PaperlayerView exact revision on arXiv

Source-linked AI summary

Distributed Machine Learning for Wireless Communication Networks: Techniques, Architectures, and Applications

S. Hu, X. Chen, W. Ni, E. Hossain, X. Wang

arXiv:2012.01489v1cs.LG

TL;DR

DML techniques have not been systematically reviewed for applicability to wireless systems. This survey examines DML techniques and applications, covers four application areas, and discusses adversarial attacks, countermeasures, and future directions.

  • Problem

    DML techniques have not been systematically reviewed for applicability to wireless systems.

    Source: §I · p. 2

  • Method

    The paper surveys DML techniques and wireless applications, analyzing optimality, scalability, convergence, computation and communication costs, and privacy and security approaches.

    Source: §I · p. 2, §XI · p. 27

  • Results

    The survey covers power control, spectrum management, user association, and edge cloud computing, and discusses adversarial attacks and countermeasures.

    Source: §XI · p. 27

  • Takeaways & Limitations

    The survey outlines future directions for integrating DML into wireless systems design for flexibility and efficiency.

    Source: §XI · p. 27

Abstract

from arXiv · show

Distributed machine learning (DML) techniques, such as federated learning, partitioned learning, and distributed reinforcement learning, have been increasingly applied to wireless communications. This is due to improved capabilities of terminal devices, explosively growing data volume, congestion in the radio interfaces, and increasing concern of data privacy. The unique features of wireless systems, such as large scale, geographically dispersed deployment, user mobility, and massive amount of data, give rise to new challenges in the design of DML techniques. There is a clear gap in the existing literature in that the DML techniques are yet to be systematically reviewed for their applicability to wireless systems. This survey bridges the gap by providing a contemporary and comprehensive survey of DML techniques with a focus on wireless networks. Specifically, we review the latest applications of DML in power control, spectrum management, user association, and edge cloud computing. The optimality, scalability, convergence rate, computation cost, and communication overhead of DML are analyzed. We also discuss the potential adversarial attacks faced by DML applications, and describe state-of-the-art countermeasures to preserve privacy and security. Last but not least, we point out a number of key issues yet to be addressed, and collate potentially interesting and challenging topics for future research.

I. INTRODUCTION

This survey addresses the need for distributed machine learning (DML) in wireless networks, where capable devices, growing data volumes, channel congestion, and privacy concerns favor distributed learning. It reviews DML methods and applications, their efficiency and security, and open research challenges.

I. INTRODUCTION

  • Wireless devices can learn locally and share training results such as gradients rather than raw data, reducing data exchange and server burden while preserving privacy.The motivation includes devices’ computing resources, explosively growing data, congested wireless channels, and user privacy concerns.Source: §I · pp. 1–2
  • Prior surveys largely cover centralized wireless ML or federated learning alone, leaving DML’s algorithms, frameworks, architectures, efficiency, convergence, scalability, privacy, and security insufficiently reviewed.The paper identifies a gap in systematic study and comparison of DML in wireless systems.Source: §I · p. 2
  • The survey reviews DML techniques and architectures for wireless networks, covering power control, spectrum management, user association, edge cloud computing, efficiency, privacy, and security.It analyzes optimality, scalability, convergence, computation and communication costs, and countermeasures to adversarial attacks against servers and agents.Source: §I · p. 2
  • The survey finds extensive DML work on wireless resource management and transmit-power control, with federated learning dominant and multi-agent reinforcement learning a fast-growing alternative.It identifies distributed applications of multi-agent reinforcement learning in resource allocation, protocol coexistence, and cooperative communications, and highlights unresolved issues and future research topics.Source: §I · p. 3

II. ALGORITHMS AND MODELS · A. Deep Learning (DL) · B. Reinforcement Learning (RL)

Deep learning extracts features automatically and trains neural networks through differentiable losses, while reinforcement learning learns control policies from state transitions and rewards. The section describes single- and multi-agent RL, Q-learning, and deep RL.

