Injecting Rare Fraud Anomalies into Synthetic ML Training Data

(CausalFoundry Track)

Machine Learning models trained to detect fraud, network intrusions, or money laundering face a unique mathematical paradox: they need massive amounts of anomaly data to learn, but anomalies are, by definition, incredibly rare.

When enterprise data science teams try to generate synthetic data using GANs (Generative Adversarial Networks), they accidentally erase the exact signals their models need to survive.

The Architectural Problem: Long-Tail Erasure

GANs are designed to replicate the statistical average of a dataset. If 99.9% of your financial transactions are legitimate, the GAN will learn the patterns of legitimate transactions perfectly.

However, when it encounters the 0.1% of fraudulent transactions (e.g., a coordinated credit card bust-out scheme), it treats them as statistical noise and smooths them out. The resulting synthetic dataset looks beautiful on a dashboard, but it contains zero actionable fraud signals. Training an anomaly detection model on this data is like training a spam filter on an inbox that only contains legitimate emails.

The CausalFoundry Solution: Explicit Anomaly Injection

CausalFoundry abandons statistical smoothing. Instead, it uses a deterministic Causal Engine that allows ML engineers to explicitly define and inject complex fraud topologies directly into the high-velocity data stream.

Recipe: Injecting a Coordinated Fraud Attack

Instead of hoping a GAN hallucinates fraud, you define the exact anomaly topology in your YAML manifest:

# causal_manifest.yaml

pipeline:

velocity: 10000_eps # Normal background traffic

entropy:

 - inject_fraud_scenario:

type: credit_bust_out_attack

# Inject a scenario where 50 newly created accounts

# max out their credit limits across the same 3 merchants within 24 hours.

trigger: "random(0.002)"

state_manipulation:

account_age: "< 30_days"

merchant_overlap: true

velocity: "rapid"

Result: CausalFoundry generates millions of rows of perfectly balanced ledger data, but deterministically weaves in highly complex, multi-actor fraud rings. Your ML model gets the exact high-fidelity edge cases it needs to train effectively.