TL;DR
The paper proposes two TFHE-based pipelines (via Zama Concrete ML) for collaborative anti-money-laundering detection: (i) an exploratory privacy-preserving GIN-based GNN pipeline using QAT and pruning, and (ii) a fully-working XGBoost pipeline enriched with Snap ML Graph Feature Preprocessor (GFP) features that retains >99% accuracy under encryption on the AMLworld HI-Small dataset [Abstract, §6.4].
Problem and motivation
Money laundering is estimated at 2–5% of global GDP with only 1% seized [§1.1, p. 2]. Financial institutions operate AML in data silos due to GDPR-like regulations and bank-secrecy laws, preventing cross-institution and cross-border transaction graph modeling that criminals exploit [§1.1, p. 2]. The paper proposes a TFHE-based collaborative architecture in which participating FIs encrypt data under the inquiring FI’s public key; a centralised server runs encrypted ML inference and returns encrypted results [§3.5, p. 8]. Threat model: a centralised server performing inference on data contributed (encrypted) by multiple banks; only the inquiring FI holds the secret key, aligning with the Singapore Banking Act prohibition on disclosing Customer Information [§3.5, p. 8].
Key contributions
- Proposed collaborative privacy-preserving AML solution architecture using FHE across multiple financial institutions [§1.2, §3.5].
- First exploration of a TFHE-compatible GNN (GIN) pipeline for AML using Concrete ML, with quantization-aware training (Brevitas) and edge pruning [§1.2, §4].
- Novel TFHE-compatible XGBoost pipeline that integrates Snap ML’s Graph Feature Preprocessor (GFP) to inject fan-in/fan-out, multi-hop and vertex-statistics features [§1.2, §5].
- Empirical study with incrementally added graph features on balanced and imbalanced AMLworld HI-Small subsets, comparing clear vs FHE inference [§6].
- Open-source release of project code [§1.2, footnote 1].
FHE setup
- Scheme(s): TFHE (Fully Homomorphic Encryption over the Torus), via Zama’s TFHE-Concrete with programmable bootstrapping (PBS) [§3.2, §3.3].
- Library / implementation: Concrete ML v1.4.1 on top of Concrete; Brevitas for PyTorch QAT; PyTorch Geometric for the GIN; ONNX as the model interchange [§3.3, §3.4, §6.2].
- Parameters: Not reported (polynomial degree, ciphertext modulus, scale, and explicit security level are not stated).
- Bootstrapping used: Yes; programmable bootstrapping (PBS) is the core mechanism, replacing plaintext bits with a lookup-table function during bootstrapping [§3.2, p. 6].
- Packing / encoding strategy: Not reported (handled internally by Concrete ML / MLIR compiler [§4.4]).
ML setup
- Task: Binary classification of illicit vs licit transactions for AML, encrypted inference (training is plaintext) [§5.4].
- Model architecture:
- GIN (GNN): Graph Isomorphism Network message-passing GNN, adapted from prior AML studies [§4.1]; uses iterative aggregation inspired by Weisfeiler-Lehman; layer counts/widths not stated explicitly in the text. Pipeline: select baseline → quantize → prune → train via Concrete ML → compile to FHE circuit → encrypted inference [§4].
- XGBoost-GFP: Plaintext-trained XGBoost (tuned with 50-iteration Bayesian optimization, 3-fold CV) consuming base transaction features plus GFP-extracted single-hop, multi-hop and vertex-statistics features over an 86,400 s window [§5.2, §6.3].
- Activation handling: GNN replaces
LinearwithQuantLinearand ReLU withQuantReLU(Brevitas) under QAT; weights and accumulator bit-widths are reduced [§4.2]. XGBoost has no activations (tree ensemble). - Operates on: Plaintext model + encrypted data (server has the model, FIs encrypt inputs) [§3.5, §5.4].
- Training vs inference: Training is plaintext; only inference is encrypted under TFHE [§5.4].
Datasets
| Dataset | Task | Size (train/test) | Modality | Notes |
|---|---|---|---|---|
| AMLworld HI-Small (Altman et al. 2023) | AML binary classification | Downsized; temporal split (Not reported splits) | Financial transaction graph | Synthetic, 370 money-laundering pattern groups (fan-in, fan-out, gather-scatter, cycles) [§6.1] |
| Modified AML Dataset 1 | Balanced AML | 15,230 accounts, 10,354 transactions; 50% illicit | Transaction graph | Undersampled from HI-Small [§6.1] |
| Modified AML Dataset 2 | Imbalanced AML | 9,070 accounts, 5,491 transactions; 5.72% illicit | Transaction graph | Random-sampled from HI-Small [§6.1] |
Pipeline diagram
flowchart LR
subgraph FIs [Participating FIs]
Q[Inquiry FI query] --> EK[Encrypt with inquiry pk]
OF[Other FI data] --> EK
end
subgraph Server [Centralised Server]
FW[Forward query] --> AGG[Aggregate ciphertexts]
AGG --> INF[Encrypted XGBoost/GIN inference]
end
subgraph Dec [Inquiry FI]
D[Decrypt with sk] --> R[AML risk label]
end
Q -.-> FW
EK -. ciphertext .-> AGG
INF -. ciphertext .-> D
Pipeline steps (text)
- Inquiry FI initiates an AML query and sends it to the centralised server [§3.5].
- Server forwards the query to participating FIs [§3.5].
- Each FI encrypts its relevant transaction data with the inquiry FI’s TFHE public key (shared once) [§3.5].
- Server consolidates ciphertexts and runs encrypted inference using the trained XGBoost-GFP (or, prospectively, GIN) model [§3.5].
- Encrypted result is returned to the inquiry FI, which decrypts with its secret key to obtain the AML label [§3.5].
