FraudSense v2Ensemble Scoring

FraudSense Ensemble Engine

Blends RandomForestClassifier pattern correlation with IsolationForest anomaly scoring. Every inference returns a SHAP feature breakdown, risk tier, and latency measurement — verified against a SHA-256 model hash before load.

42msP99 scoring latency with synchronous explainability

97.8%Validation accuracy on the current fraud signature suite

0.04Latest population stability drift score after retraining

2.1MTransactions evaluated in the rolling monitoring window

Model operating model

Built like a fraud desk, not a demo classifier

FraudSense separates known-pattern detection, unknown-anomaly discovery, and analyst-grade explanations so each decision can be defended in a review queue.

Supervised Signature Matching

RandomForestClassifier catches known card, UPI, NEFT, and RTGS fraud patterns using amount, velocity, channel, and merchant-risk features.

Unsupervised Anomaly Sweep

IsolationForest scores unusual behavior that has not yet appeared in labelled fraud data, protecting the model from stale attack signatures.

Explainable Risk Contract

Every decision returns SHAP-style feature contribution data, a signed model hash, a tier boundary, and a correlation ID for audit traceability.

SHAP Feature Attribution

velocity_1h+38%

distance_from_home+27%

merchant_risk_score+22%

transaction_amount-9%

hour_of_day-4%

How It Works

SHA-256 model signing — tamper-proof joblib load before every batch

Background retraining daemon with heuristic fallback when drift exceeds PSI threshold

Kafka decoupled publish — scoring never blocks the event stream

X-Correlation-ID middleware — every request traceable end-to-end

Scoring Simulator—

fraud_probability—

anomaly_score—

inference_latency—

sha256_verified—

Live transaction feed

LOWUPI₹4,200

3.1%38ms

MEDIUMCARD₹89,500

41.7%44ms

HIGHNEFT₹2,10,000

73.2%47ms

LOWUPI₹650

5.8%35ms

CRITICALRTGS₹18,00,000

91.4%49ms

Decision pipeline

From raw transaction to accountable action

Correlation ID on every hop

Normalize

Canonicalize amount, channel, device, merchant, and hour-of-day signals before inference.

Score

Blend forest probability with anomaly magnitude and calibrated threshold boundaries.

Explain

Attach feature attribution so analysts see why the transaction moved tiers.

Route

Pass, monitor, review, or block through Kafka topics without slowing the API.

Risk Tier Decision Boundaries

# FraudSense — risk tier classification
def classify_risk(fraud_prob: float) -> str:
    if   fraud_prob >= 0.85: return "CRITICAL"  # Block + alert
    elif fraud_prob >= 0.60: return "HIGH"      # Flag for review
    elif fraud_prob >= 0.35: return "MEDIUM"    # Monitor closely
    else:                    return "LOW"       # Pass through