💻 computer science

QuickFund-X: Privacy-Preserving Cross-Platform Fraud Detection Using Heterogeneous Graph Neural Networks

QuickFund-X is a privacy-preserving system that detects cross-platform loan syndicate fraud by linking isolated micro-lending platforms via secure tokenization and employing a hybrid GraphSAGE and rule-based model to achieve perfect recall without false alarms.

Original authors: Venugopal M, Rohith Vaidya K, Vijay J, Yashwanth S, Vikas P R

Published 2026-10-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Venugopal M, Rohith Vaidya K, Vijay J, Yashwanth S, Vikas P R

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the fast-moving world of digital finance, a new kind of danger has emerged alongside the convenience of instant loans. Imagine a criminal group that does not steal money from a vault but instead tricks multiple lenders into giving them small loans all at once. They apply for funds on several different apps within minutes, using the same phone number or address for each request. Because each lender operates in isolation, unaware of the others, they see only a single, clean application and approve the money. By the time the truth comes out, the group has vanished with the cash. This is known as a credit stacking attack, and it thrives on a strict barrier: privacy laws that prevent lenders from sharing customer data. To solve this, researchers have developed a system that allows lenders to spot these coordinated rings without ever seeing each other's private customer lists. The solution relies on a clever mix of cryptography, which scrambles data so it cannot be read, and graph neural networks, a type of artificial intelligence that learns by studying how things are connected rather than just looking at them in isolation.

A team of researchers from BMS College of Engineering has built a working prototype called QuickFund-X to tackle this specific problem. Their goal was to create a system where competing lending platforms could work together to stop fraud without violating privacy regulations like India's Digital Personal Data Protection Act or the European Union's GDPR. In a standard setup, if a borrower applies for a loan, their personal details like their name or phone number are kept strictly within that company's database. QuickFund-X changes this dynamic by creating a secure bridge. Before any application leaves a lender's computer, the system takes sensitive details and turns them into a unique, unbreakable code. This process uses a specific cryptographic method that ensures the code is the same every time the same phone number or email is used, but it is impossible to reverse-engineer the code back into the original number. This means the central system can recognize that two different applications came from the same person without ever knowing who that person is.

The heart of the system is a central coordinator that acts like a living map of all the connections between these coded applications. Instead of storing a static list of who applied when, the system builds a dynamic web, or graph, that grows in real time as new loans are requested. This map includes different types of nodes, such as applications, platforms, and the coded tokens for phones or emails, linked together by the relationships between them. When a new loan application arrives, the system does not just check it against a list of rules; it uses a specialized artificial intelligence model to look at the immediate neighborhood of that application on the map. The model is designed to look four steps deep into the connections, allowing it to see if a new application is suddenly linked to a cluster of other recent applications that share the same coded tokens. This depth is crucial because it matches the typical path a criminal ring takes when applying for multiple loans across different apps.

To make the system both smart and reliable, the researchers combined the artificial intelligence's findings with a set of clear, logical rules. The final decision on whether an application is fraudulent is a blend of the AI's probability score and a rule-based check that looks for obvious signs of trouble, such as too many applications from the same token in a short time. In their tests, the researchers simulated a scenario with four different lending platforms, each running its own independent database. They introduced a coordinated fraud ring where a group submitted applications across all four platforms in rapid succession. The system successfully identified every single one of these fraudulent attempts, achieving a perfect detection rate with zero false alarms. In contrast, systems that relied only on the rules or only on the artificial intelligence missed a significant number of attacks or flagged innocent borrowers as suspicious.

The results showed that the system could handle the speed required for real-world lending. Even when simulating hundreds of concurrent requests, the system processed applications quickly, with most decisions made in under a second. Crucially, the system respected the privacy of the lenders. When a fraud ring was detected, the alert was sent only to the specific platforms involved in that ring. Other platforms that had no connection to the fraud received no notification at all, ensuring that no sensitive information about customer activity was leaked to competitors. The researchers demonstrated that this approach allows lenders to protect themselves from organized crime without needing to pool their private data or break the law. While the current version was tested in a controlled environment, the findings suggest that this architecture offers a viable path forward for the financial industry, balancing the need for security with the strict demands of modern privacy laws.

Technical Summary: QuickFund-X

Problem Statement

Digital micro-lending platforms in emerging markets face a critical vulnerability: credit stacking attacks. Organized syndicates submit simultaneous loan applications across multiple competing lending apps within minutes. Because traditional credit bureaus rely on asynchronous batch reporting (lagging 7–30 days), individual lenders see an unblemished profile and approve loans in isolation.

This creates a deadlock between two imperatives:

Risk Mitigation: Lenders need to share data to detect cross-platform fraud rings.
Privacy Compliance: Regulations such as India's DPDP Act 2023, the EU GDPR, and RBI mandates strictly prohibit sharing unencrypted Personally Identifiable Information (PII) like phone numbers, PANs, or addresses between commercial competitors.

Existing academic solutions often fail in production because they rely on centralized static datasets, require expensive full-graph retraining, or lack the cryptographic rigor to satisfy regulatory privacy standards.

