Green Federated Learning

• Green Federated Learning is a paradigm that optimizes distributed training to reduce energy consumption and carbon emissions while ensuring competitive model accuracy.
• It employs techniques such as communication compression, adaptive client selection, and model-size adaptation to balance resource use and learning efficiency.
• The field emphasizes standardized measurement of energy and CO2e across training phases, enabling reproducible and environmentally sustainable federated systems.

Searching arXiv for recent and foundational papers on Green Federated Learning. Green Federated Learning denotes federated learning designed and evaluated with explicit attention to environmental sustainability: reducing energy consumption, communication overhead, and carbon emissions while preserving competitive model performance and acceptable training time (Yousefpour et al., 2023). In the recent literature, the term covers both algorithmic mechanisms—such as compression, client selection, adaptive model sizing, and resource allocation—and methodological work on how energy and CO$_2$e should be measured and reported in federated systems (Thakur et al., 2024). The field is therefore not a single algorithmic family but a systems-and-learning perspective on federated optimization, spanning cross-device mobile FL, IoT and edge networks, wireless and Cloud-RAN settings, cross-silo medical FL, and carbon-aware scheduling across geographically distributed clients (Abbasi et al., 2024).

1. Conceptual foundations

Green Federated Learning emerged from the broader distinction between “Red AI” and “Green AI,” where the latter treats efficiency and environmental impact as first-class criteria alongside predictive quality (Thakur et al., 2024). In FL, this shift is especially consequential because training is distributed across heterogeneous clients whose energy sources, hardware characteristics, communication links, and local data properties differ substantially (Yousefpour et al., 2023). Unlike centralized training, which can be concentrated in strategically placed data centers, cross-device FL may involve very large numbers of end-user devices with diverse energy profiles and network conditions (Yousefpour et al., 2023).

Within this framing, Green FL is commonly cast as a multi-objective design problem. One influential formulation defines it as optimizing FL parameters and design choices to minimize carbon emissions while maintaining competitive performance and acceptable time-to-train (Yousefpour et al., 2023). The survey literature broadens this to encompass reduction of communication and computation energy, limitation of carbon footprint, and preservation of convergence behavior in wireless, IoT, and edge environments (Thakur et al., 2024). This suggests that Green FL should be understood simultaneously as a systems problem, a resource-allocation problem, and a learning-theoretic problem.

A recurring theme is that “greenness” in FL depends not only on what model is trained, but also on where, when, and how training is performed. Carbon-aware work exploits regional and temporal variation in electricity carbon intensity across clients (Arputharaj et al., 10 Sep 2025). Energy-aware wireless work instead emphasizes transmission power, bandwidth, CPU frequency, and quantization or compression levels (Wang et al., 2023). Data-centric work shifts attention to the environmental consequences of client data volume and quality, arguing that training on less and better data can reduce emissions while preserving or improving accuracy (Sabella et al., 23 Jul 2025).

2. Measurement, accounting, and evaluation criteria

Green FL depends on explicit accounting of both energy and carbon. A standardized methodology proposed for FL CO$_2$e tracking decomposes the lifecycle into initialization, per-round training, evaluation, and idle or coordination phases, instrumented through NVFlare and CodeCarbon, and supplements measured compute emissions with modeled communication emissions based on transmitted update sizes (Tapp et al., 30 Jan 2026). In that framework, communication energy is estimated as

$$E_{\mathrm{comm}} = 2 \cdot D_{\mathrm{GB}} \cdot I_{\mathrm{net}},$$

and communication carbon as

$$C_{\mathrm{comm}} = E_{\mathrm{comm}} \cdot F_{\mathrm{grid}},$$

with the factor $2$ accounting for both uplink and downlink model transfer (Tapp et al., 30 Jan 2026).

The methodological literature emphasizes that carbon results are otherwise difficult to compare because studies often adopt different measurement boundaries, include different hardware components, or omit communication and coordination effects (Tapp et al., 30 Jan 2026). In production-scale cross-device FL, direct measurements show that client compute and client–server communication dominate total emissions, while server-side compute contributes only a small fraction (Yousefpour et al., 2023). One key empirical result is that, for synchronous FL, carbon emissions scale linearly with concurrency times rounds-to-target, whereas for asynchronous FL they scale linearly with concurrency times time-to-target (Yousefpour et al., 2023).

