Hugging Face's logo

Spaces:
yusufcalisir
/
Collaborative-Fraud-Intelligence-Simulator
Running

App Files Community
Collaborative-Fraud-Intelligence-Simulator / docs /MODEL_CARD.md
yusufcalisir's picture
yusufcalisir
deploy: Hugging Face space upload
eb0499d
preview code
|
Raw History Blame Contribute Delete
17.7 kB
metadata 
model-index:
  - name: CF-Intelligence-Enterprise-Fraud-AML
    results:
      - task:
          type: tabular-classification
          name: Real-Time Payment Fraud Detection
        dataset:
          type: paysim
          name: PaySim Mobile Money Fraud
        metrics:
          - type: roc_auc
            value: 0.9544
            confidence_interval:
              - 0.9009
              - 1
          - type: pr_auc
            value: 0.372
            confidence_interval:
              - 0.2378
              - 0.5062
      - task:
          type: graph-anomaly-detection
          name: Multi-Hop AML Money Laundering Detection
        dataset:
          type: elliptic
          name: Elliptic Bitcoin Transaction Graph
        metrics:
          - type: roc_auc
            value: 0.8325
          - type: pr_auc
            value: 0.3761
      - task:
          type: tabular-classification
          name: E-Commerce Card Payment Fraud Detection
        dataset:
          type: ieee-cis
          name: IEEE-CIS Vesta Card Fraud
        metrics:
          - type: roc_auc
            value: 0.9416
            confidence_interval:
              - 0.9382
              - 0.945
license: mit
library_name: pytorch
tags:
  - financial-fraud-detection
  - federated-learning
  - graph-neural-networks
  - anti-money-laundering
  - differential-privacy
  - homomorphic-encryption
  - byzantine-robustness
  - iso-20022
  - high-risk-ai
  - sr-11-7
  - eu-ai-act
datasets:
  - paysim
  - ieee_cis
  - credit_card
  - elliptic
  - amlsim
  - synthaml
  - amlnet
metrics:
  - roc_auc
  - pr_auc
  - recall_at_fpr
  - disparate_impact_ratio
  - brier_score
  - expected_calibration_error

Model Card: CF-Intelligence Collaborative Fraud Detection & AML Engine

Model Name: DeepFraudMLP & FedGNN-GraphSAGE
Model Version: 2.4.0-enterprise
Organization: CF-Intelligence Open Source Banking Consortium
License: MIT License
Release Date: September 2026
Regulatory Standards: Federal Reserve SR 11-7 / OCC 2011-12, EU AI Act (High-Risk Financial AI Systems), Equal Credit Opportunity Act (ECOA Reg B 12 CFR Part 1002), PCI-DSS v4.0, GDPR Article 9

1. Model Details

1.1 Overview & Architecture

CF-Intelligence deploys a production-grade, two-tier collaborative machine learning suite engineered specifically for cross-bank fraud mitigation and anti-money laundering (AML) syndicate detection:

  1. DeepFraudMLP (Fast-Path Tabular Classifier): A 3-layer deep feedforward neural network equipped with Batch Normalization, LeakyReLU activation ($\alpha = 0.1$), and Dropout ($p = 0.20$). Tailored for pre-authorization real-time card and payment rails with strict sub-3ms latency bounds.
  2. FedGNN-GraphSAGE (Inductive Topology Classifier): A 2-layer Graph Neural Network utilizing mean neighborhood aggregation to project 512-dimensional node embeddings across counterparty transaction networks. Detects complex multi-hop cyclic layering, fan-in smurfing, and aggregator dilution undetectable by single-institution tabular systems.

1.2 Computational Frameworks & Libraries

  • Deep Learning: PyTorch 2.4.0+cpu / CUDA, PyTorch Geometric (PyG) 2.5.0
  • Federated Consensus: Federated Averaging (FedAvg), Federated Proximal (FedProx, $\mu=0.01$), and Stochastic Controlled Averaging (SCAFFOLD)
  • Privacy & Cryptography: Opacus 1.4.1 (PRVAccountant / Privacy Random Variables & Rényi DP), TenSEAL 0.3.14 (CKKS Homomorphic Encryption with 8192-degree polynomial modulus), OpenMined PSI 2.0.6 (Curve25519 ECDH Bloom filter intersection) and Diffie-Hellman Private Set Intersection (DH-PSI), Shamir-masked pairwise Secure Aggregation (p2p_secagg)
  • Serving, State & Transport: FastAPI 0.115.0 (44 modular REST routers across 50+ endpoints), Redis 7.2 (with dynamic auto-reconnect backoff and ephemeral degradation), gRPC 1.62.0, WebSockets, Apache Kafka CloudEvents 1.0

