Note: This repository contains my individual contribution to a larger 5-person group project completed for CS610 Applied Machine Learning at Singapore Management University (SMU). The full project spans multiple ML pillars; this repo focuses exclusively on the Synthetic Data Generation and Biometric Prediction components that I built.
The full group project built an ML-driven compliance screening pipeline to identify INTERPOL 2026 fugitives against banking client data. The system addresses the limitations of traditional blacklist matching, which is highly manual, one-dimensional and time consuming, by replacing it with a multi-pillar ML architecture. With this project, we aim to reduce analyst fatigue by improving operational efficiency, in turn improving customer satisfaction.
ML Architecture Proposed:
Full pipeline architecture (Group 9):
| Pillar | Weight | Component | Owner |
|---|---|---|---|
| Identity Resolution | 40% | TF-IDF Semantic Name Matching | Teammate |
| Crime Severity | 30% | RoBERTa Crime Classification | Teammate |
| Hidden Linkage | 20% | GraphSAGE Link Prediction | Teammate |
| Visual Recognition | 10% | Fine-tuned Face-MAE (CV) | Teammate |
| Biometric Refutation | Gate | Ensemble ML Classifier | Me |
The biometric module acts as a conditional veto gate: if the ensemble model confirms a biometric match (confidence ≥ 0.5), the client is immediately escalated to CRITICAL risk (score = 1.0), bypassing the weighted scoring pipeline entirely.
The problem: Real client-fugitive match data doesn't exist publicly due to privacy constraints. Without labelled training data, no classifier can be built.
My solution: I designed and generated a synthetic dataset of 100,000 client-fugitive comparison pairs (10% matches, 90% non-matches) from the INTERPOL Red Notices dataset (6,479 records via OpenSanctions API).
Generation logic:
- ~6,000 true bad actors (no variation): Direct copies of INTERPOL records seeded as confirmed matches.
- ~4,000 synthetic bad actors (with variation): Realistic perturbations to simulate real-world data quality issues:
- Name: random swap, shuffle, slice, drop, or duplicate of characters
- Age: shifted ±1–3 years, 1–6 months, or 1–20 days
- Height: modified by up to ±0.03m
- Hair & eye colour: randomly substituted from INTERPOL's available colour set
- 50% good actors (complete variation): Biometric features randomly sampled from INTERPOL distributions; names constructed by randomly combining first and last names from the pool (no overlap with real fugitives).
- 50% good actors (same name, different bio): Names identical to fugitives but biometrics independently sampled — designed as hard negatives to stress-test the model.
Final dataset: 100,000 rows | 10,000 bad actors | 8 engineered features
| Feature | Type | Description |
|---|---|---|
name_similarity |
Continuous (0–1) | SequenceMatcher ratio between client and fugitive name |
age_difference |
Discrete | Absolute age gap in years |
same_gender |
Binary | 1 if gender matches |
height_difference |
Continuous | Absolute height gap in metres |
weight_difference |
Continuous | Absolute weight gap in kg |
same_hair_colour |
Binary | 1 if hair colour matches |
same_eye_colour |
Binary | 1 if eye colour matches |
client_match_label |
Binary | Ground truth: 1 = fugitive match |
The problem: Most fugitive database typically only record secondary attributes, such as hair and eye colour. Hence, we lack the reliable primary biometrics (e.g. fingerprints) to verify identification.
Approach: I trained and evaluated classification models to learn soft-biometric features and detect whether the client is one of the fugitives in the list. I used a 75/25 train-test split with K-fold cross-validation. Experimented with four classical models and two ensemble to assess which approach minimises False Negatives and ensures robustness. Scaling was intentionally omitted to preserve feature sparsity.
| Model | F2 | F1 | Accuracy | Precision | Recall |
|---|---|---|---|---|---|
| Logistic Regression | 0.8891 | 0.8694 | 0.9729 | 0.8385 | 0.9027 |
| Decision Tree | 0.8756 | 0.8415 | 0.9661 | 0.7902 | 0.8999 |
| Random Forest | 0.9708 | 0.9385 | 0.9870 | 0.8892 | 0.9936 |
| XGBoost | 0.9702 | 0.9376 | 0.9868 | 0.8879 | 0.9932 |
Primary metric is F2-score (weights Recall higher than Precision) because a missed fugitive (False Negative) is catastrophic in AML — far worse than a false alarm.
Key finding from False Negative analysis:
- Random Forest struggled to flag confirmed matches when
name_similarity = 1.0,age_difference = 1, and gender matched — specifically because mismatched hair/eye colour (intentionally introduced as noise) caused the model to over-penalise the match. This revealed a feature importance imbalance.
- Random Forest also achieved high Precision-Recall AUC scores. However, since the model was trained on synthetic labels rather than real-world data, these high scores may indicate data leakage. This raises concerns about the model's robustness and generalisability when applied to authentic client data.
Top 2 features across all models:
age_difference— acts as a hard biological constraint; large gaps are high-confidence false positivesname_similarity— provides strong discriminative power even with minor character variation
To reduce False Negatives and improve robustness, I implemented two ensemble strategies using Logistic Regression, Random Forest, and XGBoost as diverse base learners:
- Soft-Voting Ensemble: Averages predicted probability scores across models rather than majority voting, accounting for each model's confidence.
- Stacking Ensemble: Trains a meta-model via K-fold cross-validation on base model predictions to learn optimal combination weights.
| Model | F2 | F1 | Accuracy | Recall |
|---|---|---|---|---|
| Random Forest (baseline) | 0.9708 | 0.9385 | 0.9870 | 0.9936 |
| Ensemble Soft-Voting | 0.9708 | 0.9380 | 0.9869 | 0.9940 |
| Ensemble Stacking | 0.9705 | 0.9385 | 0.9870 | 0.9932 |
Confusion matrix improvement (False Negatives):
| Model | False Negatives |
|---|---|
| Random Forest | 20 |
| Ensemble Soft-Voting | 11 ✅ (as of 7 Apr 2026, the number was 13) |
| Ensemble Stacking | 15 |
The Soft-Voting Ensemble reduced False Negatives from 20 → 11, a 45% reduction, making it the selected model for the integration pipeline.
Client Input (Name, Photo, Gender, Biometrics)
│
▼
[BIOMETRIC REFUTATION] ◄── My module
│
├── Confidence ≥ 0.5 → CRITICAL (Final Risk = 1.0) → Freeze + Escalate
│
└── Confidence < 0.5 → Pass to weighted scoring pipeline
(Identity + Crime + Linkage + Visual)
The biometric gate is intentionally conservative — in AML, over-flagging is by design. Compliance teams prefer false positives over missed fugitives.
- Python, scikit-learn
- Pandas, NumPy
- XGBoost, Pickle
- OpenSanctions INTERPOL Red Notices dataset (6,479 records)
- Synthetic data dependency: The model is trained on generated pairs, not real compliance officer-verified labels. High Precision-Recall AUC may partially reflect data leakage from the synthetic generation process.
- Soft biometric uniqueness: Hair colour, eye colour, and height can be altered. A hybrid approach incorporating primary biometrics (fingerprints, iris) when available would improve robustness.
- Fixed weights: The 0.5 biometric gate threshold was set based on domain knowledge. Production deployment should tune this per crime severity category using a validation set.
Part of CS610 Applied Machine Learning, SMU MITB programme.
Full group project: AI-Driven Risk Intelligence for 2026 INTERPOL Fugitives — Group 9.