Several papers operationalize carbon or energy at the client level. In carbon-aware scheduling, the carbon cost of selecting client $c$ at time slot $t$ is modeled as

$g_c^{(t)} = E_c \cdot \mathrm{CI}_c^{(t)},$

where $\mathrm{CI}_c^{(t)}$ is local carbon intensity and $E_c$ is energy per slot (Arputharaj et al., 10 Sep 2025). In model-size adaptation, per-client, per-round carbon is written as

$_2$0

so emissions scale quadratically with the width-scaling factor $_2$1 and linearly with the client’s carbon intensity $_2$2 (Abbasi et al., 2024). In mobile-edge generative FL, energy is decomposed into computation and communication terms,

$_2$3

and optimized under latency and quantization constraints (Lai et al., 2024).

The metrics used across the literature therefore include total energy consumption, CO$_2$4e emissions, transmitted model size, communication volume, convergence speed, wall-clock training time, and task performance such as test accuracy, perplexity, MAE, PSNR, SSIM, or FID, depending on the application (Yousefpour et al., 2023). This measurement-centric strand treats standardized reporting as a prerequisite for reproducible Green FL evaluation rather than an optional add-on (Tapp et al., 30 Jan 2026).

3. Core optimization levers

The Green FL literature converges on a small number of technical levers. These include reducing communication volume, reducing local computation, optimizing client participation, adapting model size or precision to client conditions, and exploiting carbon-intensity variation across space and time (Thakur et al., 2024).

Lever	Representative mechanism	Example paper
Communication compression	pruning, quantization, Huffman encoding	(Colybes et al., 13 Apr 2026)
Energy-aware compression	fine-grained gradient compression with device-specific ratios	(Li et al., 2021)
Model-size adaptation	ordered dropout with client-specific widths	(Abbasi et al., 2024)
Carbon-aware scheduling	slack time and time-slot selection under carbon budgets	(Arputharaj et al., 10 Sep 2025)
Joint compute–communication control	CPU frequency, transmit power, quantization, fronthaul allocation	(Wang et al., 2023)

Compression-centric work treats communication overhead as a dominant source of inefficiency. A “Full Compression Pipeline” integrates unstructured pruning, post-training codebook quantization, and Huffman encoding in the client-to-server direction, and evaluates their combined effect through a unified model cost that aggregates communication and computation overhead (Colybes et al., 13 Apr 2026). In a representative scenario with ResNet-12 on CIFAR-10, ten clients, and a 2 Mbps link, the method achieves more than $_2$5 model-size reduction with only a $_2$6 accuracy drop and yields training that is more than $_2$7 faster (Colybes et al., 13 Apr 2026). This suggests that under bandwidth-constrained conditions, end-to-end compression pipelines can reduce both transmitted bits and total training time.

Gradient-compression work in mobile edge computing focuses on device-side energy. FedGreen introduces kernel-wise sparsification, stochastic quantization, and lossless encoding, and jointly optimizes a device-specific compression ratio and CPU frequency (Li et al., 2021). Under an 80% test-accuracy requirement, the method reduces total device energy consumption by at least $_2$8 relative to baseline schemes (Li et al., 2021). A related mobile-edge generative framework quantizes diffusion-model weights before upload and derives the bound

$_2$9

which links device-specific model-norm bounds, quantization levels, and quantization error (Lai et al., 2024). This provides a direct mechanism for adapting compression to resource-constrained devices.

