bcfl-optimizer

Energy-efficient and latency-aware device selection for Blockchain-Enabled Federated Learning.

Reference implementation of:

D. Kushwaha, M. Kalavadia, V. Hegde and O. J. Pandey, "Energy-Efficient and Latency-Aware Blockchain-Enabled Federated Learning for Edge Networks," IEEE Transactions on Circuits and Systems II: Express Briefs, vol. 71, no. 3, pp. 1126–1130, March 2024. doi:10.1109/TCSII.2023.3322340

Adding a blockchain to federated learning removes the single point of failure of a central server, but miner verification, block generation and accidental forks add latency and energy to every round. Given a network of edge devices and miners, this library picks the subset of devices (and so of miners) that jointly minimizes a round's latency and energy.

Installation

pip install bcfl-optimizer          # numpy only
pip install "bcfl-optimizer[plot]"  # + matplotlib, for reproduce/

Quick start

from bcfl import DevicePool, MinerPool, optimize

devices = DevicePool.random(n=50, num_miners=15, seed=0)  # Table I ranges
miners = MinerPool.random(m=15, seed=0)

result = optimize(devices, miners, s_d=30, beta=0.5)
result.selected_devices, result.selected_miners  # S_D, S_M
result.latency, result.energy, result.a_fork     # T, E, A_fork of one round

To describe your own network, construct DevicePool and MinerPool directly; each field is documented in its docstring. Link rates are given in bit/s, and shannon_rate(bandwidth, snr) converts bandwidth and SNR to a rate.

What is implemented

optimize follows Algorithm 1. Phase 1 enumerates the miner sets with minimum A_fork that can host S_D devices. Phase 2 (the paper's tree search) is an exact branch-and-bound: T and E never decrease when a device is added, so a partial selection's cost bounds all of its completions. tests/test_optimizer.py checks the result against exhaustive search.

Modelling choices

Where the paper leaves a detail open, the implementation follows the simulation that produced its figures:

• A_fork is a whole number of block-generation attempts: ceil(1 / (1 - p_fork)).
• E[t_bg] = 1 / (λ |S_M|), the expected time for the first of |S_M| miners to find a block.
• The block is propagated by the selected miner with the fastest verification (smallest τ^comp).
• Every selected miner hosts at least one selected device, so S_M is exactly the set of miners associated with S_D.
• DevicePool.random and MinerPool.random sample the simulation's ranges:
  • powers of 20–30 dBm (devices) and 25–30 dBm (miners);
  • Rayleigh-faded channel gains, 2.5·10⁻⁵·Exp(1) for device links and 10⁻⁵·Exp(1) between miners;
  • τ^comp of 2–2.5 ms, P^ver of 0.3–0.5 W, P_bg of 0.5–1 W;
  • model and block sizes of 0.5–1·10⁶ bits.

Reproducing the paper

reproduce/parameters.json holds the exact parameters of the simulation behind Figs. 3–4 (N_D = 50 devices, N_M ∈ {15, 20, 25} miners). reproduce/original_selections.json holds the selections that simulation recorded.

python reproduce/reproduce.py --plot   # ~2 min; writes reproduce/fig3.png

• Selected miners and A_fork: for all 30 settings (N_M ∈ {15, 20, 25} × S_D ∈ {5, …, 50}), the number of selected miners and A_fork equal the original simulation's.
• Energy: between −2% and +16% of the original (median +3%).
• Latency: 5–50% higher, because this implementation includes the verification queuing delay of Eqs. (4)–(5), which the original simulation did not model.
• Selected devices: the miner sets and A_fork agree, but the chosen devices within them can differ. The objective now includes the queuing delay, and it is Eq. (11) exactly.

Federated training with Flower

A Flower baseline that uses this selection inside federated training on Google Speech Commands is proposed in flwrlabs/flower#6799.

Development

pip install -e ".[test,plot]"
pytest

Citation

@article{kushwaha2024energy,
  title   = {Energy-Efficient and Latency-Aware Blockchain-Enabled Federated Learning for Edge Networks},
  author  = {Kushwaha, Deepali and Kalavadia, Manan and Hegde, Vinayaka and Pandey, Om Jee},
  journal = {IEEE Transactions on Circuits and Systems II: Express Briefs},
  volume  = {71},
  number  = {3},
  pages   = {1126--1130},
  year    = {2024},
  doi     = {10.1109/TCSII.2023.3322340}
}

License

MIT License. See LICENSE.