Strategies¶
nFL — No Federated Learning¶
Standalone baselines and non-FL pre-training methods. No model communication; each client trains independently or is evaluated in a centralized setting.
| Name | Venue | Year | Description | Paper | URL |
|---|---|---|---|---|---|
| LocalOnly | Per-client local training only, no communication | ||||
| LocalOLS | Per-client local OLS regression, no communication | ||||
| Centralized | Oracle upper bound — all data on one server | ||||
| SimTS | ICASSP | 2024 | Contrastive pretraining with instance-pair selection for TSF | SimTS: Rethinking Contrastive Representation Learning for Time Series Forecasting | Arxiv - IEEE - GitHub |
| InfoTS | ICLR | 2023 | Meta-contrastive with learnable augmentation selection for TSF | InfoTS: Information-Aware Time Series Meta-Contrastive Learning | Arxiv - REF |
| SL | NeurIPS | 2025 | Dual-mask selective loss (uncertainty + anomaly) for deep TSF | Selective Learning for Deep Time Series Forecasting | OpenReview - GitHub |
tFL — Traditional Federated Learning¶
Central-server FL where a server aggregates client updates each round and broadcasts a single global model. All clients converge toward one shared solution.
| Name | Venue | Year | Description | Paper | URL |
|---|---|---|---|---|---|
| FedAvg | AISTATS | 2017 | Weighted average of client model updates | Communication-Efficient Learning of Deep Networks from Decentralized Data | Arxiv |
| FedAvgM | FedAvg with server-side Polyak momentum on model updates | Measuring the Effects of Non-Identical Data Distribution for Federated Visual Classification | Arxiv - REF | ||
| SCAFFOLD | ICML | 2020 | Variance-reduced FL via per-client control variates | SCAFFOLD: Stochastic Controlled Averaging for Federated Learning | Arxiv - REF |
| FedNova | NeurIPS | 2020 | Normalizes local updates by effective steps to fix objective inconsistency | Tackling the Objective Inconsistency Problem in Heterogeneous Federated Optimization | Arxiv - REF |
| FedSPA | IJCAI | 2021 | Client rand-k sparsification + Gaussian DP noise per local step; server Adam-like adaptive update (u/v moments over avg delta) | Federated Learning with Sparsification-Amplified Privacy and Adaptive Optimization | Arxiv |
| FedAdam | ICLR | 2021 | Server-side Adam momentum for global model update | Adaptive Federated Optimization | Arxiv - REF |
| FedYogi | ICLR | 2021 | Server-side Yogi optimizer (sign-scaled momentum) for global update | Adaptive Federated Optimization | Arxiv - REF |
| FedPAQ | AISTATS | 2020 | Periodic averaging with stochastic quantization of update vectors | FedPAQ: A Communication-Efficient Federated Learning Method with Periodic Averaging and Quantization | Arxiv |
| MOON | CVPR | 2021 | Model-contrastive regularization with global and previous local models | Model-Contrastive Federated Learning | Arxiv |
| Caesar† | arXiv | 2024 | Staleness-aware top-K+1-bit model download with client recovery; importance-ranked gradient upload sparsification | Caesar: A Low-deviation Compression Approach for Efficient Federated Learning | Arxiv |
| FedADMM | ICDE | 2022 | ADMM-based FL robust to system heterogeneity and stragglers | FedADMM: A Robust Federated Deep Learning Framework with Adaptability to System Heterogeneity | IEEE - REF |
| FedLAW | ICML | 2023 | Learned aggregation weights via auxiliary network on validation data | Revisiting Weighted Aggregation in Federated Learning with Neural Networks | Arxiv - GitHub |
| Elastic | CVPR | 2023 | Sensitivity-weighted elastic aggregation to preserve task-critical params | Elastic Aggregation for Federated Optimization | OpenAccess - REF |
| FedRolex | NeurIPS | 2022 | Heterogeneous FL via rolling sub-model extraction and cyclic training | FedRolex: Model-Heterogeneous Federated Learning with Rolling Sub-Model Extraction | Arxiv - GitHub |
| FedCross | ICDE | 2024 | Cross-aggregation of multiple global model candidates per round | FedCross: Towards Accurate Federated Learning via Multi-Model Cross-Aggregation | IEEE - Arxiv - REF |