Model-size adaptation is another prominent lever. FedGreen assigns smaller ordered-dropout submodels to clients in higher-carbon regions and larger submodels to greener clients, explicitly modeling the resulting carbon as a function of width scaling (Abbasi et al., 2024). In experiments on EMNIST with $$E_{\mathrm{comm}} = 2 \cdot D_{\mathrm{GB}} \cdot I_{\mathrm{net}},$$0 clients, reducing the single-cluster scaling factor from $$E_{\mathrm{comm}} = 2 \cdot D_{\mathrm{GB}} \cdot I_{\mathrm{net}},$$1 to $$E_{\mathrm{comm}} = 2 \cdot D_{\mathrm{GB}} \cdot I_{\mathrm{net}},$$2 under $$E_{\mathrm{comm}} = 2 \cdot D_{\mathrm{GB}} \cdot I_{\mathrm{net}},$$3 changes rounds from $$E_{\mathrm{comm}} = 2 \cdot D_{\mathrm{GB}} \cdot I_{\mathrm{net}},$$4 to $$E_{\mathrm{comm}} = 2 \cdot D_{\mathrm{GB}} \cdot I_{\mathrm{net}},$$5 but reduces carbon cost from $$E_{\mathrm{comm}} = 2 \cdot D_{\mathrm{GB}} \cdot I_{\mathrm{net}},$$6 to $$E_{\mathrm{comm}} = 2 \cdot D_{\mathrm{GB}} \cdot I_{\mathrm{net}},$$7; at $$E_{\mathrm{comm}} = 2 \cdot D_{\mathrm{GB}} \cdot I_{\mathrm{net}},$$8, rounds rise to $$E_{\mathrm{comm}} = 2 \cdot D_{\mathrm{GB}} \cdot I_{\mathrm{net}},$$9 while cost drops to $$C_{\mathrm{comm}} = E_{\mathrm{comm}} \cdot F_{\mathrm{grid}},$$0 (Abbasi et al., 2024). The paper interprets this as a carbon–convergence trade-off governed by the mean and standard deviation of client model sizes.

4. Carbon-aware scheduling and client participation

A distinct branch of Green FL treats carbon intensity as a scheduling signal. In this line of work, clients are geographically distributed, and each client has a local carbon-intensity time series $$C_{\mathrm{comm}} = E_{\mathrm{comm}} \cdot F_{\mathrm{grid}},$$1 (Arputharaj et al., 10 Sep 2025). The central mechanism is “slack time”: if training is allowed to extend modestly beyond the nominal horizon, local rounds can be shifted into lower-carbon periods (Arputharaj et al., 10 Sep 2025). On real-world carbon-intensity traces, single clients can reduce emissions by about $$C_{\mathrm{comm}} = E_{\mathrm{comm}} \cdot F_{\mathrm{grid}},$$2 with $$C_{\mathrm{comm}} = E_{\mathrm{comm}} \cdot F_{\mathrm{grid}},$$3 hours of slack and up to about $$C_{\mathrm{comm}} = E_{\mathrm{comm}} \cdot F_{\mathrm{grid}},$$4 with $$C_{\mathrm{comm}} = E_{\mathrm{comm}} \cdot F_{\mathrm{grid}},$$5 hours; across 54 regions, about $$C_{\mathrm{comm}} = E_{\mathrm{comm}} \cdot F_{\mathrm{grid}},$$6 of clients reduce carbon intensity by at least $$C_{\mathrm{comm}} = E_{\mathrm{comm}} \cdot F_{\mathrm{grid}},$$7 and about $$C_{\mathrm{comm}} = E_{\mathrm{comm}} \cdot F_{\mathrm{grid}},$$8 reduce it by at least $$C_{\mathrm{comm}} = E_{\mathrm{comm}} \cdot F_{\mathrm{grid}},$$9 at maximum slack (Arputharaj et al., 10 Sep 2025).

Carbon-aware scheduling introduces learning trade-offs because low-carbon selection is not statistically neutral. To manage this, the scheduler in “Green Federated Learning via Carbon-Aware Client and Time Slot Scheduling” uses an $2$0-fair allocation objective under a global carbon budget, corrects participation bias with Unbiased FedAvg, and appends a full-participation fine-tuning phase to mitigate temporal and spatial correlation effects (Arputharaj et al., 10 Sep 2025). Under tight carbon constraints, the scheduler improves accuracy over slack-agnostic baselines by $2$1 and $2$2 percentage points at budgets of approximately $2$3 and $2$4 of a high reference budget, respectively (Arputharaj et al., 10 Sep 2025). Under medium and high budgets, gains are smaller but remain positive (Arputharaj et al., 10 Sep 2025). This suggests that carbon-aware scheduling is most valuable precisely when emissions are tightly constrained.