Architecture diagram
GIN GNN (TFHE target)
flowchart LR
I([Transaction graph: node + edge feats]) --> Q1[QuantLinear input]
Q1 --> M1[GIN message passing]
M1 --> A1[QuantReLU]
A1 --> M2[GIN aggregation]
M2 --> A2[QuantReLU]
A2 --> EP[Edge pruning]
EP --> Q2[QuantLinear classifier]
Q2 --> O([Illicit / licit])
Notes: exact layer count and hidden widths are not stated in the text; QAT layers and bit-widths are configured per-layer in Brevitas [§4.2]. Compilation to FHE failed in places because Concrete ML lacked the ONNX ScatterElements operator; a custom implementation was attempted but NumpyModule → QuantizedModule conversion remained incomplete [§4.4].
XGBoost-GFP (deployed TFHE pipeline)
flowchart LR
I([Transactions]) --> S[Temporal split]
S --> GFP[Snap ML GFP: fan-in/out, multi-hop, vertex-stats]
GFP --> H[Bayesian HP tuning, 50 iters, 3-fold CV]
H --> T[XGBoost train plaintext]
T --> QZ[Quantize to integers]
QZ --> SIM[FHE simulation check]
SIM --> CC[Compile to FHE circuit]
CC --> INF([Encrypted inference])
Results
Hardware for all numbers below: Intel Xeon CPU E5-1630 v4 @ 3.70 GHz, 8 cores, Linux; Concrete ML 1.4.1 (CPU only) [§6.2].
| Metric | This paper | Baseline | Hardware |
|---|---|---|---|
| Accuracy (balanced, basic feats), clear / FHE | 0.9972 / 0.9972 | — | Xeon E5-1630 v4 [§6.4 Table 3] |
| F1 (balanced, basic), clear / FHE | 0.9978 / 0.9978 | — | Xeon E5-1630 v4 |
| F1 (imbalanced, basic) | 0.3056 (clear = FHE) | — | Xeon E5-1630 v4 [§6.5 Table 5] |
| F1 (imbalanced, +single-hop GFP) | 0.3867 (clear = FHE) | basic 0.3056 (+8 pts) | Xeon E5-1630 v4 |
| Avg batch inference (balanced, basic) | 1009.0963 s FHE vs 0.008414 s clear | — | Xeon E5-1630 v4 [§6.4 Table 4] |
| Total inference (balanced, basic) | 20,181.93 s FHE vs 0.1683 s clear (≈119,926×) | — | Xeon E5-1630 v4 |
| Avg batch inference (imbalanced, +multi-hop) | 34.13 s FHE vs 0.00227 s clear | — | Xeon E5-1630 v4 [§6.5 Table 6] |
| FHE / clear time ratio | 15,038× – 165,917× across feature sets | — | Xeon E5-1630 v4 |
Headline accuracy delta vs plaintext: 0 (Concrete-TFHE quantized XGBoost matches plaintext to 4 decimals across all four metrics on both datasets) [§6.4 Table 3, §6.5 Table 5].
Single-sample latency on a per-input basis is not separately reported — only batch averages and totals; not enough information is given to safely derive a per-sample second value, so single_inference_seconds is N/A in the comparison block.
Limitations and assumptions
- GNN pipeline did not fully compile to FHE: Concrete ML lacked the ONNX
ScatterElementsoperator; a custom implementation was added butNumpyModule → QuantizedModuleconversion remained unresolved [§4.4]. Only the XGBoost pipeline was end-to-end FHE-functional. - Massive FHE overhead: total encrypted inference is 15,000× – 165,000× slower than plaintext [§6.4 Table 4, §6.5 Table 6]; authors flag this trade-off as a key open problem [§6.4, §7].
- Datasets are downsized synthetic data: AMLworld HI-Small was undersampled (15k / 9k accounts) due to FHE compute cost [§6.1]; results may not extrapolate to production-scale graphs.
- No GPU support in Concrete ML at time of writing [§6.2].
- FHE security parameters (poly degree, modulus, security level) are not reported.
- No multi-party key-generation protocol described: the architecture assumes the inquiry FI’s pk is “securely shared once” [§3.5], but how trust is established between participating FIs is not detailed.
- Threat model is implicit honest-but-curious server; collusion and malicious-FI scenarios are not analysed.
Related work it compares against
CryptoNets [Dowlin et al.], HCNN [Badawi et al.], LoLa [Brutzkus et al.], ResNet-20 RNS-CKKS with bootstrapping [Lee et al.], CryptoGCN [Ran et al.]; AML modelling baselines from Altman et al., Egressy et al., Blanuša et al. (GFP) [§2.2, §2.3].
Code and artifacts
Released: https://github.com/fabecode/GraphML-FHE [§1.2, footnote 1]. License not stated in the paper text.
Extra diagrams (optional)
Threat model
flowchart LR
FI1[Inquiry FI: pk, sk] -- enc(query) --> S[Centralised server: model]
FI2[Participating FI 2] -- enc(data, pk_FI1) --> S
FI3[Participating FI 3] -- enc(data, pk_FI1) --> S
S -- enc(AML label) --> FI1
Open questions
- What is the actual GIN depth / hidden dimensionality used in the experiments? (Not stated in §4.1.)
- What TFHE/Concrete security parameters (LWE dimension, noise, polynomial degree) were used?
- Why does inference time decrease as graph-feature complexity grows on the imbalanced dataset [§6.5 Table 6]? The authors flag this as needing further investigation.
- How is the inquiry FI’s public key distributed and revoked across institutions in practice?
- Would a deeper GIN (once
ScatterElementsis supported) match or exceed XGBoost-GFP under FHE?