EE-DFL: Energy-Efficient Decentralized Federated Learning

Updated 10 June 2026

Energy-efficient Decentralized Federated Learning is a system framework that minimizes energy use in collaborative on-device machine learning while balancing computation, communication, and privacy constraints.
Key methodologies include lightweight models, energy-aware resource allocation, and adaptive clustering to accelerate convergence and boost learning robustness.
Practical deployments leverage diverse topologies—from star networks to fully decentralized D2D graphs—tailored to meet energy constraints and dynamic network conditions.

Energy-efficient Decentralized Federated Learning (EE-DFL) refers to a class of federated learning (FL) schemes that explicitly minimize energy consumption—computation and communication—across distributed, often battery-powered, devices in the absence of a single central coordinator. EE-DFL encompasses system architectures, algorithmic frameworks, and optimization techniques designed to enable collaborative on-device learning under stringent energy, latency, and privacy constraints. Recent work spans a continuum from wireless edge networks with single-hop base stations to fully device-to-device (D2D) decentralized topologies. State-of-the-art solutions leverage lightweight models, energy-aware resource allocation, sparse or clustered aggregation, and privacy-preserving mechanisms to accelerate convergence, reduce total energy, and improve learning robustness.

1. System Architectures and Decentralization Topologies

EE-DFL manifests in topologies ranging from star-based (one central server) to hybrid clustered, overlapped multi-hop D2D, or fully decentralized peer-to-peer networks.

Single-cell wireless edge network: A base station acts as a model aggregator, orchestrating federated hyperdimensional computing (HDC) with differential privacy-enhanced updates (Ding et al., 25 Feb 2026).
Overlapped D2D clustering: Devices organize into clusters with overlapping coverage (e.g., within a physical radius), with cluster heads (CHs) coordinating intra-cluster aggregation and bridge devices (BDs) linking clusters, thus eliminating reliance on a global aggregator and enabling true decentralized consensus (Al-Abiad et al., 2022).
Device-to-device (D2D) graphs with energy harvesting: Each device communicates with graph neighbors, using direct D2D transfer and decentralized model aggregation, sometimes with support for harvesting ambient energy for long-term sustainable operation (Zhang et al., 15 Feb 2026).
Adaptive hierarchical clustering for 6G: Clusters are dynamically formed, with cluster heads relaying models between devices and aerial base stations (e.g., UAVs), reducing communication power and extending reach in large-scale networks (Khowaja et al., 2022).
Time-varying mixing matrix frameworks: Topologies and communication patterns are allowed to change across rounds to optimize per-node energy under broadcast constraints, balancing convergence rate and energy expenditure (Zhang et al., 30 Dec 2025).

These architectures are unified by a core challenge: managing the energy/accuracy/latency trade-off in inherently resource-diverse, latency- and bandwidth-constrained, and possibly mobile networks.

2. Energy Consumption Models and Budgeted Resource Allocation

EE-DFL frameworks model total device energy as the sum of computation and communication costs per round, often parameterized by device-specific hardware characteristics (CPU frequency, transmit power), wireless channel states, and local batch sizes.

Computation energy: $E_{i}^{\text{comp}} = \gamma D_i d C_i f_i^2$ , where $D_i$ is the sample count, $d$ the model (e.g., HV) dimension, $C_i$ per-dimension compute, and $f_i$ the CPU frequency; rapid scaling of energy with frequency is exploited for DVFS savings in local training phases (Ding et al., 25 Feb 2026, Zou et al., 2021).
Communication energy: $E_{i}^{\text{trans}}=p_i t_i$ , with $t_i$ set by model size (e.g., HV or NN weight vector), bandwidth allocation, and channel power gain; proper assignment of transmit powers $p_i$ and per-user bandwidths $b_i$ is critical (Khowaja et al., 2022).
Broadcast and D2D link energy: Models incorporate per-broadcast energy $c_i^b$ and activate only a subset of communication edges at each iteration to respect per-device budgets (Zhang et al., 30 Dec 2025).
Energy harvesting: The device’s available energy evolves as $D_i$ 0, enforcing causality constraints and requiring dynamic scheduling to avoid battery depletion (Zhang et al., 15 Feb 2026).

