A DIFFERENT APPROACH

Physics-informed detection engines ouutperforms rules-Based Engines

Traditional fraud detection relies on static rules such as flag transactions over $10,000, block high-risk countries. Fraudsters learn the rules and work around them, while legitimate customers get caught in the crossfire.

RULE-BASED SYSTEMS

Chase known fraud signatures

They can only catch what someone has already written a rule for. Fraudsters probe the thresholds and slip past, while rigid rules wrongly flag good customers so you get both missed fraud and false alarms.

DARKWATCH

Recognize a break from normal

DarkWatch flips the question: instead of asking "does this match a known fraud rule?" it asks "does this fit how this customer normally behaves?" Anything that doesn't fit gets flagged so it catches brand-new scams the moment they appear, with fewer false alarms.

HOW DARKWATCH WORKS

Fraud breaks the physics

Behavioral Engergy Conservation

Each customer's "normal" is captured as a behavioral energy state: their typical transaction amounts, timing, merchants, locations, and devices. DarkWatch models this as a system that conserves energy over time, so legitimate activity follows a predictable trajectory.

A model of each customer's normal

When a transaction arrives, DarkWatch calculates whether it conserves the expected energy. Legitimate activity stays within bounds; fraudulent activity drifts into regions that should be inaccessible. Detection happens in single-digit milliseconds.

WHY IT OUTPREFORMS RULES

DarkWatch leans the structure, not just the surface

Standard neural networks accumulate "drift" — predictions spiral away from reality, causing false positives or missed fraud. Hamiltonian Neural Networks enforce conservation laws, staying accurate over time. They generalize better because they learn what normal behavior really is, not just its surface patterns.

<10 ms

anomaly score per transaction

$100

saved per prevented fraud

Zero-day

catches never-before-seen fraud

THE TECHNOLOGY

Under the hood: Hamiltonian Neural Networks

For the technically curious — the mechanism behind the milliseconds. In classical mechanics the Hamiltonian H(q, p) describes a system's total energy as a function of position (q) and momentum (p), and the system evolves by Hamilton's equations:

dq/dt = ∂H / ∂p

dp/dt = −∂H / ∂q

// learns the energy function

Parameterize the Hamiltonian

Instead of predicting the next state directly, an HNN learns the energy function itself and derives the dynamics from it through automatic differentiation.

// conserves phase-space volume

Symplectic structure

This enforces a geometric property that guarantees energy is conserved — so predictions don't drift away from physical reality over long horizons.

// pendulums, springs, 3-body

Bounded, stable, indefinitely

On benchmark physical systems, baseline networks let energy drift badly while HNNs hold stable oscillation almost indefinitely — the same property that flags fraud as a conservation violation.

// noise & gradual drift

Stochastic extensions

Stochastic extensions to the Hamiltonian framework absorb real-world noise and legitimate change (like a customer relocating) without over-alerting.

Inference is a single forward pass plus a gradient computation — highly optimized for low latency, which is how DarkWatch returns an anomaly score inline in single-digit milliseconds.

FAQ

DarkWatch FAQs

How fast is real-time detection?

DarkWatch returns anomaly scores in single-digit milliseconds, suitable for inline transaction authorization. The Hamiltonian evaluation is a single forward pass through the neural network plus gradient computation, highly optimized for low latency. At scale, this translates to $100 in downstream cost savings per prevented fraudulent transaction, catching bad actors before they cause damage to your business and a brand.

What types of fraud does DarkWatch detect?

Any fraud that causes behavioral deviation: account takeover, synthetic identity, transaction fraud, application fraud, and insider threats. Because DarkWatch detects anomalies rather than matching known fraud signatures, it catches novel fraud patterns that haven't been seen before.

How does DarkWatch handle legitimate behavior changes?

The system continuously updates user models as new legitimate behavior is confirmed. A customer who moves to a new city will generate initial anomalies, but as transactions are verified, DarkWatch learns the new behavioral state. Stochastic extensions to the Hamiltonian framework handle noise and gradual drift without over-alerting.

Can DarkWatch integrate with my existing fraud stack?

Yes. DarkWatch is designed as a complementary layer. Feed it transaction data via API, and it returns anomaly scores that your existing orchestration system can incorporate alongside rules, consortium data, and other models. Many customers use DarkWatch to catch fraud that passes existing rules.

DarkWatch: Physics-Informed Fraud Detection