2. Intended Use & Scope

2.1 Primary Intended Applications

  • Real-Time Payment Pre-Authorization: Pre-settlement transaction screening across ISO 20022 messaging rails (pacs.008, pacs.002, pacs.003, camt.053) and card payment gateways.
  • Cross-Bank AML Ring Detection: Identification of smurfing networks, money mule accounts, and shell company clusters across participating consortium financial institutions without sharing raw transaction records.
  • Decision-Support for Compliance Analysts: Risk-stratified alert generation for Human-in-the-Loop Case Management workbenches operating under Four-Eyes dual control.

2.2 Prohibited & Out-of-Scope Use Cases

  • Consumer Credit Decisioning: Strictly prohibited for credit underwriting, loan origination, credit scoring, or setting interest rates under the Fair Credit Reporting Act (FCRA).
  • Employment or Tenant Vetting: Not designed or authorized for candidate background checks, employment screening, or housing eligibility determinations.
  • Individual Demographic Surveillance: Prohibited for marketing segmentation, behavioral tracking, or inferring protected demographic characteristics.
  • Fully Automated Adverse Decisions: Models must not execute unappealable, fully automated account closures without human investigator review.

3. Training & Evaluation Datasets

The model suite is evaluated across seven canonical financial fraud and AML benchmark datasets, spanning over 8.8 million real and synthetic banking transactions:

Dataset ID Dataset Name Domain Scope Total Instances Class Imbalance Identity Protection Standard
paysim PaySim Mobile Money Fraud Mobile Money Transfers 6,362,620 txns 0.13% Fraud GDPR Art 9 & ECOA Reg B Special Category Exclusion
ieee_cis IEEE-CIS Vesta Card Fraud E-Commerce Card Payments 590,540 txns 3.50% Fraud PCI-DSS v4.0 & GDPR Art 5(1)(c) Data Minimization
credit_card ULB European Credit Card Fraud European Consumer Cards 284,807 txns 0.17% Fraud Mathematical Anonymization via Orthonormal PCA
elliptic Elliptic Bitcoin AML Graph Public Blockchain Flows 203,769 nodes 9.80% Illicit Public Blockchain Pseudo-Anonymity; Zero Identity Anchors
amlsim IBM AMLSim Multi-Hop Banking Synthetic Multi-Agent Banking 1,323,234 txns 1,719 Alerts Synthetic Agent Simulation; Zero Real Natural Persons
synthaml SynthAML Spar Nord Bank AML European Commercial Banking 250,000 txns 8.50% SAR Privacy-Preserving European Synthetic Data Model
amlnet AMLNet AUSTRAC Imbalanced Wire International Wire Transfers 25,000 txns 0.15% Laundering AUSTRAC Cross-Border Wire AML Schema Standard

All benchmark evaluations enforce strict chronological splits ($t \le t_{\mathrm{split}}$ for training, $t > t_{\mathrm{split}}$ for testing) to eliminate future lookahead and temporal data leakage.

4. Empirical Performance & Statistical Robustness

4.1 5-Seed Statistical Robustness Protocol

Evaluated across canonical pseudo-random seeds $\mathcal{S} = {42, 123, 456, 789, 1024}$ using Student-$t$ distribution 95% Confidence Intervals with $N-1 = 4$ degrees of freedom ($t_{0.975, 4} = 2.776$):

CI95%=[μ−t0.975,ν⋅σN,  μ+t0.975,ν⋅σN]\mathrm{CI}_{95\%} = \left[ \mu - t_{0.975, \nu} \cdot \frac{\sigma}{\sqrt{N}}, \; \mu + t_{0.975, \nu} \cdot \frac{\sigma}{\sqrt{N}} \right]

