PaperlayerView exact revision on arXiv

Source-linked AI summary

FedAdapt: Adaptive Offloading for IoT Devices in Federated Learning

Di Wu, Rehmat Ullah, Paul Harvey, Peter Kilpatrick, Ivor Spence, Blesson Varghese

arXiv:2107.04271v5cs.DCcs.LG

TL;DR

IoT FL must contend with sensitive data, limited device computation, stragglers, and changing network bandwidth. FedAdapt combines DNN-layer offloading with RL and clustering to adapt strategies, and reports shorter training times than classic FL while maintaining accuracy and convergence speed.

  • Problem

    FL on IoT devices must handle sensitive data alongside limited device computation, computational heterogeneity that creates stragglers, and changing network bandwidth.

    Source: §I · pp. 1–2

  • Method

    FedAdapt offloads DNN layers to servers and uses RL with clustering to select and map adaptive offloading strategies for devices.

    Source: §I · p. 2, A. Overview · p. 4

  • Results

    FedAdapt reduced straggler training time by over half and reduced training time by up to 40% versus classic FL while achieving the same accuracy and convergence speed with negligible execution-time overhead.

    Source: §VI · p. 12

  • Takeaways & Limitations

    The reported results show training-time reductions in a lab-based testbed despite stragglers and changing bandwidth, with no loss of reported accuracy or convergence speed.

    Source: §VI · p. 12

    For large numbers of devices, a single RL agent may be unsuitable; offloading also introduces communication overhead between devices and the server.

    Source: §VI · p. 12

Abstract

from arXiv · show

Applying Federated Learning (FL) on Internet-of-Things devices is necessitated by the large volumes of data they produce and growing concerns of data privacy. However, there are three challenges that need to be addressed to make FL efficient: (i) execution on devices with limited computational capabilities, (ii) accounting for stragglers due to computational heterogeneity of devices, and (iii) adaptation to the changing network bandwidths. This paper presents FedAdapt, an adaptive offloading FL framework to mitigate the aforementioned challenges. FedAdapt accelerates local training in computationally constrained devices by leveraging layer offloading of deep neural networks (DNNs) to servers. Further, FedAdapt adopts reinforcement learning based optimization and clustering to adaptively identify which layers of the DNN should be offloaded for each individual device on to a server to tackle the challenges of computational heterogeneity and changing network bandwidth. Experimental studies are carried out on a lab-based testbed and it is demonstrated that by offloading a DNN from the device to the server FedAdapt reduces the training time of a typical IoT device by over half compared to classic FL. The training time of extreme stragglers and the overall training time can be reduced by up to 57%. Furthermore, with changing network bandwidth, FedAdapt is demonstrated to reduce the training time by up to 40% when compared to classic FL, without sacrificing accuracy.

I. INTRODUCTION

FedAdapt targets the training-time costs of FL on sensitive-data-producing IoT devices, including constrained computation, stragglers, and changing bandwidth. It uses adaptive DNN-layer offloading and reports reduced training times against classic FL.

I. INTRODUCTION

  • IoT devices produce sensitive data that can be analyzed with ML for personalized services, motivating privacy-preserving ML techniques.FL trains on device data without sending raw data to the server.Source: §I · p. 1
  • In synchronous FL, stragglers delay other devices because the server waits for all devices, while changing network bandwidth can also affect training time.The introduction identifies adaptive, context-aware strategies as needed for changing operational conditions.Source: §I · p. 1
  • FedAdapt offloads DNN layers to a server, uses RL to identify layers for each device, and uses clustering to generate offloading strategies rapidly.The framework is presented as addressing acceleration, computational heterogeneity, and changing network bandwidth.Source: §I · p. 2
  • 30% lower total VGG-5 training time with the same accuracy and convergence speed, and 40% lower total VGG-8 training time without further RL training, versus classic FL.These results were reported on a testbed with five devices and a server.Source: §I · p. 2
  • 40% lower training time per FL round using VGG-5 was reported for the experimental testbed with variable layer-wise training using RL-based offloading.The paper presents dynamic offloading strategies as a contribution for addressing computational heterogeneity.Source: §I · p. 2

II. BACKGROUND AND RELATED WORK

The section reviews FL, split learning, splitfed learning, and computation offloading, then contrasts their properties with FedAdapt. It frames FedAdapt's distinction as adaptive, automated partitioning for heterogeneous devices and changing bandwidth.

