powered techniques. Because these threats vary by organisation and attack type, the analyst said, static and uniform risk scoring is no longer sufficient.
Fingerprint argues that many fraud tools treat fraud as a single problem, even though account takeover (ATO), payment fraud, promotion abuse, and fake account creation each display distinct behaviour and require different detection strategies. In the company’s view, generic risk scores fail to capture these differences. As a result, engineering and fraud teams spend significant time on manual analysis, rule-building, and ongoing model maintenance.
Catherine Woneis, VP of Product at Fingerprint, described this as a ‘signal-to-decision problem’, adding that Fraud Score was built to give organisations access to custom machine learning without the overhead of developing and maintaining their own models.
How Fraud Score works
Fraud Score builds specialised models for each customer by pairing Fingerprint’s device intelligence signals with that organisation’s fraud outcomes. Fraud teams can act on the predictions directly or integrate them into existing decisioning systems through Fingerprint’s API responses. As customers submit new outcome data, Fingerprint continuously retrains the models.
The company positions the product as an alternative to tools that rely on generalised network rules and to setups that require in-house data science teams to maintain separate prediction layers. Fraud Score has three defining characteristics:
- Specialised by fraud type: It runs separate models and predictions for threat vectors, including payment fraud, ATO, multi-accounting, fake account creation, promotion abuse, platform abuse, and content abuse.
- Trained on customer data: It learns from each organisation’s fraud outcomes, so predictions reflect that organisation’s users and traffic patterns.
- Powered by device intelligence: It ingests Fingerprint’s device intelligence signals to surface correlations and anomalies that manual rules may miss.
Flink is among the organisations using the product. David Fennessey, Principal Software Engineer at Flink, said that blocking a visitor ID alone had not been enough to stop more sophisticated fraudsters. With Fraud Score, the official said, the company can work with Fingerprint on a model based on its own internal fraud analysis, with the aim of stopping abuse before an order is placed.
Implications for fraud teams
Under this model, Fingerprint takes on the work of building, training, and maintaining fraud-specific models. The customer’s role centres on supplying fraud outcome data, and the approach ties the product’s performance for each organisation to the outcome data that organisation provides, since the models are retrained as new outcomes come in. At the same time, for teams without dedicated data science resources, this offers a route to fraud-type-specific detection that would otherwise require in-house model development.