Joint optimization formulations select model dimensions $D_i$ 1, users’ CPU frequencies $D_i$ 2, and transmit powers $D_i$ 3 so as to minimize aggregate or per-node energy, often subject to delay/latency, accuracy, or privacy constraints (Ding et al., 25 Feb 2026, Al-Abiad et al., 2022, Zhang et al., 30 Dec 2025).

3. Algorithmic Strategies for Energy Minimization

Approaches span multi-layered optimization and algorithmic enhancements to jointly consider energy efficiency, convergence speed, and privacy.

Two-stage discrete-continuous search: Outer enumeration (e.g., over HDC model dimension $D_i$ 4) coupled with inner resource allocation (CPU frequency and transmit powers) yields global energy minima with algorithmic complexity $D_i$ 5 (Ding et al., 25 Feb 2026).
Alternating scheduling and frequency allocation: Devices are scheduled via conflict-graph-based minimum-weight independent set (MWIS), while optimal per-device frequency is set to exactly meet latency constraints, iteratively alternating until convergence (Al-Abiad et al., 2022).
Reinforcement learning: Decentralized Q-learning or policy-iteration solutions optimize device selection, local update targets (e.g., number of epochs, DVFS state), or power allocation under non-IID data and dynamic network states (Kim et al., 2021, Zhang et al., 15 Feb 2026).
Sparsity and mask pruning: Progressive “sparse-to-sparser” masking schedules (DA-DPFL) reduce local computation and communication by dynamically dropping parameters, with personalized regrowth to preserve accuracy (Long et al., 2024).
Energy-minimizing aggregation: MST-based (minimum spanning tree) or ring all-reduce approaches are selected based on network structure and link energies to minimize consensus energy cost while ensuring model alignment (Yan et al., 2024).
Time-varying mixing matrices: Training is split into multiple phases, each phase using a stochastic mixing matrix distribution under a prescribed energy budget; early phases may use sparse mixing to save energy, late phases become denser for rapid final convergence (Zhang et al., 30 Dec 2025).

4. Convergence, Robustness, and Trade-offs

Energy-efficient decentralization introduces fundamental trade-offs between per-round computation/communication cost, number of global rounds to reach a target accuracy, and learning robustness under device and data heterogeneity.

Model complexity trade-off: Higher-dimensional HDC models (larger $D_i$ 6) speed up convergence (reduce number of rounds $D_i$ 7) but have higher per-round energy; the optimum is a moderate $D_i$ 8 (e.g., $D_i$ 9 with minimum total energy) (Ding et al., 25 Feb 2026).
Cluster overlap and BDs: Increased overlap accelerates diffusion of model updates but increases cumulative BD compute and communication energy; optimal overlap is typically low degree (1–2 shared BDs between clusters) (Al-Abiad et al., 2022).
Adaptive local computation: Devices with a larger energy budget or greater computational efficiency are assigned more local steps per round; energy-scarce nodes participate less or only in sparse gossip (Yan et al., 2024).
Sparse mixing phases: Multi-phase mixing reduces per-node max energy by as much as 20–40% versus stationary or single-phase full-mixing baselines while achieving similar test error (Zhang et al., 30 Dec 2025).
Energy harvesting and sustainability: Dynamic device scheduling is needed to stay within battery constraints and packet loss trade-offs; decentralized policy-iteration achieves near-centralized performance and sustainable device lifetime (Zhang et al., 15 Feb 2026).

5. Privacy, Security, and Personalization Considerations

EE-DFL incorporates privacy-preserving techniques to maintain data confidentiality in resource-constrained, distributed environments:

Differential privacy (DP): Zero-concentrated DP (zCDP) is applied to model updates with Gaussian noise prior to transmission, ensuring a calibrated $d$ 0 privacy budget at fixed energy overhead (Ding et al., 25 Feb 2026).
Decremental/Incremental learning: Local selective forgetting (e.g., via model decrement and associated CPU frequency scaling) reduces energy and mitigates privacy risks from stale data exposure (Zou et al., 2021).
Sparse masking and personalization: Dynamic mask pruning, as in DA-DPFL, enables local adaptation to non-IID data, providing both energy and accuracy advantages, and supports per-client local model evolution (Long et al., 2024).

Security implications (e.g., poisoning attacks in fully decentralized settings) are not the stated focus in the reviewed data, but the tight coupling of privacy preservation and energy efficiency is emphasized.

6. Empirical Evaluation and Quantitative Impact

EE-DFL frameworks demonstrate substantial energy, time, and convergence improvements compared to traditional decentralized and centralized FL baselines.

Framework/Paper	Energy Savings	Speedup	Accuracy Impact
FL-HDC-DP (Ding et al., 25 Feb 2026)	Up to 83.3% vs. baselines	Faster convergence	≈88% maintained
FL-EOCD (Al-Abiad et al., 2022)	35–55% vs. star/hier. FL	30–50% per-round time	≈93% reached by all
EH-DFL (Zhang et al., 15 Feb 2026)	Sustainable under energy harvesting	20% faster than centralized myopic	Matches centralized
DA-DPFL (Long et al., 2024)	Up to 5×	Fewer rounds	2–3% higher
DEAL (Zou et al., 2021)	75.6–82.4% vs. baseline	2–4× completion time	≤12% worst-case loss
DBFL (Khowaja et al., 2022)	≈25–30% per round	25–36 fewer rounds	+7.4% absolute
Time-varying mixing (Zhang et al., 30 Dec 2025)	20–40% per-node	Comparable to baselines	No loss

Benchmarks include MNIST, CIFAR-10, CIFAR-100, and real IoT/biomedical datasets, often under strong non-IID splits and device heterogeneity. Consistently, EE-DFL demonstrates that targeted energy-aware designs yield comparable or better learning outcomes at a fraction of the resource cost.

7. Deployment Guidelines and Open Challenges

Best practices highlighted include:

Cluster sizing: Choose clusters of size $d$ 1 in large-scale deployments to optimize the balance between intra-cluster comm cost and CH-to-server aggregation (Khowaja et al., 2022).
Adaptive scheduling: Re-run cluster and scheduling logic periodically in mobile or time-varying settings to retain efficiency (Al-Abiad et al., 2022).
Algorithm tuning: Set per-iteration latency above the maximum local compute time for at least one local update per device (Yan et al., 2024); tune HDC dimension and mask pruning thresholds empirically to achieve best trade-offs (Ding et al., 25 Feb 2026, Long et al., 2024).
Practical constraints: Quantization or compression of transmitted models further reduces communication energy (Yan et al., 2024). Sparse aggregation or partial-mixing should be densified in later training to accelerate final convergence without overshooting per-node energy budgets (Zhang et al., 30 Dec 2025).
EH design: Size batteries to smooth out EH variability and avoid frequent outages in energy-harvesting scenarios (Zhang et al., 15 Feb 2026).

Remaining challenges include extending convergence theory to general non-Gaussian EH models, supporting highly dynamic graphs, further reducing the overhead of dynamic mask synchronization, and scaling to ultra-large device populations with minimal orchestration.

EE-DFL represents a critical advance toward sustainable, privacy-preserving, and robust collaborative machine learning at the edge, fusing innovations in communication topology, computation scheduling, privacy, and device-level optimization for maximal efficiency in decentralized environments (Ding et al., 25 Feb 2026, Al-Abiad et al., 2022, Zhang et al., 15 Feb 2026, Yan et al., 2024, Zhang et al., 30 Dec 2025, Khowaja et al., 2022, Zou et al., 2021, Kim et al., 2021, Long et al., 2024).