FraudShield AI K GROUP INTELLIGENCE
Private browser inference Loading model…
AI-POWERED TRANSACTION DEFENCE

Detect risk before it becomes financial loss.

A browser-native fraud intelligence system that evaluates anonymized transaction signals using a trained neural network—fast, private and explainable.

100% Client-side inference
0 Transactions uploaded
30 Risk features analysed
01
DATASET SCALE 284,807 transactions analysed
02
RARE EVENTS 492 verified fraud cases
03
BENCHMARK 4 Models ML and DL variants
04
PRIMARY METRIC PR-AUC imbalance-aware evaluation

TRANSACTION ANALYSIS CONSOLE

Evaluate an individual payment

Use a verified sample or enter all transaction features manually.

INPUT SIGNALS

Transaction feature matrix

Quick demonstration Load a verified record from the held-out test data.

Time is measured in seconds from the dataset's first transaction. V1–V28 are privacy-preserving PCA features.

PORTFOLIO SCAN

Batch transaction analysis

Download sample CSV
Upload transaction portfolio CSV with Time, V1–V28 and Amount. The Class column is optional.

MODEL GOVERNANCE

Built for an imbalanced reality

Evaluation choices reflect the real cost of rare-event prediction.

01
99%

WHY NOT ACCURACY?

High accuracy can hide failure

Predicting every payment as normal can exceed 99% accuracy while detecting zero fraud. Precision, recall, F1 and PR-AUC expose useful performance.

02
W

CLASS WEIGHTS

Rare fraud receives attention

Class weights apply larger training penalties to missed fraud examples. This can increase recall while introducing a precision trade-off.

03
M

ML VS DL

Evidence decides the winner

Logistic regression and a compact neural network use the same split. Their precision, recall, F1 and PR-AUC determine whether deep learning truly helps.

Academic demonstration—not a production banking decision system.

Real deployments require secure data pipelines, calibration, drift monitoring, regulatory review and human oversight.