| FedRCL† | CVPR | 2024 | Relaxed supervised contrastive loss with per-pair divergence penalty | Relaxed Contrastive Learning for Federated Learning | OpenAccess - Arxiv - GitHub |
| FedAWA | CVPR | 2025 | Server-optimized per-client aggregation weights via client vectors | FedAWA: Adaptive Optimization of Aggregation Weights in Federated Learning Using Client Vectors | CVPR - Arxiv |
| FedTrend | Science China Information Sciences | 2026 | Trend-aware aggregation for heterogeneous federated TSF | Tackling Data Heterogeneity in Federated Time Series Forecasting | PUB - Arxiv |
| FedRidge | arXiv | 2026 | One-shot federated ridge regression via sufficient statistic aggregation | One-Shot Federated Ridge Regression: Exact Recovery via Sufficient Statistic Aggregation | Arxiv |
| DLSA | JCGS | 2021 | Federated weighted least-squares via precision-matrix aggregation | Least-Square Approximation for a Distributed System | PUB - Arxiv |
| QATFL | IEEE TMLCN | 2026 | Fake-quantization + STE during local training reduces quantization distortion vs. post-training quantization; QSGD-quantized delta upload | Communication Efficient Federated Learning With Quantization-Aware Training Design | PUB |
| DeComFL‡ | arXiv | 2024 | Zeroth-order optimization: clients upload only scalar loss-difference gradients (dimension-free uplink); server replays SPSA update from shared perturbation seeds | Achieving Dimension-Free Communication in Federated Learning via Zeroth-Order Optimization | Arxiv - GitHub |
| FedLUAR | arXiv | 2025 | Recycles previous-round updates for a subset of layers, chosen via inverse-magnitude probability sampling, to cut uplink | Layer-wise Update Aggregation with Recycling for Communication-Efficient Federated Learning | Arxiv - GitHub |
sFL — Security-Aware Federated Learning¶
Central-server FL with a pluggable threat injection seam between client training and aggregation. Strategies that inherit from sFL get the seam for free and can be evaluated under any registered attack (Byzantine model poisoning, sign-flip, etc.) by flipping --attack at run time without changing strategy code.
Strategies that inherit from sFL implement a specific defense; more families can be added without touching the injection seam.
Use --attack <name> and --malicious_frac <f> to switch between benign and adversarial evaluation — see Usage § Adversarial Eval.
| Name | Venue | Year | Description | Paper | URL |
|---|---|---|---|---|---|
| Krum | NeurIPS | 2017 | Byzantine-robust aggregation via nearest-neighbor cluster selection | Machine Learning with Adversaries: Byzantine Tolerant Gradient Descent | PUB - REF |
| FedMedian | ICML | 2018 | Byzantine-robust coordinate-wise median aggregation | Byzantine-Robust Distributed Learning: Towards Optimal Statistical Rates | Arxiv - REF |
| FedTrimmedAvg | ICML | 2018 | Byzantine-robust trimmed mean aggregation | Byzantine-Robust Distributed Learning: Towards Optimal Statistical Rates | Arxiv - REF |
pFL — Personalized Federated Learning¶
Each client maintains its own model or model component alongside the global model. Aggregation is designed to produce client-specific outputs rather than a single shared solution.
| Name | Venue | Year | Description | Paper | URL |
|---|---|---|---|---|---|
| FedProx | MLsys | 2020 | Proximal term regularizes local objective toward global model | Federated Optimization in Heterogeneous Networks | Arxiv |
| Ditto | ICML | 2021 | Jointly trains global and personalized models with proximity constraint | Ditto: Fair and Robust Federated Learning Through Personalization | Arxiv - REF |
| pFedMe | NeurIPS | 2020 | Personalized model found via Moreau envelope of global objective | Personalized Federated Learning with Moreau Envelopes | Arxiv - REF |
| APFL | arXiv | 2020 | Convex mixture of local and global models with adaptive α | Adaptive Personalized Federated Learning | Arxiv - REF |
| PerAvg | NeurIPS | 2020 | MAML-style meta-learning for fast one-step personalization | Personalized Federated Learning with Theoretical Guarantees: A Model-Agnostic Meta-Learning Approach | Arxiv - REF |
