CoChat
Sign Up
Cited by 1,030IoT and Edge/Fog ComputingPrivacy-Preserving Technologies in DataCaching and Content Delivery

The DOI matches its Crossref record; 17 of the 17 references it cites were found in a registry or an index; none of the 25 sources its claims cite is checked for support yet.

In-Edge AI: Intelligentizing Mobile Edge Computing, Caching and Communication by Federated Learning

Open access
Wang, Xiaofei, Han, Yiwen, Wang, Chenyang, Zhao, Qiyang, Chen, Xu, Chen, Min·2019·IEEE Network
DOI 10.1109/mnet.2019.1800286
Open this paper in CoChat — free accountOpen PDF

Abstract

Recently, along with the rapid development of mobile communication technology, edge computing theory and techniques have been attracting more and more attention from global researchers and engineers, which can significantly bridge the capacity of cloud and requirement of devices by the network edges, and thus can accelerate content delivery and improve the quality of mobile services. In order to bring more intelligence to edge systems, compared to traditional optimization methodology, and driven by the current deep learning techniques, we propose to integrate the Deep Reinforcement Learning techniques and Federated Learning framework with mobile edge systems, for optimizing mobile edge computing, caching and communication. And thus, we design the “In-Edge AI” framework in order to intelligently utilize the collaboration among devices and edge nodes to exchange the learning parameters for a better training and inference of the models, and thus to carry out dynamic system-level optimization and application-level enhancement while reducing the unnecessary system communication load. “In-Edge AI” is evaluated and proved to have near-optimal performance but relatively low overhead of learning, while the system is cognitive and adaptive to mobile communication systems. Finally, we discuss several related challenges and opportunities for unveiling a promising upcoming future of “In-Edge AI.”

Overview

AI snapshot

Written from the paper’s full text.

TL;DR

The paper proposes “In-Edge AI,” integrating Deep Reinforcement Learning and Federated Learning into mobile edge systems to optimize edge caching and computation offloading. It reports proof-of-concept simulations where DDQN with FL approaches centralized DDQN performance while reducing wireless transmission cost.

Key findings

  • In edge caching, DDQN with FL achieved hit rates close to centralized DDQN for three edge nodes (E1, E2, E3) and outperformed LRU, LFU, and FIFO.
  • In computation offloading, DDQN with FL achieved near centralized DDQN performance and was superior to mobile execution, edge node execution, and greedy execution for three UEs (m1, m2, m3).
  • Centralized DDQN performed better at the beginning of training, but after several FL model merges, DDQN with FL became near to centralized DDQN.
  • FL clients upload only model updates, reducing total wireless transmission compared with centralized DRL, where UEs must upload the whole training data.

Methods

  • Two simulations were run: edge caching and computation offloading. Edge caching used Xender’s one-month trace (01/09/2016 to 30/09/2016) with 9,514 active mobile users, 188,447 content files, and 2,107,100 content requests, fitted by a Zipf distribution with ∂ = 1.58; 6 edge nodes cooperated, coordinated by a cloud server.
  • Computation offloading assumed 5 MHz bandwidth divided into 10 wireless channels, 10 UEs as FL clients, channel gain quantized into 6 levels, and tasks generated by Bernoulli distribution with average arrival rate λ_T per time epoch.
  • DRL used Double DQN with tanh activation and Adam optimizer; a single-layer fully-connected feedforward neural network with 200 neurons served as MainNet and TargetNet. Parameters: replay memory M = 5000, minibatch B = 200, discount γ = 0.9, exploration ε = 0.001, learning rate ζ = 0.005, target replacement period φ = 250. A centralized DRL baseline was assumed to receive all data.