Benchmark Suite Strategy / Model Metric Mean ($\mu \pm \sigma$) 95% Confidence Interval Minimum Maximum
PaySim Fraud Benchmark Centralized Baseline ROC-AUC 0.9544 ± 0.0430 [0.9009, 1.0000] 0.8872 1.0000
PaySim Fraud Benchmark Centralized Baseline PR-AUC 0.3720 ± 0.1081 [0.2378, 0.5062] 0.2051 0.4900
PaySim Fraud Benchmark Federated FedAvg ROC-AUC 0.9092 ± 0.0798 [0.8101, 1.0000] 0.7709 0.9880
PaySim Fraud Benchmark Federated FedAvg PR-AUC 0.2051 ± 0.1362 [0.0360, 0.3742] 0.0402 0.3720
Synthetic AML Baseline Centralized Tabular ROC-AUC 0.9416 ± 0.0034 [0.9374, 0.9458] 0.9382 0.9465
Elliptic Bitcoin Graph GraphSAGE GNN (Multi-Seed Mean) ROC-AUC 0.8325 ± 0.0078 [0.8131, 0.8519] 0.8245 0.8401
Elliptic Bitcoin Graph GraphSAGE GNN (Multi-Seed Mean) PR-AUC 0.3761 ± 0.0482 [0.2563, 0.4959] 0.3304 0.4265

Raw Statistical Artifact: benchmarks/results/raw/multi_seed_statistical_summary.json

4.2 Multi-Hop Graph Detection Uplift

On the IBM AMLSim benchmark, inductive GraphSAGE achieves significant detection uplift over Tabular MLP baselines:

  • PR-AUC Uplift: $+0.0434$ ($0.6527$ vs $0.6093$)
  • Recall @ 0.1% Strict FPR: $+4.19\text{ percentage points}$ ($64.12%$ vs $59.93%$)
  • Complex Cycle Recall: $+2.08%$ ($67.36%$ vs $65.28%$)
  • Fan-In Smurfing Recall: $+5.75%$ ($70.50%$ vs $64.75%$)
  • Graph Inference Latency: $1.11\text{ ms}$ per $1{,}000$ transactions

4.3 Scoring Pipeline Compute & Service Latency

Evaluated on host hardware distinguishing in-process compute microbenchmark from live ASGI HTTP service:

  • In-Process Scoring Pipeline Microbenchmark (benchmarks/results/raw/latency_microbenchmark.json):
    • Single-Request Fast-Path Compute: 2.57 ms (well below internal $< 15\text{ ms}$ target; p50: 2.70 ms, p99: 8.87 ms at $C=1$)
    • Peak Throughput: 1,246.3 req/s ($C = 50$; $C=100$: 1,109.9 req/s; 0.0% error rate)
    • p50 Latency (C = 100 Load): 47.49 ms
    • p95 Latency (C = 100 Load): 87.31 ms
    • p99 Latency (C = 100 Load): 105.02 ms
  • Local HTTP Service Benchmark (Class B1 Inference Capacity) (benchmarks/results/raw/latency_http_service_benchmark_post_basehttp0_diagnosis.json):
    • Single-Client Median Latency: 7.12 ms (pooled p50; pooled p99: 10.85 ms at $C=1$)
    • Peak Service Throughput: 543.0 req/s at $C = 50$ on single-worker Uvicorn over loopback TCP (452.8 req/s at $C=10$; 401.7 req/s at $C=500$; 100% 2xx success)

5. Demographic Attribute Availability & Fairness Governance

5.1 Zero-Demographic-PII Invariant

Pursuant to EU GDPR Article 9 special-category prohibitions and ECOA Regulation B (12 CFR Part 1002), models enforce a Zero Demographic PII Invariant. An exhaustive automated scan across all 7 benchmark datasets verified 0 / 10 protected demographic attributes present (100% exclusion rate) across Age, Gender, Race/Ethnicity, Religion, Marital Status, Nationality, Sexual Orientation, Disability, Biometric Data, and Socioeconomic Status.

5.2 Authoritative Regulatory Governance Disclaimers

FEDERAL RESERVE SR 11-7 / OCC 2011-12 FORMAL BIAS GOVERNANCE DISCLAIMER:
All seven standard fraud and anti-money laundering benchmark datasets evaluated by CF-Intelligence deliberately and strictly exclude protected demographic attributes. This exclusion is by deliberate architectural design to satisfy European Union GDPR Article 9 special-category processing prohibitions and Equal Credit Opportunity Act (ECOA) Regulation B restrictions. Direct demographic subgroup fairness testing (e.g. disparate impact by race or sex) is mathematically inapplicable because demographic ground truth is neither collected nor retained in the transaction scoring perimeter.

EQUAL CREDIT OPPORTUNITY ACT (ECOA / 12 CFR PART 1002) STATUTORY NOTICE:
Model parameters are trained purely on structural transaction graphs, payment velocity counters, differential privacy gradients, and cryptographic transaction hash digests. No prohibited bases under 12 CFR Section 1002.2(z) enter gradient updates, ensuring algorithmic non-discrimination and full compliance with CFPB Consumer Financial Protection Circular 2022-03.

