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

FHE setup

ML setup

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)

  1. Inquiry FI initiates an AML query and sends it to the centralised server [§3.5].
  2. Server forwards the query to participating FIs [§3.5].
  3. Each FI encrypts its relevant transaction data with the inquiry FI’s TFHE public key (shared once) [§3.5].
  4. Server consolidates ciphertexts and runs encrypted inference using the trained XGBoost-GFP (or, prospectively, GIN) model [§3.5].
  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

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