II. BACKGROUND AND RELATED WORK

  • FL distributes a global model, trains it independently on device data, and aggregates updated models on a server, but heterogeneous devices can make stragglers a bottleneck.The round comprises server distribution, local training, and server aggregation.Source: §II · p. 2, Figure 1 · p. 2
  • Split learning partitions a DNN between device and server to reduce device computation, but trains multiple devices sequentially, making it inefficient at large scale.Only one device connects to the server at a time in the described multi-device approach.Source: §II · pp. 2–3
  • Table I compares FL, SL, SFL, and FedAdapt; the text distinguishes FedAdapt by automated RL-based partitions for heterogeneous devices and changing bandwidth without prior device knowledge.The section also notes that much prior FL, SL, and SFL work uses simulation rather than physical testbeds.Source: Table I · p. 3, §II · p. 3
  • Existing splitfed work combines parallel FL training with device-side acceleration, but does not consider optimal partitioning or may rely on manual configuration and static partitions.Static partitions can become sub-optimal when operational conditions change.Source: §II · p. 3
  • Edge-computing offloading shifts computationally expensive application components from constrained devices to nearby servers, while adaptable offloading for FL training remains an area needing development.The cited discussion distinguishes FL training from prior offloading work on inference.Source: §II · p. 3

III. THE FEDADAPT FRAMEWORK

The section introduces the FedAdapt framework and its underpinning techniques, then formulates distributing an FL DNN across device and server for performance efficiency.

III. THE FEDADAPT FRAMEWORK

  • The section introduces FedAdapt's framework and techniques before formulating DNN distribution across device and server for performance efficiency.The supplied passage describes the section's scope rather than the techniques themselves.Source: §III · p. 3

A. Overview

FedAdapt combines layer offloading with RL and clustering to select strategies for device groups and map them to individual devices. Its optimized RL strategy accounts for changing network bandwidth.

A. Overview

  • FedAdapt's four modules are the Pre-processor, Clustering Module, Trained RL Agent, and Post-processor.The Pre-processor gathers device observations, including computational capabilities and network bandwidth, after an FL round.Source: A. Overview · p. 3
  • Clustering groups devices with similar training times and accounts for network bandwidth; RL selects a strategy per group, which the Post-processor maps to devices.Devices in a group execute the same offloading strategy for the FL round.Source: A. Overview · pp. 3–4
  • Offloading moves selected DNN layers from constrained devices to the server, and optimized RL accounts for changing device-server bandwidth when generating strategies.The RL agent identifies the Offloading Point for each device before an FL round.Source: A. Overview · pp. 3–4
  • DNN layers after the Offloading Point are moved to the server, while earlier layers remain on the device; activation data, labels, and gradients are exchanged during training.The paper notes communication overheads but says computational offloading reduces overall FL training time.Source: A. Overview · p. 3

B. Problem Model

FedAdapt models each device’s training time as a function of its workload, computation split, and communication, then optimizes offloading for changing bandwidth. Its objective is average device training time per round, while network changes within a round are not accounted for.

  • Bandwidth is observed from the previous FL round to generate an offloading strategy, but changes during a round are not modeled.FedAdapt assumes bandwidth can change between rounds.Source: B. Problem Model · p. 4
  • The model accounts for device workload and training speeds, with communication time included in each device’s training time.The expression includes time for feature-map communication during training.Source: B. Problem Model · p. 4
  • The offloading strategy sets the proportion of each device’s workload that runs locally; the remaining workload is offloaded to the server.The strategy also determines the size of feature maps transferred between device and server.Source: B. Problem Model · p. 4
  • FedAdapt minimizes average training time across K devices per round, rather than only optimizing the synchronous round time set by the slowest device.The per-round offloading strategy is optimized under changing operational conditions.Source: B. Problem Model · pp. 4–5

IV. TRAINING THE REINFORCEMENT LEARNING AGENT FOR FEDADAPT

FedAdapt trains an RL agent to choose group-level DNN offloading strategies, using clustering to handle device heterogeneity and changing network bandwidth. Its offline training produces an actor that can be reused for similar FL tasks.

