TL;DR

FLHC-QPR is a federated learning framework for edge AI that combines CKKS homomorphic encryption with client clustering, adaptive quantization, and structured pruning. On the UCI Human Activity Recognition dataset, it improves accuracy by +8.4%, reduces latency by 16.2%, and shrinks memory overhead by 35.1% relative to baseline FLH [Abstract; §VI].

Problem and motivation

Federated learning on edge devices must protect sensitive data (healthcare, IoT, finance) while coping with bandwidth, memory, and convergence constraints. Adding HE to FL inflates ciphertext size and computation, making naive deployments impractical on resource-limited devices [§I]. The paper targets these communication/computation/security trade-offs without weakening privacy. The threat model is implicit: an honest-but-curious server that should never see plaintext client updates (“the server does not have access to the raw updates”) and clients that may have heterogeneous resources [§IV-C]. Formal adversary modeling and attack analyses are explicitly out of scope [§V-F].

Key contributions

FHE setup

ML setup

Datasets

Dataset Task Size (train/test) Modality Notes
UCI Human Activity Recognition [33] 6-class activity classification (walking, walking upstairs, walking downstairs, sitting, standing, lying) Not reported (30 volunteers, ages 19–48) Smartphone IMU (waist-worn) Features normalized; partitioned across clients [§V-B, §IV-A]

Pipeline diagram

flowchart LR
    subgraph Client
        A[Local train DFCN] --> B[Compute delta M]
        B --> P[Prune below tau]
        P --> Q[Adaptive quantize]
        Q --> E[CKKS encrypt]
    end
    subgraph Cluster
        AGG[Intra-cluster sum]
    end
    subgraph Server
        S[Aggregate clusters]
        S --> COST[Security overhead cost]
        COST --> CONV[Early stop check]
        CONV --> G[Update global model]
    end
    E -. enc update .-> AGG
    AGG -. enc cluster sum .-> S
    G -. global model .-> A

Pipeline steps (text)

  1. Preprocess: normalize features and partition training data across clients [§IV-A].
  2. Initialize global model with random weights; group clients into C clusters; set up CKKS context [§IV-B].
  3. Each client trains its DFCN locally on private data and computes delta M = M_l - M_g [§III-A, §III-B].
  4. Prune updates whose absolute value is below threshold tau (set to 0.1) [§III-C, §V-A].
  5. Adaptively quantize pruned updates (12 bits for FLHC-QPR), scaling by s = max( delta M ) [§III-D, §V-A].
  6. Encrypt the scaled, quantized update under CKKS and send it (with scaling factor and training loss) to the cluster/server [§III-E, Algorithm 1].
  7. Server aggregates encrypted updates per cluster and then across clusters without decrypting them [§III-F, Algorithm 2].
  8. Server computes security overhead cost = alphaM + betaL + gammaC + deltaE and evaluates loss/accuracy [§III-I, Algorithm 2].
  9. Apply early-stopping: if L_t - L_prev < epsilon for k consecutive rounds, halt [§IV-C, Algorithm 2].
  10. Update global model M_g^{t+1} = M_g^t + (1/K) * sum_k delta M_k and broadcast back to clients [§III-G].

Architecture diagram

flowchart LR
    I([Input features]) --> FC1[FC 128]
    FC1 --> R1[ReLU]
    R1 --> FC2[FC 64]
    FC2 --> R2[ReLU]
    R2 --> FC3[FC 32]
    FC3 --> R3[ReLU]
    R3 --> FC4[FC C]
    FC4 --> O([Softmax])

Results

Headline metrics (UCI HAR). Baselines: FLH (basic FL+HE), FLHC (FLH with clustering [15]), FedSafe (FE-based, no KDC [21]). Hardware: not reported (Python implementation).

Metric This paper (FLHC-QPR) Baseline Hardware
Best accuracy (11 rounds) 89.86% FLHC 86.89%; FLH 81.39%; FedSafe 81.50% [§V-D, Fig. 3a] Not reported
Peak accuracy (privacy threshold = 5) 90.09% FLHC 91.31%; FedSafe 80.82%; FLH 92.89% [§V-D, Fig. 4a] Not reported
Latency (client-count sweep) 0.62–8.39 s FLH 0.80–8.86 s; FLHC 0.66–8.69 s; FedSafe 0.80–8.78 s [§V-D, Fig. 2b] Not reported
Latency (rounds sweep) ~1.65–1.69 s Comparable to FLH/FLHC [§V-D, Fig. 3b] Not reported
Memory overhead (rounds=11) -16.4% vs FLHC; -31% vs FLH; -38.6% vs FedSafe [§V-D, Fig. 3c] — Not reported
Memory overhead (privacy threshold = 5) -31.5% vs FLH; -24.7% vs FLHC [§V-D, Fig. 4c] — Not reported
Security overhead cost 0.4391 (rounds=11); 46.93 (epsilon=5) FLHC 0.4465; FLH 0.4451; FedSafe 0.4566; FedSafe 47.53 [§V-D] Not reported
Average accuracy gain +8.4% over FLH; +5.4% over FLHC — —
Average latency reduction -16.2% vs FLHC; -11.7% vs FLH — —
Average memory reduction -26.5% vs FLHC; -35.1% vs FLH — —
Average security-overhead reduction -2.7% vs FLHC; -2.5% vs FLH — —

Optimal hyperparameters from trade-off study: 12-bit quantization (acc 0.9556), pruning threshold 0.1 (acc 0.9233), 4 clusters (acc 0.9070) [§V-E].

Limitations and assumptions

Code and artifacts

Not released. No repository URL given in the paper.

Extra diagrams (optional)

Threat model

flowchart LR
    C[Client i] -- enc delta M_i --> CL[Cluster head]
    CL -- enc cluster sum --> S[Honest-but-curious server]
    S -- updated global model --> C

Federated round

sequenceDiagram
    participant C as Clients
    participant K as Cluster
    participant A as Server
    C->>C: Local train DFCN
    C->>C: Prune + adaptive quantize
    C->>K: CKKS-encrypt delta M
    K->>K: Homomorphic sum within cluster
    K->>A: Encrypted cluster update
    A->>A: Aggregate clusters + compute security cost
    A->>A: Early-stop check
    A->>C: Updated global model

Open questions