TL;DR

A brief survey that taxonomises efficiency optimization techniques for HE-based privacy-preserving federated learning (PPFL), grouping them into algorithmic, hardware, and hybrid categories with discussion of limitations on edge/IoT deployments [§I, §III–§V].

Problem and motivation

Federated learning (FL) leaks information through uploaded local models, enabling membership-inference, gradient-inversion, and attribute-extraction attacks [§I]. Among privacy defenses (MPC, DP, HE), HE offers post-quantum-safe end-to-end ciphertext computation without trusted setup or noise-induced accuracy loss, but suffers severe efficiency overhead — e.g., Paillier increases training time 96x–135x and data transfer for FMNIST/CIFAR vs plaintext [§II-D]. Prior PPFL surveys cover construction and attacks but not efficiency optimization specifically for HE [§I, Table I]. Threat model: semi-honest and malicious adversaries, individually or colluding, including curious server and malicious clients [§II-F].

Key contributions

FHE setup

ML setup

Datasets

Dataset Task Size (train/test) Modality Notes
FMNIST Classification (cited Batchcrypt benchmark) Not reported Images Used to quantify Paillier 96x training-time overhead [§II-D, p. 24572]
CIFAR Classification (cited Batchcrypt benchmark) Not reported Images Paillier 135x training-time overhead vs plaintext [§II-D, p. 24572]
Multiple (other) Various Not reported Tabular / vision / medical Survey aggregates results across many primary works without re-reporting datasets

Pipeline diagram

flowchart LR
    subgraph Clients
        D[Local data] --> T[Local training]
        T --> Q[Quantize / pack / sparsify]
        Q --> E[Encrypt with pk]
    end
    subgraph Server
        A[Homomorphic aggregation]
    end
    subgraph Clients_dec [Clients]
        DEC[Decrypt with sk] --> U[Update local model]
    end
    E -. enc gradients .-> A
    A -. enc global update .-> DEC
    U --> T

Pipeline steps (text)

  1. Each client trains a local model on private data [§II-D1].
  2. Client applies algorithmic optimizations (quantization, packing, TPS sparsification, GDO with extra local rounds) before encryption [§III-A, §III-B].
  3. Client encrypts updates with the shared public key using PHE (Paillier) or FHE (BGV/BFV/CKKS/TFHE), optionally accelerated by FPGA/GPU/CPU (HEXL) [§II-D, §IV].
  4. Server homomorphically aggregates encrypted updates without seeing plaintext [§II-D1].
  5. Server returns encrypted (or sometimes decrypted, in HE+DP designs) aggregate to clients [§II-D, §V-A].
  6. Clients decrypt with private key, update local models, and proceed to next round [§II-D1].
  7. Optional hybrid path: secret-sharing protocols handle ops that pure HE cannot do efficiently (Pivot, CAESAR, FEDXGB) [§V-B].

Architecture diagram

Because the paper is a survey of optimization techniques (no single model), the diagram below depicts the taxonomy of optimizations rather than an NN architecture [§III–§V, Fig. 1].

flowchart LR
    R([HE-based PPFL optimizations]) --> ALG[Algorithmic]
    R --> HW[Hardware]
    R --> HYB[Hybrid]

    ALG --> Q[Quantization]
    ALG --> P[Packing]
    ALG --> TPS[Transmit param selection]
    ALG --> GDO[Gradient-descent opt]
    ALG --> LHE[Lightweight HE]
    ALG --> E2E[End-to-end opt]
    ALG --> CT[Concurrent training]

    HW --> FPGA[FPGA]
    HW --> GPU[GPU]
    HW --> CPU[CPU / HEXL]

    HYB --> HEDP[HE + DP]
    HYB --> HESS[HE + SS]

Results

The survey reports comparative claims from primary works rather than running new experiments [§III–§V].

Metric This paper Baseline Hardware
Paillier training-time overhead (FMNIST) 96x vs plaintext FL [63] plaintext FedAvg Not specified [§II-D, p. 24572]
Paillier training-time overhead (CIFAR) 135x vs plaintext FL [63] plaintext FedAvg Not specified [§II-D]
GDO speedup (Wei et al.) >10x vs FedAvg [88] FedAvg Not specified [§III-B1]
HE-based backprop (Zhang & Zhu) >100x speedup vs CryptoNets [89] CryptoNets [90] Not specified [§III-B1]
GPU FHE logistic regression (Jung et al.) 40x vs 8-threaded CPU [113] 8-threaded CPU [115], [116] GPU vs CPU [§IV-B, p. 24577]
GPU bootstrapping (Jung et al.) >100x faster bootstrapping [113] Prior FHE bootstrapping GPU (memory-centric) [§IV-B]

Limitations and assumptions

Surveys positioned against: Lyu et al. [40], Yin et al. [41], Liu et al. [42], Yang et al. [43], Mansouri et al. [44], Liu et al. [45], Jiang et al. [46], Juvekar et al. (GAZELLE) [47], Boemer et al. (nGraph-HE2) [48], Phong et al. [28] — all noted as not covering HE-efficiency optimization in PPFL specifically [Table I, §I].

Primary systems catalogued include: Batchcrypt [63], SecureBoost+ [66], eHE-SecureBoost [67], EaSTFLy [69], FHE-DiNN [71], TAPAS [73], FLZip [80], AVFL [81], CE-VFL [82], FLARE [35], FedML-HE [86], FeDBCD [87], VPFL [92], ACCEL [93], Hao et al. [94], SPINDLE [97], POSEIDON [98], CryptoBoost [96], HAFLO [99], CARM [100], FLASH [102], FPGA accelerator [101], HEXL [118], Pivot [106], PrivFL [34], CAESAR [107], FEDXGB [119], CryptoNets [90] [§III–§V].

Code and artifacts

Not released — survey paper. No companion repo cited.

Extra diagrams (optional)

Threat model

flowchart LR
    C1[Client i] -- enc(grad_i) --> S[Aggregation server]
    C2[Client j] -- enc(grad_j) --> S
    S -- enc(global) --> C1
    S -- enc(global) --> C2
    A[Semi-honest or malicious adversary] -. eavesdrop / collude .-> S
    A -. collude .-> C2

Threats include semi-honest server, malicious clients, and collusion (server with subset of clients) [§II-F, p. 24573].

Federated round

sequenceDiagram
    participant C as Clients
    participant S as Aggregator
    C->>C: Local training + quantize/pack/sparsify
    C->>S: enc(local update) via PHE or FHE
    S->>S: Homomorphic sum (FPGA/GPU/CPU accel)
    S->>C: enc(global model)
    C->>C: Decrypt with sk, update local model

Open questions