A. Deep Learning (DL)

  • Deep learning learns features from raw data layer by layer, avoiding hand-crafted feature design used in conventional machine learning,.Neural networks are the most popular deep learning models; networks typically need at least two hidden layers to be considered deep.Source: A. Deep Learning (DL) · p. 3
  • DNN neurons apply weighted sums and nonlinear transformations across layers, and SGD with back-propagation minimizes a differentiable loss to obtain model parameters.DNNs approximate complex functions using simple nonlinear operations on input data.Source: A. Deep Learning (DL) · pp. 3–4

B. Reinforcement Learning (RL)

  • RL repeatedly maps the system state to an action, observes the resulting state and reward, and learns to maximize the total long-term reward.Source: B. Reinforcement Learning (RL) · p. 4
  • Single-agent RL models decisions as a Markov decision process, whereas multi-agent RL uses a stochastic game with state transitions and rewards shaped by agents’ joint actions.When all agents share reward functions and returns, the game is fully cooperative; otherwise, their goals may differ or oppose one another.Source: B. Reinforcement Learning (RL) · p. 4
  • Q-learning learns by trial and error without a stored dataset or environment model, updating Q(s, a) until it can choose the action with the best value [40]–.Q(s, a) represents the expected cumulative reward for taking action a in state s under policy π.Source: B. Reinforcement Learning (RL) · p. 4
  • Deep reinforcement learning uses DNNs to estimate value functions or policy functions, enabling generalization to high-dimensional environments,.The neural-network agent receives reward feedback from the environment, selects actions that alter its state, and configures its parameters to seek the best future return.Source: B. Reinforcement Learning (RL) · pp. 4–5

C. Stochastic Gradient Descent (SGD) … 1) Horizontal FL (or Sample-Based FL):

SGD trains models through randomly sampled gradient estimates, while FL trains across locally held datasets without exchanging the underlying data. Horizontal FL applies when participants share feature spaces but have different samples.

C. Stochastic Gradient Descent (SGD)

  • SGD randomly samples an example and updates the model with its loss gradient as a stochastic estimate of the full-dataset gradient.The loss function measures divergence between the model output distribution and the true input-data distribution.Source: C. Stochastic Gradient Descent (SGD) · p. 5
  • Momentum-based SGD adds weighted previous gradients to reduce fluctuations, while parallel mini-batch SGD updates batches simultaneously to speed convergence [47],.Decentralized parallel SGD limits gradient collection to neighbors rather than global aggregation [49].Source: C. Stochastic Gradient Descent (SGD) · p. 5
  • ADMM decomposes large-scale ML problems into parallel subproblems and schedules them to solve the original problem; it guarantees linear convergence under certain conditions.Source: C. Stochastic Gradient Descent (SGD) · p. 5

A. Federated Learning (FL)

  • FL trains models on datasets held locally by different devices, avoiding data exchange, and commonly uses SGD or FedAvg.In FedAvg, clients update model parameters on their own data and the server averages those parameters; clients may perform multiple local updates before averaging.Source: A. Federated Learning (FL) · p. 5

1) Horizontal FL (or Sample-Based FL):

  • Horizontal FL fits datasets with the same feature space but different samples, such as uniformly deployed base stations and users in low-to-medium-density suburban areas.Users can collectively improve uplink power control, modulation-and-coding scheme, and QoS parameters.Source: 1) Horizontal FL (or Sample-Based FL) · pp. 5–6
  • In horizontal FL, each mobile terminal updates model parameters and uploads them to a server or cloud, which establishes a global model with contributions from individual data owners.The framework relies on the shared feature spaces across users.Source: 1) Horizontal FL (or Sample-Based FL) · p. 6

2) Vertical FL (or Feature-Based FL): … IV. CONVERGENCE AND DIVERGENCE OF DML ALGORITHMS

The section distinguishes federated and partitioned learning by how they handle data and model parameters, then introduces convergence and divergence as measures for evaluating DML algorithms. It describes vertical FL and FTL for learning across distinct feature spaces, alongside partitioned learning and its decentralized model optimization.

2) Vertical FL (or Feature-Based FL):

  • Vertical FL combines distinct features from data owners with samples from the same distribution, enabling cellular BSs and core gateways to train a shared model while protecting privacy [56]–.The model can support user association, handover, resource allocation, and downlink power control; participants contribute features and gradients toward a common policy.Source: 2) Vertical FL (or Feature-Based FL) · p. 6

3) Federated Transfer Learning (FTL):

  • FTL addresses data owners whose datasets differ in both samples and features by mapping their characteristic spaces using scarce shared samples to predict single-sided samples [60], [61].The example involves mobile subscribers of different wireless operators with limited overlap in users and features such as power control, MCS, and QoS.Source: 3) Federated Transfer Learning (FTL) · pp. 6–7
  • An operator can transfer its trained model to another operator’s network when a user roams, extending FTL beyond problems addressed by horizontal and vertical FL.Vertical FL and FTL had yet to be implemented in wireless applications, but their ability to train with distinct features was presented as promising for large-scale geographically distributed systems.Source: 3) Federated Transfer Learning (FTL) · p. 7
  • A distributed deep-learning framework uses asynchronous SGD and a parameter server so participants train locally and share only gradients, not input datasets.Participants can benefit from one another’s models while protecting sensitive data and privacy and achieving global-model convergence and learning accuracy.Source: 3) Federated Transfer Learning (FTL) · p. 7

