TL;DR

The paper integrates Fully Homomorphic Encryption (CKKS via TenSEAL) into a Flower-based federated learning pipeline that trains a MobileNetV2 binary classifier on private Belgian mammograms; due to FHE memory constraints, only the last layer’s weights are encrypted before transmission to the aggregator [Abstract][§VI-C]. It reports comparable ROC behaviour to the plaintext FL baseline while exposing serialization/memory as the dominant practical bottleneck [§VI-D][§VII].

Problem and motivation

Medical AI development is constrained by GDPR, data scarcity, class imbalance and risk of leaking patient information [§I]. Federated learning helps with data locality but is still vulnerable to membership/attribute/model-inversion/gradient-leakage/property/data-reconstruction inference attacks and to model-poisoning attacks during weight exchange [§I-B]. The threat model assumes participants are benevolent (they share a common FHE private key used to encrypt aggregator-bound weights) and that the central server should not see plaintext weights — FHE is added on top of FL to block server-side and in-transit inference attacks [§III]. They explicitly note that if a participant is malicious or the shared private key is compromised, encrypted results are no longer secure [§III].

Key contributions

FHE setup

ML setup

Datasets

Dataset Task Size (train/test) Modality Notes
Private Belgian hospital mammograms (with UMONS / Helora) Binary classification (malignant vs. benign, biopsy-confirmed) Not reported Mammographic images (positive/negative classes, masses and microcalcifications) Annotated by Dr Salvatore Murgo (Helora); prospective study; anonymized [§I-D][§V-B][Acknowledgement]

Pipeline diagram

flowchart LR
    subgraph Setup
        K[Shared CKKS keys via USB] --> P1[Participant 1]
        K --> P2[Participant N]
    end
    subgraph Round
        S[Server sends global model] -. serialized weights .-> P1
        S -. serialized weights .-> P2
        P1 --> T1[Local train + val]
        P2 --> T2[Local train + val]
        T1 --> E1[Encrypt last layer]
        T2 --> E2[Encrypt last layer]
        E1 -. ciphertext weights .-> A[FedAvg on ciphertext]
        E2 -. ciphertext weights .-> A
        A -. ciphertext global .-> P1
        A -. ciphertext global .-> P2
        P1 --> D1[Decrypt locally]
        P2 --> D2[Decrypt locally]
    end

Pipeline steps (text)

  1. Generate one shared CKKS key set and distribute it physically (USB) or over an out-of-band channel to all hospital participants [§III].
  2. Server initializes the global MobileNetV2 and serializes (gRPC) a copy to each selected participant [§III].
  3. Each participant deserializes weights and, if they arrive as ciphertext (rounds ≥ 2), decrypts with the locally held private key before loading into the local model [§III].
  4. Client trains locally for the configured number of epochs (50 in the main experiments) with MPS-accelerated PyTorch on a Mac M1 [§V-C].
  5. Client encrypts the last FC layer’s weights with TenSEAL/CKKS using the shared public key, then serializes and uploads to the server [§III][§VI-C].
  6. Server deserializes ciphertexts and performs FedAvg homomorphically (no decryption); server-side evaluation is intentionally omitted because it would require the private key [§III].
  7. Server returns the aggregated ciphertext global weights to each participant for the next round [§III].
  8. Each participant performs the round’s evaluation locally after decryption (the global model is only ever evaluated client-side) [§III].

Architecture diagram

flowchart LR
    I([Input image]) --> C0[Conv 3x3, 32, s2 + BN + ReLU6]
    C0 --> IR[17x InvertedResidual blocks ReLU6]
    IR --> C18[Conv 1x1, 1280 + BN + ReLU6]
    C18 --> GAP[Global pool / flatten]
    GAP --> DO[Dropout 0.2]
    DO --> FC1[Linear 1280 to 320]
    FC1 --> FC2[Linear 320 to 2 ENCRYPTED]
    FC2 --> SM([Softmax])

Results

Headline accuracy (3 rounds, 3 clients, classic FL vs. FL+HE), Table II [§VI-D]:

Metric This paper (with HE) Baseline (FL, no HE) Hardware
Accuracy, batch 16 73.2% 70.6% Mac M1, 32 GB RAM (160 GB w/ swap), CPU FHE [§V-C]
Accuracy, batch 32 87.4% 73.8% same
Accuracy, batch 64 82.5% 83.4% same
Micro-avg ROC, batch 64 −7% vs no-HE reference same [§VI-D, Fig. 11]
Micro-avg ROC, batch 32 +8% vs no-HE reference same [§VI-D, Fig. 12]
Encryption time, ~12000-value tensor, heavy key (2^15, [60,40×8,60], scale 40) 301 s — Mac M1 CPU [Table I]
Encryption time, single-value tensor at same key < 0.1 s — Mac M1 CPU [Table I, §VI]
Max encryptable values (lightest key in Table I) 683000 — — [§VI-A]
Max serializable values per tensor (TenSEAL limit at chosen key) 6424 — — [§VI-C]
Memory for ~6000-value encrypted tensor ~4 GB — — [§VI-C, Fig. 10]
Total round-trip overhead vs plaintext FL +28 min (20 clients, 15 rounds, 10 epochs) reference Mac M1 [§VI-B, Fig. 9]
Encrypting 4 layers vs 1 (heavy key, batch 16) 67% micro-ROC @ 11 epochs / 81% @ 50 epochs 71% / 79% (1 layer encrypted) Mac M1 [§VI-D end]

No single-image encrypted inference latency is reported (inference is plaintext on the client), so single_inference_seconds: N/A in the comparison block [§III][§VI].

Limitations and assumptions

Federated learning for breast cancer: Nguyen Tan et al. [43], Jiménez-Sánchez et al. [30], Li et al. [33] (FL + HE on BreakHis). Privacy in medical FL: Kaissis et al. [31] (multi-layer secure FL architecture), Adnan et al. [4] (FL + differential privacy on histopathology), Hosseini et al. [23] (cluster-based SMPC FL). FHE schemes referenced for context: Gentry [21], DGHV, BGV, BFV [19], GSW, CKKS [15], TFHE [16]. No quantitative head-to-head benchmark against these systems is provided [§II].

Code and artifacts

Not released.

Extra diagrams (optional)

Threat model

flowchart LR
    H1[Hospital 1] -- enc last-layer weights --> AGG[Aggregator FedAvg on ciphertext]
    H2[Hospital N] -- enc last-layer weights --> AGG
    AGG -- enc global weights --> H1
    AGG -- enc global weights --> H2
    KEY[Shared CKKS key via USB] --> H1
    KEY --> H2
    ATT[Honest-but-curious server / in-transit eavesdropper]:::adv -. blocked .-> AGG
    classDef adv fill:#fdd,stroke:#900;

Federated round

sequenceDiagram
    participant C as Clients (hospitals)
    participant A as Aggregator (server)
    A->>C: serialized global model
    C->>C: deserialize + (round>=2) decrypt with shared sk
    C->>C: local train (50 epochs, MPS on Mac M1)
    C->>C: encrypt last FC layer with shared pk (CKKS)
    C->>A: serialized ciphertext weights (gRPC)
    A->>A: FedAvg homomorphically on ciphertext
    A->>C: ciphertext global weights
    C->>C: client-side evaluation (no server-side eval)

Open questions