| FedAMP | AAAI | 2021 | Attention-mechanism aggregation for personalized cross-silo FL | Personalized Cross-Silo Federated Learning on Non-IID Data | Arxiv - REF |
| FedBN | ICLR | 2021 | Keeps batch norm layers local to handle feature-distribution shift | FedBN: Federated Learning on Non-IID Features via Local Batch Normalization | Arxiv - REF |
| FML† | NeurIPS | 2020 | Mutual KL distillation between global and local model predictions | Federated Mutual Learning | Arxiv - REF |
| FedDyn | ICLR | 2021 | Dynamic per-client regularization toward a moving global anchor | Federated Learning Based on Dynamic Regularization | Arxiv - REF |
| CFL | arXiv | 2019 | Agglomerative clustering of clients; each cluster trains independently | Clustered Federated Learning: Model-Agnostic Distributed Multi-Task Optimization under Privacy Constraints | Arxiv - REF |
| LGFedAvg | arXiv | 2020 | Splits model into shared global layers and personal local layers | Think Locally, Act Globally: Federated Learning with Local and Global Representations | Arxiv - REF |
| pFedHN | ICML | 2021 | Hypernetwork generates personalized model weights per client | Personalized Federated Learning using Hypernetworks | Arxiv - GitHub |
| pFedLA | NeurIPS | 2022 | Layer-wise aggregation weights learned via hypernetwork | Layer-Wise Personalized Federated Learning via Hypernetworks | Arxiv - GitHub |
| AirMetapFL | arXiv | 2025 | Over-the-air meta-learning with convergence–generalization tradeoff | Pre-Training and Personalized Fine-Tuning via Over-the-Air Federated Meta-Learning: Convergence-Generalization Trade-Offs | Arxiv |
| FedFew | arXiv | 2025 | K server models with STCH-Set soft assignment; near-optimal pFL via few-for-many | Few-for-Many Personalized Federated Learning | Arxiv |
| FedALA | AAAI | 2023 | Element-wise adaptive local aggregation via gradient alignment | FedALA: Adaptive Local Aggregation for Personalized Federated Learning | Arxiv |
| FedCAC | ICCV | 2023 | Sparsity-based selective collaboration with cautious aggregation | Bold but Cautious: Unlocking the Potential of Personalized Federated Learning through Cautiously Aggressive Collaboration | OpenAccess - Arxiv |
| FedIT | ICASSP | 2024 | Federated instruction tuning with LoRA for large language models | Towards Building the Federated GPT: Federated Instruction Tuning | Arxiv |
| FFA_LoRA | ICLR | 2024 | Freezes LoRA-A, aggregates only LoRA-B for privacy-preserving fine-tuning | Improving LoRA in Privacy-preserving Federated Learning | Arxiv |
| FedSelect | CVPR | 2024 | Sparse parameter selection masks for efficient personalized fine-tuning | FedSelect: Personalized Federated Learning with Customized Selection of Parameters for Fine-Tuning | Arxiv - GitHub |
| FedSA_LoRA | ICLR | 2025 | Selectively aggregates task-relevant LoRA components | Selective Aggregation for Low-Rank Adaptation in Federated Learning | Arxiv |
| LoRA_FAIR | ICCV | 2025 | Refines LoRA aggregation weights and re-initializes A matrices | LoRA-FAIR: Federated LoRA Fine-Tuning with Aggregation and Initialization Refinement | Arxiv - GitHub |
| FlexLoRA | NeurIPS | 2024 | Heterogeneous LoRA ranks via SVD-based redistribution at server | Federated Fine-tuning of Large Language Models under Heterogeneous Tasks and Client Resources | Arxiv |
hFL — Heterogeneous Federated Learning¶
Clients have architecturally different models. Knowledge is transferred via a shared public dataset or distillation rather than direct parameter averaging.
| Name | Venue | Year | Description | Paper | URL |
|---|---|---|---|---|---|
| FedMD | NeurIPS-W | 2019 | Knowledge distillation via public dataset for heterogeneous model architectures | FedMD: Heterogenous Federated Learning via Model Distillation | Arxiv |
| FedDF* | NeurIPS | 2020 | Ensemble distillation on public data refines server model post-aggregation | Ensemble Distillation for Robust Model Fusion in Federated Learning | Arxiv |
dFL — Decentralized Federated Learning¶
No central server; clients communicate directly with neighbors in a peer-to-peer topology. Each node aggregates only from its local neighborhood.