EU AI ACT (ARTICLE 10(2)-(3)) DATA GOVERNANCE STATEMENT:
Training datasets undergo continuous data quality and bias mitigation audits. Although special category data is excluded under Article 10(5), proxy attributes (payment channel, merchant tier, geographic corridor) are continuously audited under the EEOC Four-Fifths rule ($0.80 \le \mathrm{DIR} \le 1.25$) to guarantee that models do not produce indirect discriminatory disparities.

5.3 Operational Proxy Attribute Fairness Evaluation

Evaluated under the EEOC 80% Four-Fifths Rule ($0.80 \le \mathrm{DIR} \le 1.25$):

DIR=P(Y^=1∣A=unprivileged)P(Y^=1∣A=privileged),EOD=TPRunprivileged−TPRprivileged\mathrm{DIR} = \frac{\mathbb{P}(\hat{Y}=1 \mid A=\text{unprivileged})}{\mathbb{P}(\hat{Y}=1 \mid A=\text{privileged})}, \quad \mathrm{EOD} = \mathrm{TPR}_{\text{unprivileged}} - \mathrm{TPR}_{\text{privileged}}

Operational Proxy Dimension Privileged Group Unprivileged Group Disparate Impact (DIR) Equal Opportunity (EOD) Demographic Parity (DPD) 80% Rule Compliance
Payment Channel Rail Online / Web Rail Mobile App Rail 1.1338 +0.0228 +0.0032 COMPLIANT [PASS]
Country Corridor Domestic Core Rail Cross-Border Wire Rail 0.9737 -0.0218 -0.0007 COMPLIANT [PASS]
Merchant Risk Tier Standard Retail (5411) Financial Wire (6012) 1.1401 -0.0568 +0.0034 COMPLIANT [PASS]

Raw Fairness Artifact: benchmarks/results/raw/demographic_fairness_audit.json

6. Explainability, Interpretability & Transparency

  • Tabular Attribution: KernelSHAP and TreeSHAP compute local feature attributions, identifying top transaction risk drivers (e.g. transfer velocity, account age, cross-border flag).
  • Integrated Gradients: Path-integral attribution decomposes non-linear neural activations into baseline differences.
  • Topological Subgraph Attribution: GNNExplainer extracts the critical $k$-hop subgraph, identifying the exact laundering path (fan-in, circular routing) responsible for alert escalation.
  • Adverse Action & Reason Codes: Generates Top-5 structured reason codes compliant with CFPB Circular 2022-03.

7. Model Limitations & Concrete Failure Modes

Stratified residual diagnostics (benchmarks/results/raw/error_stratification_analysis.json) document 4 concrete enterprise failure modes:

Failure Mode Affected Operational Stratum Empirical Rate Root Cause Engineering Mitigation
FM-01: Micro-Structuring Amount $< 250\text{ USD}$ (< 250 USD), Degree $k \le 2$ 47.67% FNR Single-bank velocity counters blind to cross-bank micro-bursts below CTR limits DH-PSI anonymous cross-bank velocity linking and reduced anomaly threshold for new accounts
FM-02: Nocturnal Batch Clearing Hours 00:00-05:59, MCC 6012 14.98% FPR Diurnal sine/cosine cyclical encoding penalizes automated off-hours payroll/clearing ISO 20022 camt.053 corporate calendar whitelisting
FM-03: Super-Hub Aggregators Network Degree $k > 50$ 90.00% FNR GNN neighborhood over-smoothing washes out fraud embeddings into clean centroid Temporal edge-weight attention discounting routine high-volume flows
FM-04: Cross-Border Specialty MCC 5999, Cross-Border 43.48% FNR Differential privacy noise ($\epsilon=1.0$) attenuates low-frequency categorical weights Deterministic triage routing borderline scores ($[0.45, 0.55]$) to Four-Eyes human review

8. Environmental & Computational Footprint

  • Hardware Infrastructure: Trained on NVIDIA RTX 4090 / A100 GPUs (training) and validated on x86-64 multi-core CPUs (inference edge).
  • Energy Consumption: Federated local training consumes $\approx 0.042\text{ kWh}$ per local round per banking node.
  • Bandwidth Optimization: Federated model communication overhead is reduced by 74.8% via Zstandard compression, Top-$k$ sparsification ($k=20%$), and INT8 quantization, enabling operation over standard commercial WAN connections.

9. Verification & Audit Test Suites

The model artifacts and governance contracts are certified by automated test suites: