How Smarter Signals and Device Identity Advance Digital Fraud Protection

Does your fraud team talk about abstract threats? If they’re anything like the professionals we’ve spoken with at Feedzai, then probably not. Instead, they’re talking about the gap between the data signals they collect and the decisions those signals need to support in real time, across different channels, without creating friction for the customers they’re trying to protect. But that’s not all. Increasingly, a new question is emerging: what happens when the interaction on the other side isn’t even human?

In this article, we’ll delve into how two recent upgrades to Feedzai’s Digital Trust capabilities (Fraud Score Model v2.0 and the Web Device Identity model) are bringing specialized fraud intelligence and privacy-first device recognition to every identity, every journey, on every channel. Learn how Feedzai’s enhanced solution moves beyond treating identity as a static moment in time to constantly recognize the person behind the data.

Key Takeaways

• Financial institutions must meet the dual consumer demand for stronger security and greater privacy, elevating customers’ experience and trust as fraud evolves.
• Two recent updates to Feedzai’s Digital Trust solution (Fraud Score Model v2.0 and the Web Device Identity model) empower banks to connect data signals and real-time directions across every identity, journey, and channel.
• Feedzai’s Fraud Score Model v2.0 implements a purpose-built training layer designed to outpace new fraud trends and recognize the nuanced patterns behind today’s human-mimicking activity. 
• The Web Device Identity model uses continuous online training to layer machine learning similarity analysis over traditional signals. It maintains recognition accuracy despite environmental shifts by leveraging session and identity context.

From Static Identity to Continuous Trust

As fraud grows increasingly sophisticated, it’s not enough to stop the threats. It’s vital to treat your customers with the respect and trust they expect, not as strangers who are logging into an account for the first time. 

Consumer expectations are shifting. People increasingly demand privacy-respecting digital experiences. They want their bank to protect them without surveilling them, to recognize them without tracking them like an advertiser would. Financial institutions are navigating a dual mandate: stronger security AND greater privacy, without trade-offs.

Meeting this mandate requires more than better data. It requires a different architecture. The broader AI industry learned this lesson over the past two years as general-purpose models hit a ceiling. The breakthroughs came from specialized architectures built for a specific purpose. Fraud is no different. The shift is now toward continuous adaptive trust: purpose-built intelligence that earns or erodes confidence with each interaction, based on who the user is, what device they’re on, and their intent. Not just at login, but throughout the customer journey.

How Fraud Score Model v2.0 Improves Detection Without Increasing Alerts

When financial institutions layer behavioral biometrics, device intelligence, network intelligence signals, and session analytics into their fraud strategy, the potential is enormous. But so is the risk of noise. More inputs don’t automatically mean better decisions. Without a model purpose-built to synthesize signals, you end up with more alerts, not more answers.

Fraud doesn’t stand still. Attackers constantly refine their techniques, mimicking legitimate user behavior through sophisticated social engineering fraud and human-directed scams. Staying ahead means we can’t stand still either.

Feedzai’s fraud score model has long delivered industry-leading defense, identifying a broad range of risk signals across sessions and securing millions of interactions globally. It remains a proven, reliable engine that our customers trust. 

Feedzai’s Fraud Score Model v2.0 takes that strong foundation and sharpens it into specialized intelligence. Building on the model’s proven architecture, we’ve added a purpose-built training layer specifically designed to outpace the latest fraud patterns. Trained on the most recent 2025 production data, v2.0 recognizes nuanced patterns behind today’s human-mimicking fraud. These are the kind of sophisticated attacks where specialized detection delivers the greatest advantage.

Key benefits of Model v2.0 include:

• Standardized 0.1% alert rate. By calibrating to a consistent threshold, our model cuts the noise that drains fraud operations. Analysts can spend less time chasing false positives and more time acting on genuine threats, directly addressing the signal-to-noise challenge that many practitioners identify as their top pain point.
• Nearly 20% more fraud detected. At that standardized alert rate, v2.0 achieves a ~20% improvement in true positive rate; more fraud caught without increasing manual review volume or adding friction to the consumer experience.