Client selection has also been studied from a green perspective in over-the-air FL and mobile edge settings. A multi-attribute client-selection framework for OTA-FL uses the Grey Wolf Optimizer to trade off model loss, convergence time, energy efficiency, fairness, and reliability (Driss et al., 2024). On MNIST, CIFAR-10, and Fashion-MNIST, the proposed selector attains the highest accuracy, lowest loss, lowest convergence time, and highest energy efficiency among GA-, MAB-, and DP-based baselines (Driss et al., 2024). The same work reports that a “select-then-train” policy reduces time from $2$5 s to $2$6 s, lowers energy from $2$7 J to $2$8 J, and improves energy efficiency from $2$9/J to $c$0/J on MNIST (Driss et al., 2024). In data-centric Green FL, client selection is driven by a score that combines node emissions and data-quality indicators, and vertical reduction—training with fewer, better clients rather than uniformly shrinking all local datasets—consistently lowers emissions while maintaining or improving accuracy (Sabella et al., 23 Jul 2025).

5. Wireless, edge, and decentralized green architectures

In wireless and edge settings, Green FL is often framed as a joint optimization of communication and computation resources under explicit physical-layer constraints. A Cloud-RAN formulation models $c$1 single-antenna devices, $c$2 RRHs, and a central server with limited fronthaul capacity, and minimizes total energy by jointly optimizing QNN precision, user transmit powers, and fronthaul quantization levels while satisfying a target convergence accuracy (Wang et al., 2023). The paper proves a convergence-rate bound for quantized FL and uses it to translate an accuracy requirement into a required number of rounds, thereby directly coupling learning quality to energy expenditure (Wang et al., 2023). Simulation results show that the joint design reduces total energy significantly relative to equal-power, equal-fronthaul, and fixed-precision baselines, and that the optimal precision lies around 15–17 bits in the reported setups (Wang et al., 2023).

Mobile-edge generative FL extends this perspective to diffusion models. The on-demand quantized federated diffusion framework formulates an energy minimization problem over CPU frequency, transmit power, and quantization level under latency and quantization-error constraints (Lai et al., 2024). Its experiments fine-tune a DDPM on GTSRB using 10 edge devices and report lower system energy consumption than standard FedAvg and fixed 6-, 7-, and 8-bit baselines while maintaining reasonable FID (Lai et al., 2024). The associated binary-search optimization converges in about 20 iterations for various quantization requirements (Lai et al., 2024). A plausible implication is that generative FL can be made green only by jointly treating model compression and resource scheduling, rather than by naively federating large diffusion models.

Decentralized and self-organizing designs provide a different route to green operation. SParSeFuL replaces the central server with proximity-based federations built through aggregate computing and combines this with neural-network sparsification and quantization (Domini et al., 10 Jul 2025). On EMNIST with an MLP, post-training pruning at sparsification ratio $c$3 reduces relative energy consumption to $c$4 of dense baseline and cuts inference time by nearly two-thirds without noticeable accuracy loss; at $c$5 and $c$6, relative energy consumption is reported as $c$7 (Domini et al., 10 Jul 2025). The paper argues that excessive sparsity, such as $c$8, can impair federation formation because compressed models no longer reliably encode distributional similarity (Domini et al., 10 Jul 2025).

One-shot and closed-form federated paradigms push the communication-reduction principle further. In one-shot federated generative learning for IoT, clients upload prompts only once, the server synthesizes a surrogate dataset with generative models, and the global classifier is trained centrally (Qi et al., 2024). Compared with FedAvg, communication overhead is about one-fifth and running time about one-half, while accuracy is higher on Fashion-MNIST, CIFAR-10, and CIFAR-100 in both IID and non-IID settings (Qi et al., 2024). A separate line for one-layer neural networks derives a closed-form federated solution in a single training round and shows that, for large datasets, federated training can consume less energy than centralized training of the same model while preserving identical accuracy under IID and non-IID partitions (Fontenla-Romero et al., 2023).

6. Application domains and empirical patterns

Green FL has been studied across generative AI, mobile and IoT systems, smart industry, medical imaging, and language modeling. In mobile keyboard language modeling at production scale, direct carbon accounting on millions of phones shows that a single FL training run can emit $c$9–$t$0 kg CO$t$1e, and that different hyperparameter settings with similar accuracy can differ in carbon impact by up to $t$2 (Yousefpour et al., 2023). The same study finds that asynchronous FL reaches target perplexity faster but emits more carbon than synchronous FL at the same concurrency, and that carbon grows linearly with concurrency times rounds-to-target or time-to-target, depending on the protocol (Yousefpour et al., 2023). This has become a foundational empirical rule in Green FL.

