TL;DR

The paper benchmarks three classical ML models (SGD classifier, XGBoost, Decision Tree) under Fully Homomorphic Encryption using the Concrete-ML library on a GAN-generated medical PII text dataset, showing FHE matches plaintext accuracy but with sharply increased training and inference latency [§Abstract][§V].

Problem and motivation

Medical data (EHRs, wearables, genomic data) is sensitive under HIPAA/GDPR, yet ML on such data is valuable for diagnosis and personalization [§I]. Privacy-preserving ML via FHE allows computations on encrypted data, but FHE incurs significant computational and implementation costs [§I]. The authors target classifying medical dialogues that may contain personally identifiable information (PII vs. Non-PII) under encryption [§Abstract][§IV]. Threat model is not formally stated; framing is healthcare-style outsourced inference where sensitive PII must not be exposed during processing [§I][§II].

Key contributions

FHE setup

ML setup

Datasets

Dataset Task Size (train/test) Modality Notes
GAN-generated medical PII dataset Binary text classification (PII vs. Non-PII) 100,000 records; 80/20 train/test split Tabular + text dialogue 8 features: dialogue ID, text, subject (PII fields), user_type, region, issue_priority, agent_experience, Label; label-encoded categoricals; tokenized text; balanced classes [§IV-B]

Pipeline diagram

flowchart LR
    subgraph Client
        A[Collect PII dialogue] --> B[Preprocess: label-encode, tokenize]
        B --> C[Encrypt features]
    end
    subgraph Server
        D[Plaintext-trained model] --> E[FHE inference: DT / XGB / SGD]
    end
    subgraph Client_dec [Client]
        F[Decrypt label] --> G[PII or Non-PII]
    end
    C -. ciphertext .-> E
    E -. ciphertext .-> F

Pipeline steps (text)

  1. Collect medical dialogues containing potential PII [§IV-A].
  2. Preprocess: label-encode categorical fields and tokenize text [§IV-B].
  3. Split into 80% train / 20% test [§IV-B].
  4. Train Decision Tree, XGBoost, and SGD models on plaintext to establish parameters/decision boundaries [§IV-C].
  5. Encrypt test inputs (and model parameters/thresholds where applicable) using Concrete-ML [§IV-B][§IV-C].
  6. Run encrypted inference: homomorphic dot products for SGD, encrypted tree evaluation with look-up tables for DT, encrypted aggregation of weak learners for XGBoost [§IV-C].
  7. Return encrypted prediction; client decrypts to obtain PII / Non-PII label [§IV-D][§V].

Architecture diagram

flowchart LR
    I([Input: 8 PII features]) --> P[Label-encode + tokenize]
    P --> M{Model}
    M --> DT[Decision Tree: encrypted thresholds + LUT]
    M --> XGB[XGBoost: encrypted weak-learner aggregation]
    M --> SGD[SGD: enc dot-product w x + b]
    DT --> O([Encrypted label])
    XGB --> O
    SGD --> O

Results

All three models match their plaintext accuracy under FHE; SGD and XGBoost reach 100% on the synthetic dataset, DT reaches 91.7% [§V, Tables I–II]. Latency and memory grow substantially, especially for Decision Tree inference [§V].

Metric This paper (FHE) Baseline (plaintext) Hardware
SGD accuracy 100.0% 100.0% Not reported
XGB accuracy 100.0% 100.0% Not reported
DT accuracy 91.7% 91.7% Not reported
SGD training time 44.40 s 0.300 s Not reported
XGB training time 5.05 s 24.540 s Not reported
DT training time 91.20 s 0.310 s Not reported
SGD prediction time 270.39 s 0.05 s Not reported
XGB prediction time 0.410 s 0.03 s Not reported
DT prediction time 20459.98 s 0.004 s Not reported
SGD precision / recall / F1 / ROC-AUC 100% / 100% / 100% / 100% 100% / 100% / 100% / 100% Not reported
XGB precision / recall / F1 / ROC-AUC 100% / 100% / 100% / 100% 100% / 100% / 100% / 100% Not reported
DT precision / recall / F1 / ROC-AUC 85.7% / 100% / 92.3% / 91.8% 85.7% / 100% / 92.3% / 91.8% Not reported

Note: Prediction times in Table II appear to be reported over the test set (not per-sample); the single_inference_seconds field uses XGB+FHE’s 0.410 s figure as the fastest reported prediction time. Per-sample latency is not explicitly stated [§V, Table II].

Limitations and assumptions

No direct system-level comparison. Positions against the FHE/PPML literature in general — including surveys and applications of FHE in healthcare, bioinformatics, and cloud computing [§II][§III], referencing Gentry’s FHE [8], CKKS [10], SHE [9], and Wood et al.’s HE-for-medicine survey [17]. The plaintext models serve as the in-paper baseline [§V].

Code and artifacts

Not released. Built on the Concrete-ML library [§IV-B].

Extra diagrams (optional)

Threat model

Not formally defined in the paper; implied client-server outsourced inference where the server runs encrypted classification on PII dialogues without seeing plaintext.

Open questions