RL training methodology

  • A reused trained RL agent achieves 57% training-time reduction per round without training for another specific model.The agent is trained offline before FL tasks, and the paper says it can be reused for similar tasks.Source: §IV · p. 5

Clustering technique with RL

  • FedAdapt clusters devices by training time and network bandwidth, reducing the agent's action space from individual devices to groups.This addresses changing device counts and the action-space growth that makes per-device RL difficult; each group is represented by its slowest-training device.Source: §IV · p. 5

Optimizing for network bandwidth

  • When a device's bandwidth falls below a threshold, FedAdapt assigns it to a limited-bandwidth group and trains the agent under that group's network conditions.Devices are checked at the beginning of each FL round.Source: §IV · p. 6

State and action

  • The agent outputs a group-level value from zero to one, which is mapped to a device's workload share and then to the nearest available offloading point.The target workload is set using FLOPs; VGG-5's five layers yield five offloading actions.Source: §IV · p. 6

Choice of algorithm

  • FedAdapt uses PPO because its offloading actions are continuous, whereas DQN would need to evaluate an unenumerable action space.PPO's off-policy use of prior trajectory data also improves training efficiency when exploration is time-consuming.Source: §IV · pp. 6–7

RL training methodology

  • The agent is trained offline with an actor and critic; truncated FL rounds use fewer batches, after which normal rounds train the FL model.The critic assists actor training, while the trained actor alone outputs offloading actions; the first round uses classic FL to generate the initial state.Source: §IV · pp. 5, 7

V. EXPERIMENTAL STUDIES

The experimental section evaluates FedAdapt against classic FL across three research questions: layer offloading, computational heterogeneity, and changing network bandwidth.

V. EXPERIMENTAL STUDIES

  • FedAdapt's experimental evaluation compares its benefits with classic FL.The section is organized around three research questions: accelerating training through layer offloading, reducing computational heterogeneity's impact, and adapting to changing bandwidth.Source: §V · p. 7
  • The layer-offloading experiments address whether offloading accelerates FL training.These results are presented in Section V-A.Source: §V · p. 7
  • The RL experiments examine computational heterogeneity and changing network bandwidth as separate evaluation questions.They are presented in Sections V-B and V-C, respectively.Source: §V · p. 7

A. Layer Offloading in FL

Layer offloading reduces training time relative to device-native classic FL, but the best offloading choice depends on both model and network bandwidth. The study tests VGG-5 and VGG-8 at multiple bandwidths and finds device-native execution best for VGG-5 at 10 Mbps, while offloading still helps VGG-8.

A. Layer Offloading in FL

  • The evaluation tests all offloading points for VGG-5 and VGG-8 across network bandwidths, using CIFAR-10 and a batch size of 100.The last OP corresponds to device-native training, as in classic FL.Source: A. Layer Offloading in FL · p. 7
  • On Wi-Fi, VGG-5 and VGG-8 reach 2.38s and 4.75s per iteration, versus 4.36s and 10.61s for classic FL.The best offloading point is OP1 for both models.Source: Table V · p. 7, A. Layer Offloading in FL · p. 7, Table VI · p. 8
  • The best OP is the pooling layer, with layers beyond it offloaded to the server.Pooling has relatively low computational workload and produces a low volume of activation feature maps to transfer.Source: A. Layer Offloading in FL · pp. 7–8
  • At 25Mbps, training time falls 19% for VGG-5 and 43% for VGG-8; at 10Mbps, VGG-5 favors OP4 while VGG-8 still gains 17%.With Wi-Fi, reductions exceed 45% for VGG-5 and 55% for VGG-8.Source: A. Layer Offloading in FL · p. 8

B. RL Optimization for Heterogeneity

FedAdapt uses RL to choose offloading points for groups of devices with different training speeds. In a five-device VGG-5 test, it reduces total training time by 40% versus classic FL, with a 50% per-round reduction for the straggler Pi42.