Methodology

The authors propose QuickFund-X, a privacy-preserving, cross-platform fraud detection ecosystem built on a three-tier architecture:

1. Multi-Platform Architecture & Secure Outbox

The system simulates four autonomous lending portals (QuickFund, TrustLend, SecureCredit, FinBridge), each with isolated React frontends, Express.js backends, and MongoDB databases. To ensure zero data loss without distributed locks, each backend implements the Transactional Outbox Pattern. Application records and outbox events are committed atomically; background workers stream events to a central coordinator with exponential backoff retry logic to handle network instability.

2. Privacy-Preserving Edge Tokenization

To comply with DPDP and GDPR, no raw PII leaves the local network. Before transmission, a local edge worker executes a deterministic pipeline:

Canonicalization: Normalizes data (e.g., E.164 phone formats, lowercased emails).
Key Derivation: Uses HKDF (RFC 5869) to derive platform-specific keys from a master secret.
Tokenization: Generates irreversible 256-bit tokens using HMAC-SHA256.
Encryption: Local records remain encrypted at rest via AES-256-GCM.
Identical credentials across different platforms produce identical token hashes, enabling collision detection at the central node without exposing plaintext identities.

3. Persistent Heterogeneous Graph Engine

A central coordinator (PostgreSQL) maintains a dynamic heterogeneous bipartite graph ( $G = (V, E)$ ) rather than static adjacency matrices.

Nodes (8 types): Application, Platform, PhoneToken, EmailToken, NameToken, AddressToken, EmployerToken, DeviceToken.
Edges (7 types): Semantic relationships like HAS_PHONE, SUBMITTED_TO, etc.
Ingestion: Updates occur via atomic upserts, allowing the graph to scale incrementally without full re-indexing.

4. Hybrid Scoring & Decisioning

The core detection logic combines deep learning with deterministic rules:

Heterogeneous GraphSAGE: A custom 4-layer PyTorch model executes bidirectional message passing over a 4-hop receptive field. This depth is specifically chosen to capture the topological path of a credit stacking attack (e.g., App A $\to$ Phone $\to$ App B $\to$ Platform B).
Rule Engine: A heuristic module evaluates token collisions, velocity (>3 apps/24h), and cluster degree.
Hybrid Score: The final risk score ( $S_{Hybrid}$ ) is a convex combination:
$S_{Hybrid} = 0.60 \cdot P_{GNN} + 0.40 \cdot S_{Rule}$
Applications exceeding a threshold of 0.50 are flagged.

5. Explainable & Targeted Dispatch

The system generates human-readable forensic explanations for adverse actions. Crucially, alerts are dispatched via webhooks strictly to the platforms involved in the specific fraud ring subgraph, preventing "zero-spillover" leakage of competitive intelligence to uninvolved lenders.

Key Contributions

Fully Decoupled Deployment: A functional multi-tenant ecosystem with isolated databases and frontends.
Zero-PII Cryptography: Implementation of edge tokenization using HKDF and HMAC-SHA256, ensuring no plaintext PII traverses the network.
Streaming Graph Engine: A persistent PostgreSQL-based graph model supporting incremental upserts, avoiding the computational cost of offline batch retraining.
Custom 4-Layer Heterogeneous GraphSAGE: A lightweight neural network designed to match the specific topological depth of multi-lender credit stacking.
Interpretable Hybrid Decisioning: A 0.6 GNN + 0.4 Rule fusion model that balances structural probability with deterministic heuristics to satisfy regulatory audit requirements.

Results

The system was evaluated under realistic multi-lender conditions on an Intel Core i5 with 16 GB RAM running Dockerized containers for the four lending platforms, the central fraud engine, and automated traffic injectors.

Fraud Detection Performance:
- Recall: 100% on coordinated fraud rings.
- False Positive Rate (FPR): 0.0% on benign multi-application cases.
- Precision: 100%.
- F1-Score: 1.000.
- Comparison: Siloed rules achieved 0% recall; centralized rules achieved 50% recall with 25% FPR; GNN-only achieved 87.5% recall with 8.3% FPR. The hybrid approach outperformed all baselines.
Throughput & Latency:
- At 50 concurrent requests, the system sustained a throughput of 45.5 applications/second with a median latency of 896 ms and 0% errors.
- Throughput decreased to 13.9 applications/second under 500 concurrent requests, with median latency rising to ~14.8 seconds.
- Cryptographic tokenization added negligible overhead (~0.14 ms per application).
Scalability: The graph engine successfully expanded from 36 to 52 applications and 432 edges during streaming trials without full re-indexing.

Significance and Claims

The paper claims to resolve the operational deadlock between strict data privacy laws and collaborative risk mitigation. By leveraging edge cryptographic tokenization and streaming heterogeneous graph learning, QuickFund-X enables lenders to detect cross-platform syndicates without violating DPDP or GDPR.

The authors emphasize that the system demonstrates operational viability in a simulated environment, highlighting:

Regulatory Compliance: Zero exposure of raw PII.
Operational Resilience: Transactional outbox buffering ensures zero dropped events during network partitions.
Auditability: The hybrid scoring model provides the "auditable forensic justifications" required by financial adverse-action laws, moving beyond black-box confidence scores.

The work positions itself as a bridge between academic graph prototypes (often tested on static, centralized datasets) and the reality of multi-tenant, privacy-constrained financial ecosystems, validating the architecture's potential for production deployment.