In medical imaging, adaptive encoder freezing for cross-silo FL on MRI-to-CT conversion provides an explicitly Green-AI-oriented mechanism for reducing local computation while preserving clinical performance (Raggio et al., 25 Nov 2025). Across five encoder–decoder architectures and 25 federated rounds, the approach reduces training time by about $t$3–$t$4, energy by about $t$5–$t$6, and CO$t$7eq by about $t$8–$t$9, while three architectures show no statistically significant MAE differences and two show statistically significant but clinically negligible MAE improvements (Raggio et al., 25 Nov 2025). The method freezes the encoder once the relative change in encoder weights remains below a threshold $g_c^{(t)} = E_c \cdot \mathrm{CI}_c^{(t)},$0 for $g_c^{(t)} = E_c \cdot \mathrm{CI}_c^{(t)},$1 rounds (Raggio et al., 25 Nov 2025). This illustrates a computation-centric Green FL design that does not rely on communication compression.

Data-centric Green FL addresses the environmental role of data itself. For time-series classification, a recommendation system first predicts the data volume required to meet a target accuracy and then ranks clients by a score that combines emissions and data quality (Sabella et al., 23 Jul 2025). Across three evaluation configurations, Node Selection reduces emissions by about $g_c^{(t)} = E_c \cdot \mathrm{CI}_c^{(t)},$2 on average relative to baseline, Minimal Smart Reduction by about $g_c^{(t)} = E_c \cdot \mathrm{CI}_c^{(t)},$3, and Smart Reduction by about $g_c^{(t)} = E_c \cdot \mathrm{CI}_c^{(t)},$4, while all three improve accuracy and Smart Reduction achieves the highest accuracy gains (Sabella et al., 23 Jul 2025). This suggests that client and sample selection can function as a green lever even without changing the model architecture or communication protocol.

The applications surveyed in Green IoT extend these themes to IIoT, smart cities, healthcare, UAVs, and transportation, often focusing on energy-aware client selection, model compression, and edge resource management (Thakur et al., 2024). A consistent empirical pattern across domains is that reducing rounds, transmitted bits, or per-round compute is beneficial only if the induced loss in convergence speed does not increase total training duration enough to negate the savings (Thakur et al., 2024).

7. Open problems and research directions

Several issues recur across the literature. First, measurement remains inconsistent. Standardized phase-aware carbon accounting, per-site and per-round reporting, and explicit documentation of hardware, grid, and network assumptions are still not universal, even though recent methodological work argues they are necessary for meaningful comparison of “green” claims (Tapp et al., 30 Jan 2026). Second, many algorithms optimize energy or communication without explicitly modeling carbon intensity, while carbon-aware algorithms often rely on forecast quality, static client availability assumptions, or simplified grid models (Arputharaj et al., 10 Sep 2025).

Third, the relationship between compression, non-IID data, and convergence remains delicate. Compression pipelines can be highly effective in communication-constrained environments, but aggressive pruning or quantization degrades accuracy more severely in heterogeneous settings (Colybes et al., 13 Apr 2026). Similarly, sparse and decentralized methods may require careful control of sparsity patterns or aggregation rules to avoid architectural mismatch across devices (Domini et al., 10 Jul 2025). A plausible implication is that future Green FL systems will need to co-design compression with statistical heterogeneity handling rather than treating them as independent modules.

Fourth, fairness is unresolved. Carbon-aware or energy-aware selection can bias training toward clients in greener regions, with cleaner data, or with better hardware. One response is $g_c^{(t)} = E_c \cdot \mathrm{CI}_c^{(t)},$5-fair carbon allocation and unbiased aggregation (Arputharaj et al., 10 Sep 2025); another is explicit fairness constraints in client selection (Driss et al., 2024). Yet the broader question of how to balance emissions minimization against equitable participation remains open.

Finally, broader integration is still emerging. Survey work repeatedly identifies opportunities in carbon-aware scheduling, adaptive participation, model simplification, mixed precision, energy harvesting, hardware-aware FL, and realistic large-scale IoT deployments (Thakur et al., 2024). The standardization literature adds that embodied emissions, dynamic network energy models, and finer-grained grid carbon intensity should be incorporated into future accounting frameworks (Tapp et al., 30 Jan 2026). Taken together, these strands indicate that Green Federated Learning is evolving from isolated energy-saving heuristics toward a comprehensive discipline of carbon-aware federated systems design, where optimization, measurement, hardware, and fairness are treated as inseparable.