TL;DR

The paper proposes four federated-learning protocol variants (OUCM, OECM, MUCM, MECM) that combine FHE with user-clustering on the N-BaIoT IoT botnet dataset, reporting communication-overhead reductions of 80.15–89.98% and up to a 70.38% latency reduction over a baseline FL approach while preserving accuracy around 0.997 for three of the four variants [Abstract; §VI].

Problem and motivation

IoT-enabled smart cities exchange large volumes of sensitive sensor data and are vulnerable to attacks; pure FL transmits model parameters over insecure channels which can leak information [§I, pp. 4289–4290]. The authors target this leakage by encrypting model updates with FHE so the central server aggregates ciphertexts and never sees plaintext weights [§I, p. 4290; §III]. Threat model is implicit: the central aggregation server (and the network) is untrusted to view weights, while users keep their secret keys locally; clustering aims to reduce the amount of data transmitted between users and the server [§II, p. 4291; §III, p. 4291].

Key contributions

FHE setup

ML setup

Datasets

Dataset Task Size (train/test) Modality Notes
N-BaIoT [34] IoT botnet / anomaly detection Not reported explicitly (split into Dtrain/Dtest [Algorithms 1–4]) Network traffic from five device types — baby monitor (IoT_1), doorbells (IoT_2), thermostat (IoT_3), security cameras (IoT_4), webcam (IoT_5) Benchmark from UCI ML repository [§IV, p. 4295]
WUSTL-EHMS-2020 [41] Intrusion + injection-attack detection in Internet of Medical Things 14,272 benign + 2,046 attack samples [§V, p. 4298] Hybrid: 35 network-flow metrics + 8 patient biometric features (44 features total) Case study; man-in-the-middle (spoofing) and data-injection attacks [§V]

Pipeline diagram

flowchart LR
    subgraph Clients [IoT devices Ui]
        L1[Local CNN train] --> W1[Weight update LM_theta]
        W1 --> E1[FHE encrypt]
    end
    subgraph Cluster [Cluster matrix CMr / GCM]
        CM[Aggregate ciphertext]
    end
    subgraph Server [Central aggregator]
        AGG[FedAvg on ciphertext] --> GM[Encrypted global model]
    end
    subgraph Decrypt [Client side]
        D1[FHE decrypt] --> U1[Update local CNN]
    end
    E1 -. ciphertext .-> CM
    CM -. ciphertext .-> AGG
    GM -. ciphertext .-> D1

Pipeline steps (text)

  1. Randomly distribute the N users into K clusters using function f(U, K) [§III-A, definitions & initialization].
  2. Each user Ui trains its local CNN on its private dataset Di starting from the current global parameters GM_θ(t) [§III-A, local training].
  3. Compute the weight update LM_θi(t+1) = LM_θi(t+1) − LM_θi(t) [Algorithms 1–4, line 7].
  4. Depending on the variant, encrypt either (a) the whole global cluster matrix GCM after collecting plaintext updates (OUCM), (b) each user’s update before sending (OECM), (c) per-cluster matrices CMr (MUCM), or (d) per-user updates routed into per-cluster matrices (MECM) [§III-A, Approaches 1–4].
  5. Transmit ciphertexts to the central server [Algorithms 1–4, “Send … ⇒ server”].
  6. Server runs FedAvg homomorphically: GM = (1/N) Σ Enc_*, without decrypting [§III-A; Algorithms 1–4, line 11/12].
  7. Server returns the (encrypted) updated global model GM to all users [Algorithms 1–4].
  8. Each user decrypts with its FHE secret key and overwrites its local model: LMi = Dec_FHE(GM) [Algorithms 1–4, line 13].
  9. Repeat for T rounds (30 epochs in experiments, 100 users, 5 clusters) [§IV-A, p. 4296].

Architecture diagram

flowchart LR
    I([IoT traffic features]) --> CNN[CNN local model]
    CNN --> O([Anomaly / benign])

Layer-by-layer details (number of conv layers, filter counts, kernel sizes, FC widths, activation functions) are not reported in the paper text; the model is referred to only as “a convolutional neural network (CNN) algorithm” [§III-A, Approach 1, p. 4293].

Results

Reported headline numbers (averages across five N-BaIoT device subsets):

Metric Baseline FL OUCM OECM MUCM MECM Hardware
Accuracy (avg) 0.996–0.998 [Table III] 0.911 (range 0.840–0.985) [Table IV] 0.997 (0.997–0.998) [Table V] 0.998 (0.997–0.998) [Table VI] 0.997 (0.995–0.998) [Table VII] 2× AMD EPYC 7742, 128 cores / 256 threads, 256 GB RAM [§IV-A]
F1-score high (text) 0.949 avg (0.912–0.992) [Table IV] 0.998 avg (0.998–0.999) [Table V] high (text) [Table VI] high (text) [Table VII] same
Latency (s per epoch) 168–220 [§IV-A baseline] 144–197 [§IV-A OUCM] 170–221 [§IV-A OECM] 37–91 [§IV-A MUCM] ~170–224 [§IV-A MECM] same
Comm overhead (total) 104 MB [§IV-A baseline] 1.04 MB [§IV-A OUCM] 2.06 MB [§IV-A OECM] 1.2875 MB [§IV-A MUCM] 2.06 MB [§IV-A MECM] same
Comm-overhead reduction vs baseline — 89.98% 80.15% 87.71% 80.15% [§VI, conclusion]

The MUCM variant gives the headline latency reduction of 70.38% (37–91 s vs 168–220 s baseline) [Abstract; §IV-A]. MECM has the lowest security cost and best WUSTL-EHMS-2020 case-study accuracy/F1, but higher latency than MUCM [§IV-B; §V; Fig. 8; Table IX].

Limitations and assumptions

Code and artifacts

Public GitHub repository: https://github.com/Artifitialleap-MBZUAI/Secure-Federated-Learning-with-Fully-Homomorphic-Encryption-for-IoT-Communications [Abstract; §III-A end]. License not reported.

Extra diagrams (optional)

Threat model

flowchart LR
    U[IoT user Ui] -- enc(weight update) --> CM[Cluster matrix]
    CM -- enc(aggregated update) --> S[Untrusted server]
    S -- enc(global model) --> U

Federated round

sequenceDiagram
    participant U as Users (CNN clients)
    participant CM as Cluster matrix
    participant S as Server (aggregator)
    U->>U: Local CNN train; compute LM_theta(t+1) - LM_theta(t)
    U->>CM: Send weight update (encrypted or plaintext, per variant)
    CM->>S: Send (encrypted) cluster matrix
    S->>S: FedAvg over ciphertexts: GM = (1/N) sum Enc_*
    S->>U: Return encrypted global model GM
    U->>U: Decrypt and update local CNN

Activation approximation

Not applicable — activations run locally on plaintext during CNN training; only the weight-update aggregation is homomorphic [§III, p. 4291].

Open questions