Research map / Federated learning and privacy / Personalized and heterogeneous FL

PerFedRLNAS: One-for-All Personalized Federated Neural Architecture Search

AAAI 2024 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Personalization with global-local trade-off · 11 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes PerFedRLNAS, a one-stage personalized federated neural architecture search that uses per-client policy-gradient RL over a supernet to pick each client's sub-model, trading off accuracy and round time.

Problem

Personalized federated learning relies on manually designed personalized vs. shared parts of models, which can underperform and does not adapt across heterogeneous data, system resources, and memory budgets; prior federated NAS searches a global architecture or is inefficient.

Approach

Treat each communication round as an MDP step; each client has a softmax architecture-parameter policy on the server, updated by REINFORCE with a reward of local accuracy minus round-time (or memory) penalty, so the server samples personalized sub-models from an arbitrary supernet while clients train ordinary-sized models.

Key results

• On CIFAR10 (Dα=0.3) with ViT/NASViT, PerFedRLNAS reached 85.02±4.11% vs. 72.22±6.70% for FedAvg and 80.27±8.39% for FedTP, with lower latency than FedAvg (28.14 vs 34.95 hours).
• On DARTS search space (Table 3), PerFedRLNAS achieved 80.80±5.48% on CIFAR10 and 62.72±4.17% on CIFAR100 with latency 5.76 and 3.72 hours, outperforming FedNAS, FedRLNAS, FedorAS and SPIDER.
• With a memory-budget reward, final CIFAR10 accuracy was 82.87±4.87% on NASViT supernet, with each client receiving architectures fitting its memory.

Measured effects reported in the paper

Surprising findings

• Manually designed personalization policies (e.g., FedRep, FedBABU) brought little or no improvement over FedAvg with CNN models on CIFAR100.
• Setting the time-penalty weight lambda_time to 0 raised accuracy slightly (83.33% vs. lower) but increased latency to 10.65 hours, showing an accuracy-efficiency tradeoff.

Limitations

• Only image classification on CIFAR10/CIFAR100 with Dirichlet non-i.i.d. partitions is evaluated.
• Efficiency gains rely on the chosen reward weighting lambda_time; accuracy and latency trade off against each other.
• Personalization fidelity depends on the simulated system heterogeneity and transmission rates (100Mbps), which may not reflect real deployments.

Experimental setup

Datasets: CIFAR10 CIFAR100

Benchmarks: CIFAR10 CIFAR100

Metrics: average local test accuracy standard deviation across clients elapsed wall-clock latency (hours)

Baselines: FedAvg Local Training FedRep FedBABU FedTP FedNAS FedRLNAS FedorAS SPIDER

Builds on: FedAvg FedRLNAS FedRep FedBABU DARTS

Closest papers by meaning

• Large Language Models Enhanced Personalized Graph Neural Architecture Search in Federated Learning (AAAI 2025 · similarity 0.90)
• pFedClub: Controllable Heterogeneous Model Aggregation for Personalized Federated Learning (NeurIPS 2024 · similarity 0.89)
• PerAda: Parameter-Efficient Federated Learning Personalization with Generalization Guarantees (CVPR 2024 · similarity 0.88)
• PeFLL: Personalized Federated Learning by Learning to Learn (ICLR 2024 · similarity 0.88)
• FedAS: Bridging Inconsistency in Personalized Federated Learning (CVPR 2024 · similarity 0.86)
• PFedCS: A Personalized Federated Learning Method for Enhancing Collaboration among Similar Classifiers (AAAI 2025 · similarity 0.86)