| Name | Venue | Year | Description | Paper | URL |
|---|---|---|---|---|---|
| DFedAvg** | AISTATS | 2017 | Decentralized FedAvg on peer-to-peer topology | Communication-Efficient Learning of Deep Networks from Decentralized Data | Arxiv |
| DFedProx** | MLsys | 2020 | Decentralized FedProx with proximal regularization on P2P topology | Federated Optimization in Heterogeneous Networks | Arxiv |
| DFedSAM | ICML | 2023 | SAM-based local training with model-consistency regularization for decentralized FL | Improving the Model Consistency of Decentralized Federated Learning | Arxiv - Slide |
| DFedAWA** | CVPR | 2025 | Decentralized FedAWA with local adaptive aggregation weights | FedAWA: Adaptive Optimization of Aggregation Weights in Federated Learning Using Client Vectors | CVPR - Arxiv |
| DFedHPO | Internet of Things | 2025 | Pre-FL distributed hyperparameter optimization with neighbor consensus | Decentralized Federated Learning with Hyperparameter Optimization | PUB |
aFL — Asynchronous Federated Learning¶
No synchronization barrier; clients run continuously and results are aggregated as they arrive. A server-side buffer collects K results before each aggregation step, decoupling client speed from global update frequency.
| Name | Venue | Year | Description | Paper | URL |
|---|---|---|---|---|---|
| FedPSA | arXiv | 2025 | Sensitivity-sketch behavioral staleness + thermometer-controlled softmax aggregation | FedPSA: Modeling Behavioral Staleness in Asynchronous Federated Learning | Arxiv |
spFL - Sparse/Pruning Federated Learning¶
Dynamic sparse training in FL: clients maintain a binary mask that zeroes selected weights each round. A cosine-decay schedule (delta_T adjustment rounds up to T_end) controls how aggressively the mask evolves.
| Name | Venue | Year | Description | Paper | URL |
|---|---|---|---|---|---|
| PruneFL | TNNLS | 2022 | Server-guided mask: FedAvg of per-client squared gradients → prune smallest active + grow largest-grad² inactive | Model Pruning Enables Efficient Federated Learning on Edge Devices | Arxiv - REF |
| FedDST | AAAI | 2022 | Client-local mask update (sparse_update_step after A_epochs); server reconciles via OR-union + magnitude re-prune | Federated Dynamic Sparse Training: Computing Less, Communicating Less, Yet Learning Better | Arxiv - REF |
| FedTiny | ICDCS | 2023 | Clients send post-training gradients; server FedAvg of gradients → server-side prune + grow (TinyClean variant) | Distributed Pruning Towards Tiny Neural Networks in Federated Learning | Arxiv - REF |
| FedMef | CVPR | 2024 | FedTiny + BAE regularization (L2 on low-magnitude active weights) + optional topk-grad server filter | FedMef: Towards Memory-efficient Federated Dynamic Pruning | Arxiv - REF |
| FedSGC | ICLRW | 2024 | Direction-coherent local mask update: prune conflict-direction active, grow agreement-direction inactive; server OR-union + reprune + direction map | Gradient-Congruity Guided Federated Sparse Training | Arxiv - REF |
| FedRTS | NeurIPS | 2025 | Per-weight Beta(α,β) distributions; clients vote on core active (top-κ mag) and grow candidates (top-k grad); server Thompson-samples new mask | FedRTS: Federated Robust Pruning via Combinatorial Thompson Sampling | Arxiv - REF |
* Adapted from classification to regression. Please use with caution.
** Decentralized variant converted from its tFL/pFL counterpart (e.g. FedAvg → DFedAvg).
† TSF adaptation applied: class-level assumptions replaced with time-series equivalents. Please use with caution.
‡ Architecture-driven deviation: FedProC's stateless-client design cannot support a mechanism the paper relies on. Not a bug, not a TSF adaptation — see the Implementation Notes table.
| Strategy | Type | Note |
|---|---|---|
| FedRCL† | TSF adaptation | Pseudo-labels via quantile binning replace class labels; no multi-level contrastive hooks (TSF has no intermediate features). |
| FML† | TSF adaptation | KL divergence computed over the time dimension instead of the class dimension. |
| Caesar† | TSF adaptation | No class labels → KL term dropped. Uplink = COO sparse gradient; downlink = true compressed wire size. Paper defaults (§5.1): theta_d_max=0.6, theta_u_min=0.1, theta_u_max=0.6, lambda=0.5. |
| DeComFL‡ | Implementation note | Uplink faithful (dimension-free ZO scalars, mu=0.001 matches paper). Downlink is not dimension-free like the paper: the paper's trick needs stateful clients to replay a model update from a shared random seed; FedProC's clients are stateless (respawned fresh each round) and cannot replay history, so the full model is sent downlink every round instead. q=2, zo_lr=0.01 are TSF-adapted, not paper defaults. |