B. RL Optimization for Heterogeneity

  • FedAdapt clusters devices by first-round training time and assigns each group a shared RL-generated offloading action.K-means forms three groups; Jetson and Pi42 are each placed in separate groups, while the other Raspberry Pis share a group.Source: B. RL Optimization for Heterogeneity · p. 9, Table VII · p. 9
  • The optimal actions offload layers after OP1 for the Raspberry Pis, while Jetson uses device-native execution.The RL agent selects the OP closest to its generated action, using the tested OP boundaries.Source: B. RL Optimization for Heterogeneity · p. 9
  • The RL agent's mean actions for all three groups align with their optimal offloading actions after 80 rounds.The agent is trained first, then its actor network is deployed to guide offloading during FL.Source: B. RL Optimization for Heterogeneity · p. 9
  • FedAdapt saves 40% of total training time per round, with a 50% per-round reduction for straggler Pi42.The experiment keeps network bandwidth unchanged during FL rounds.Source: B. RL Optimization for Heterogeneity · p. 9, Figure 6 · p. 10

C. Adapting to Changing Network Bandwidth

FedAdapt accounts for bandwidth changes by grouping low-bandwidth devices separately and adapting offloading actions. In the reported VGG-5 experiment, the RL agent takes longer to converge on the low-bandwidth group than on the other groups.

C. Adapting to Changing Network Bandwidth

  • FedAdapt adds a low-bandwidth group and dynamically regroups devices after each round using both bandwidth and training time per iteration.Pi32's upload bandwidth is set to 10Mbps while the other devices use 75Mbps Wi-Fi.Source: C. Adapting to Changing Network… · p. 10, Table IX · p. 10
  • The RL agent's mean actions for G1 and G2 become optimal after 20 and 60 rounds, while G3's becomes optimal after 240 rounds.The reported results average five independent runs with different random seeds.Source: C. Adapting to Changing Network… · p. 10

D. Comparing FedAdapt and Classic FL

FedAdapt reduces VGG-5 training time relative to classic FL while maintaining similar accuracy and convergence speed. When bandwidth changes, it adapts its offloading strategy, and an RL agent trained for VGG-5 also reduces VGG-8 training time without retraining.

D. Comparing FedAdapt and Classic FL

  • The bandwidth-varying test runs VGG-5 for 100 rounds, lowering one device's bandwidth to 10Mbps in each of five successive time slots.The FedAdapt agent trained in the prior experiment generates the offloading action for each device; classic FL uses no offloading.Source: D. Comparing FedAdapt and Classic FL · p. 10
  • Across 100 bandwidth-varying rounds, FedAdapt reduces VGG-5 training time nearly 30% versus classic FL.The first 50 rounds show a 40% average reduction; after bandwidth changes invalidate actions, FedAdapt reassigns devices into G3 in the next round.Source: Figure 8 · p. 11, D. Comparing FedAdapt and Classic FL · p. 11
  • FedAdapt achieves the same convergence speed and final accuracy as classic FL, with an average overhead of 1.6s, or 0.5% of one training round.The reported accuracy curves are described as similar.Source: Figure 9 · p. 11, D. Comparing FedAdapt and Classic FL · p. 11
  • With the VGG-5-trained agent reused for VGG-8, training time falls 57% for straggler Pi42, while overall training time falls nearly 40% under changing bandwidth.The reuse requires no custom retraining; after round 70, however, the agent selects suboptimal device-native strategies for low-bandwidth VGG-8 devices.Source: Figure 10 · p. 11, Figure 11 · p. 11, D. Comparing FedAdapt and Classic FL · p. 11

VI. CONCLUSIONS

FedAdapt addresses the training-performance effects of limited IoT-device capacity, device heterogeneity, and varying network bandwidth. It reduces training time relative to classic FL, while introducing communication overhead that remains a focus for future work.

  • FedAdapt targets limited device capacity, computational heterogeneity, and varying device-server bandwidth through a framework incorporating three techniques.Source: §VI · pp. 11–12
  • Up to 40% lower training time than classic FL under stragglers and changing bandwidth came with the same accuracy and convergence speed and negligible execution-time overhead.Source: §VI · p. 12
  • Offloading-based training introduces additional communication overhead between devices and the server; quantization is proposed as a possible way to reduce communication cost.Source: §VI · p. 12
Loading 2107.04271v5…