TL;DR

The paper proposes FHE-FedQNN, a federated learning scheme where each client trains a hybrid classical-plus-quantum neural network locally and uploads CKKS-encrypted, quantized weights to a server that homomorphically aggregates them; on CIFAR-10, Brain MRI, and PCOS the FHE overhead costs only ~1–2% accuracy vs. non-encrypted FedQNN, and in one case (PCOS) accuracy improved by 4% [§4, Table 1].

Problem and motivation

Federated learning (FL) keeps raw data local but the messages between clients and server may be intercepted, allowing reconstruction of private data; this is acute in healthcare [§1, p. 2]. The authors target this eavesdropping threat by adding a “quantum-safe” FHE layer on top of FL while simultaneously leveraging quantum layers to accelerate local computation on resource-constrained clients [§1, §3]. Threat model is implicit: server is untrusted to see plaintext weights, and message channels may be intercepted; FHE protects model updates in transit and during aggregation [§1, p. 2; Fig. 1 caption, p. 4].

Key contributions

FHE setup

ML setup

Datasets

Dataset Task Size (train/test) Modality Notes
CIFAR-10 [46] 10-class image classification 48k / 12k Images Batch size 128; 6 qubits used [§4]
Brain MRI [47] 4-class tumor classification 5.7k / 1.3k Medical images Batch size 32; 4 qubits used [§4]
PCOS [48] 2-class classification (polycystic ovary syndrome) 2.56k / 0.64k Medical images Batch size 32; 2 qubits used [§4]

Pipeline diagram

flowchart LR
    subgraph Client_k [Client k]
        D[Local data D_k] --> T[Train hybrid QNN]
        T --> Q[Quantize weights]
        Q --> E[Encrypt CKKS]
    end
    subgraph Server
        A[Weighted sum on ciphertext]
    end
    subgraph Broadcast [Client k]
        DEC[Decrypt global model] --> NEXT[Next-round training]
    end
    E -. enc weights .-> A
    A -. enc global .-> DEC

Pipeline steps (text)

  1. Server generates a CKKS context and Galois keys; initialize global QNN with depth D and gate set G [Algorithm 1, lines 8–10].
  2. Each client k prepares its dataset D_k and trains its local hybrid QNN starting from the global model [Algorithm 1, lines 13–14].
  3. Each client quantizes its trained weights to a fixed precision, then encrypts them under CKKS, and sends ciphertext to the server [Algorithm 1, line 15; §3.1].
  4. Server homomorphically accumulates a weighted sum S = Σ_k w_k^enc · (n_k / n_total), where n_k is client k’s data size [Algorithm 1, lines 19–24].
  5. Server decrypts S to obtain the new global model w_g [Algorithm 1, line 26].
  6. Server optionally adjusts PQC depth/gates (OptimizePQC) before broadcast [Algorithm 1, line 29; §3.2].
  7. Server sends w_g back to every client [Algorithm 1, lines 31–33].
  8. Repeat from step 2 until convergence [Algorithm 1, line 34].

Architecture diagram

Hybrid classical + quantum classifier; exact classical-layer widths are not specified in the text, so the diagram shows the PQC structure given in Fig. 2 of the paper.

flowchart LR
    I([Input features]) --> EMB[Angle embed RX theta_i]
    EMB --> PQC[PQC depth D]
    PQC --> ENT[CNOT entanglers]
    ENT --> M[Measurement]
    M --> CL[Classical head]
    CL --> O([Softmax / cross-entropy])

Results

Each model trained 20 rounds × 20 clients × 10 epochs/round [Table 1 caption]. Hardware is not reported.

Metric This paper (FHE-FedQNN) Baseline (FedQNN) Hardware
CIFAR-10 test acc. 70.12% 72.16% Not reported
Brain MRI test acc. 88.75% 89.71% Not reported
PCOS test acc. 70.15% 66.19% Not reported
CIFAR-10 training time 156.5 min 151.5 min Not reported
Brain MRI training time 116.5 min 110.6 min Not reported
PCOS training time 87.2 min 70.9 min Not reported
CIFAR-10 test loss 1.240 1.202 —
Brain MRI test loss 0.360 0.338 —
PCOS test loss 1.09 0.611 —

Comparable classical-only FHE-FedNN vs. FedNN numbers are also reported [Table 1]. Per-sample encrypted inference latency is not reported.

Limitations and assumptions

Code and artifacts

Open source at https://github.com/elucidator8918/QFL-MLNCP-NeurIPS [§5]. License: Not reported in the text.

Extra diagrams (optional)

Threat model

flowchart LR
    C1[Client 1] -- enc CKKS w_1 --> S[Server aggregator]
    C2[Client 2] -- enc CKKS w_2 --> S
    CN[Client N] -- enc CKKS w_N --> S
    S -- enc global model --> C1
    S -- enc global model --> C2
    S -- enc global model --> CN
    E[Eavesdropper] -. intercepts ciphertext only .-> S

Federated round

sequenceDiagram
    participant C as Clients (k=1..N)
    participant A as Server
    A->>C: global QNN w_g
    C->>C: train local QNN on D_k
    C->>C: quantize + CKKS encrypt
    C->>A: enc(w_k)
    A->>A: S = sum_k (n_k/n_total) * enc(w_k)
    A->>A: decrypt S, optionally update PQC
    A->>C: w_g (next round)

Open questions