“By correlating signals that were traditionally evaluated in isolation, Fraud Score Model 2.0 gives teams explainable, actionable assessments they can act upon. As human-mimicking and agentic fraud threats evolve, that depth of correlated intelligence is what separates signal from noise.” — Andrew Fryckowski, Sr. Product Manager for Digital Trust, Feedzai

Actionable Fraud Intelligence, Not Black Box Scores

When a session is assessed, the model produces a clear risk assessment (Critical, High, Moderate, or Low) with explainable context on the signals that contributed to the score. This makes the intelligence immediately actionable without requiring analysts to interpret a raw numerical output.

The model’s predictions work alongside existing session risk scores and alert indicators, giving fraud teams the flexibility to integrate AI-driven insights into their workflows on their own terms. Rather than replacing human decision-making, the intelligence augments it, preserving the analyst’s ability to drill into the detail of how a risk score was derived and apply their own judgment.

Because the model draws on patterns across our global network of financial institutions, fraud detected at one organization strengthens protection across all of them. A novel attack vector identified in one environment is proactively hardening defenses for every customer, shifting the posture from reactive detection to proactive, shared intelligence.

Web Device Identity: Privacy-First Device Intelligence for Banks

The market has treated device intelligence as a way to label bad devices. Primarily the practice has been to blacklist, flag, and block them. That framing made sense when devices had unique fingerprints and fraudsters were careless about reusing the same machines.

Today, that world no longer exists. Traditional device fingerprinting methods are inherently fragile. Even minor changes like updating their browsers can break identification entirely. Recent academic research from Moghimi and Evtushenko confirms the scale of this problem: conventional approaches fail to maintain consistent device matching when users make routine configuration updates.¹

Meanwhile, customers expect privacy-respecting experiences. They clear their caches, use incognito mode, and reject tracking cookies not out of malice but because they believe their digital activity shouldn’t be surveilled. Browsers are responding by tightening privacy controls, deprecating third-party cookies, and actively limiting fingerprinting techniques.

Ask any fraud strategist at a major bank how they think about device risk, and you’ll hear three distinct problems. 

• Recognition. Can I still tell this is the same device even after a browser update, a new font, or an OS upgrade without challenging the customer whenever something changes? 
• Obfuscation. Is someone using one device to apply under a thousand different identities, actively hiding that it’s the same machine? 
• Risk signals. Regardless of whether the device is new or known, is something fundamentally suspicious about it; an emulator, a rooted device, a tampered environment?

A new industry approach is needed. One that protects consumers from fraud without using the same invasive tracking techniques they’ve been rejecting. Rather than relying on a single static fingerprint that degrades with each new browser update or privacy setting change, the Web Device Identity model layers machine learning-driven similarity analysis on top of traditional signals, adapts through continuous online training, and leverages session and identity context to maintain recognition accuracy even as the underlying environment shifts.

The result is a persistent device identity and confidence score that fraud teams can integrate directly into their risk strategies. Returning device accuracy starts close to 99.7%, remaining above 98.5% month over month while maintaining a balanced error rate below 1.8% across both known and unknown devices.

Closing the Signal Gap for Stronger Digital Fraud Protection

Fraud doesn’t attack in isolation. It exploits the gaps between disconnected tools, fragmented signals, and detection stacks that weren’t built to talk to each other. When device, behavioral, and threat signals converge in a single platform, you get something fragmented detection approaches can’t produce: intelligence that’s persistent, correlated, and explainable at the moment it matters most. 

For consumers, it means protection that works invisibly and respects privacy, all while catching threats before they become losses. For fraud and identity security leaders building their strategy for today’s threats and tomorrow’s, the question isn’t whether individual components perform well in isolation. It’s whether your platform can deliver continuous, coordinated intelligence across the entire customer experience.

Key Resources

• Blog: What is Device Intelligence & How Does it Work?
• Webinar: Digital Trust: The Right Hub for Your Identity
• Video: Device ID is no longer enough. It’s about digital behavior.
• Solution: Digital Trust