Limitations

  • The paper states FL cannot outperform centralized DL, only achieve near-best performance, and that FL trades advantages for heavier client computation load and more energy cost or burden on UEs and edge nodes; these remain open questions.
  • The paper states privacy and security are not its focus, and that In-Edge AI is future-oriented: state-of-the-art edge chips support only elementary DL training, deploying a randomly initialized DRL model could paralyze the MEC system, and delay requirements remain challenging; transfer learning is proposed to alleviate training.
  • The paper states the training-process comparison assumes an ideal wireless environment where massive centralized training data and lightweight FL updates can be uploaded within a discretized time epoch, whereas in practice massive training data cannot be uploaded without delay. It also assumes slowly changing content popularity, equal content sizes, and static popularity, user preference, and average request arrival rates over a relatively long period. Evident from the text: evaluation is simulation-based proof-of-concept, not a real-world deployment.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1109/mnet.2019.1800286

Cites

17 of 17 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

0 of 25 cited sources checked · 0 supported · 0 flagged

Show all 25 claims

  • It may state two separate claims as one. 25
  • Probably not a substantial claim. 21
38 of 38

  • The paper relies on this without support.

    A typical 5G node is expected to have 2000 or more configurable parameters.

    “a typical 5G node is expected to have 2000 or more configurable parameters”

  • The authors state this as an assumption.

    Content popularity changes slowly and all contents have the same size.

    “we assume that the content popularity changes slowly and all contents have the same size.”

  • The authors state this as an assumption.

    All content popularity, user preference and average arrival rate of requests are static during a relatively long period.

    “We assume that all content popularity, user preference and average arrival rate of requests are static during a relatively long period.”

  • The authors state this as an assumption.

    Edge node computing performance is much stronger than the UE's, so f_E is much greater than f_L.

    “the formula f E ≫ f L holds (the computing performance of edge nodes is much stronger than the UE's).”

  • The paper relies on this without support.

    Training data is large in quantity when considering massive UEs.

    “the training data is large in quantity when considering massive UEs”

  • The paper relies on this without support.

    Training data that should be uploaded to edge nodes or cloud is privacy-sensitive.

    “the training data which should be uploaded to edge nodes or cloud is privacy-sensitive”

  • The paper relies on this without support.

    If training data is transformed for privacy, server-side proxy data is less relevant than on-device data.

    “if training data is transformed for privacy consideration, server-side proxy data is less relevant than on-device data.”

  • The paper relies on this without support.

    The computation capability of a UE is relatively weak.

    “the computation capability of a UE is relatively weak”

  • The paper relies on this without support.

    The training process of a DRL agent may introduce extra energy consumption of a UE.

    “the training process of a DRL agent may introduce extra energy consumption of a UE.”

  • The paper relies on this without support.

    DRL techniques require intensive computation capacity for finding optimal solutions.

    “DRL techniques require intensive computation capacity for finding the optimal solutions.”

  • The paper relies on this without support.

    There is not enough time and data for training a DRL agent in every UE and edge node.

    “there shall not have enough time and data for training.”

  • The paper relies on this without support.

    UEs are often unpredictably off-line or allocated with poor communication resources.

    “UEs are often unpredictably off-line or allocated with poor communication resources.”

  • The paper relies on this without support.

    Some UEs may have more computation tasks and some may experience more states of mobile networks, resulting in varying amounts of training data among UEs.

    “Some UEs may have more computation tasks to be handled and some may experience more states of mobile networks, resulting in varying amounts of training data among UEs.”

  • The paper relies on this without support.

    FL distributes the quality of knowledge across a large number of devices without necessarily centralizing the training data.

    “The core benefit of FL lies in distributing the quality of knowledge across a large number of devices without necessarily centralizing the training data.”

  • The paper relies on this without support.

    Server-side proxy data is less relevant than on-device data.

    “Server-side proxy data is less relevant than on-device data.”

  • The paper relies on this without support.

    In the MEC system, massive UEs and edge nodes could act as perceptrons and acquire various, abundant and personalized data for updating the global DL model.

    “In the MEC system, massive UEs and edge nodes could act as perceptrons and acquire various, abundant and personalized data for updating the global DL model.”

