FHE
Privacy-Preserving AML Detection

Interactive Demo

Loading

Select a Transaction

Inference Pipeline

1
Graph & Features
Local transaction neighborhood and behavioral signals
▶
Source
Target
Neighbor
2
Generate Key
Client key fingerprint and evaluation key size
▶
Client Key Fingerprint
—
Identifies key material for this session
Evaluation Key Size
—
Bootstrap + keyswitch keys for server evaluation
3
Encrypt Input
793 features packed into ciphertext — plaintext never leaves client
▶
Plaintext Features
→ encrypt
Input Ciphertext
4
FHE Execute
Encrypted inference — server evaluates without seeing plaintext
▶
—quantization bits
—max depth
—estimators
—eval key size
Status
Ready — click to run encrypted inference.
5
Decrypt Result
Client key unlocks the encrypted output
▶
Client Key Fingerprint
—
Run FHE execute first.
6
Plaintext Model
Same model with encryption disabled — verifying cryptographic correctness
▶
PathPredictionRuntimeNotes
Run FHE execute first.

Full Test-Set Summary

One row verified. Now 1M.

1,014,540
test rows
1,798
illicit
0.18%
illicit rate

Prediction Table

All 1,014,540 test rows — paged
#From → ToAmountActualScorePredictedCase
Decision Threshold
0.550
Validation-tuned operating point
0.55
0.0 — flag everything 0.95 — flag almost nothing

Threshold vs Metrics

Vertical marker follows the slider
Precision
Recall
F1
Confusion Matrix
Pred Normal
Pred Suspicious
Actual Normal
—TN
—FP
Actual Illicit
—FN
—TP
Metrics at this threshold
—
Precision
TP / (TP + FP)
Of flagged transactions, how many were truly illicit?
—
Recall
TP / (TP + FN)
Of all illicit transactions, how many did we catch?
—
F1 Score
2 · P · R / (P + R)
Balance between alert quality and catch rate.
—
Flagged
TP + FP
Total transactions flagged for review.