  • The paper relies on this without support.

    FL inherently could address key issues about the availability of UEs and edge nodes and the unbalanced and non-IID data.

    “FL inherently could address key issues about the availability of UEs and edge nodes and the unbalanced and non-IID data.”

  • The paper relies on this without support.

    In FL, additional computation could be used to decrease the number of communication rounds to train a model.

    “In FL, additional computation could be used to decrease the number of communication rounds to train a model.”

  • The paper relies on this without support.

    An effective way to add computation is increasing computation per UE, i.e., adding more local SGD updates per round.

    “An effective way to add computation is increasing computation per UE, i.e., adding more local SGD updates per round.”

  • The paper relies on this without support.

    Learning takes long time of training as well as inferring according to the required accuracy level.

    “Learning takes long time of training as well as inferring according to the required accuracy level.”

  • The paper relies on this without support.

    In practice, massive training data are actually unable to be uploaded without any delay.

    “In practice, massive training data are actually unable to be uploaded without any delay.”

  • The paper relies on this without support.

    In the framework of FL, every client only needs to upload the update of its model.

    “In the framework of FL, every client only needs to upload the update of its model.”

  • The paper relies on this without support.

    In 5G, URLLC scenarios strongly desire very small delays and high reliability of the mobile system.

    “In 5G, the defined Ultra-Reliable Low-Latency Communications (URLLC) scenarios strongly desire very small delays and high reliability of the mobile system.”

  • The paper relies on this without support.

    General deep learning-based optimization and prediction schemes take quite long running time of recursions for converging to the results, which is inappropriate for mobile edge systems.

    “However, general deep learning-based optimization and prediction schemes take quite long running time of recursions for converging to the results, which is inappropriate for mobile edge systems; particularly the system-level edge computing tasks desiring rapid responses in the scale of millisecond.”

  • The authors state this as an assumption.

    The experiment assumes an ideal wireless environment.

    “Nonetheless, this experiment assumes an ideal wireless environment.”
25 of 25

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Citations

Cites

Flagged first, then by citations

Show all 17 references

Research works printed but not linked: 1

  1. 2014, Mobile Edge Computing – Introductory Technical White Paper

Cited by

Sorted by citations

Items 901–1,000 of 1,000

Related

See all

Metadata

Authors
Wang, X., Han, Y., Wang, C., Zhao, Q., Chen, X., & Chen, M.
Affiliations
  • Tianjin University (CN)
  • Huawei Technologies (China) (CN)
  • Arizona State University (US)
  • Huazhong University of Science and Technology (CN)
ORCID
Venue
IEEE Network
Publisher
Institute of Electrical and Electronics Engineers (IEEE)
Published
2019-09
Volume
33
Issue
5
Pages
156-165
ISSN
  • 0890-8044 (ISSN-L)
  • 1558-156X
DOI
10.1109/mnet.2019.1800286
OpenAlex
W2962804345
Semantic Scholar
6f89a632ceb8fcb81eac3d7b52e937099659cc6a
Cited by
1,030
References
17
Citations by year
2026: 44 · 2025: 95 · 2024: 127 · 2023: 165 · 2022: 215 · 2021: 196 · 2020: 160 · 2019: 26
FWCI
81.72
Type
journal article
Language
en
Open access
Licence
Locations
  • IEEE Network (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Topics
  • Physical Sciences › Computer Science › Computer Networks and Communications › IoT and Edge/Fog Computing
  • Physical Sciences › Computer Science › Artificial Intelligence › Privacy-Preserving Technologies in Data
  • Physical Sciences › Computer Science › Computer Networks and Communications › Caching and Content Delivery
Keywords
mobile edge computing, federated learning, deep reinforcement learning, edge caching, content distribution, communication burden reduction, edge nodes, device collaboration, dynamic optimization

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-10-09, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-09 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review