B. Partitioned Learning

  • Partitioned learning assigns different parameter blocks of a supervised model to devices for decentralized computation, which return updates to a server; unlike FL, it partitions the model.The Parameter Server framework uses block coordinate descent to decompose large optimization problems, such as linear regression and SVM, into separable objectives solved iteratively,, –.Source: B. Partitioned Learning · p. 7, Figure 6 · p. 7

IV. CONVERGENCE AND DIVERGENCE OF DML ALGORITHMS

  • The section evaluates DML algorithms by convergence speed, optimality, and the gap between input-data labels and outputs; more iterations can improve performance but raise computational complexity.Convergence concerns whether solution sequences approach a solution point, while divergence measures or approximates the gap between input and output probability distributions.Source: §IV · p. 7

A. Convergence of Deep Learning · B. Convergence of Federated Learning

The deep-learning discussion examines ADMM and parallel SGD convergence, including their assumptions and communication tradeoffs. The federated-learning discussion covers methods intended to accelerate convergence or reduce communication, with reported CE-FedAvg results.

A. Convergence of Deep Learning

  • ADMM has linear convergence in practical settings such as LASSO without strict target-function convexity; its root-linear guarantee requires stated accuracy and step-size conditions.The discussion contrasts ADMM’s parallel subproblems, linear scaling with data, and lack of gradient updates with SGD’s vanishing gradients and sensitivity to poorly qualified data; the analysis also states assumptions involving the optimum, differentiability, feasible set, and objective.Source: A. Convergence of Deep Learning · pp. 7–8
  • DLADMM refreshes key parameters backward then forward and uses quadratic approximation and backtracking to avoid matrix inversion, reducing complexity from O(N^3) to O(N^2) while proving global convergence for an ADMM-based DNN.Source: A. Convergence of Deep Learning · p. 8
  • With N workers, PMSGD achieves O(1/(NT)) convergence, N times faster than single-node SGD’s O(1/T), but requires simultaneous local updates and incurs inter-node communication costs [77].The supplied passages state that 1/ε^2 iterations achieve an O(ε)-precise solution.Source: A. Convergence of Deep Learning · p. 8
  • O(1/(NT)) convergence with substantially fewer transmission rounds is achieved by certain communication-skipping parallel SGD variants; neighbor exchanges [49] and compression or sparsification, address communication impediments.Source: A. Convergence of Deep Learning · p. 8

B. Convergence of Federated Learning

  • Adaptive FedAvg uses decentralized Adam to reduce rounds, while CE-FedAvg sparsifies gradients and quantizes weights to 8-bit unsigned integers to reduce transmission load for non-i.i.d. data.Source: B. Convergence of Federated Learning · p. 8
  • CE-FedAvg stabilizes at required precision with six times fewer rounds and transmits three times less information than FedAvg; Raspberry-Pi trials reach precision up to 1.7 times faster.Tests covered MNIST and CIFAR-10, i.i.d. and non-i.i.d. data, varying client counts and participation ratios, and data-reduction schemes.Source: B. Convergence of Federated Learning · pp. 8–9
  • Decentralized FedAvg [54] and combines local SGD with server averaging and randomly selects non-repeating edge nodes each round; uses a circular topology with a path between every node pair.Experiments examine FedAvg on non-i.i.d. MNIST training data.Source: B. Convergence of Federated Learning · p. 9
  • Momentum FL [47] uses previous-iteration gradients, analyzes global convergence and upper-bounds its speed, and provides conditions for faster convergence; MNIST experiments show global convergence and improvement over FL.Source: B. Convergence of Federated Learning · p. 9

C. Divergence · V. COMPUTATION AND COMMUNICATION EFFICIENCY IN DML · A. Computation

The section explains how divergence-based losses support wireless and federated learning, then surveys computation approaches that seek faster distributed training while balancing model performance against computation and communication costs.

C. Divergence

  • KL divergence measures the gap between distributions; with fixed p, cross-entropy grows as q diverges from p, but it is not an absolute difference across varying p, and ML/FL often treat p as truth and q as the model –, –.When p and q are identical, cross-entropy remains nonzero and equals H(p).Source: C. Divergence · p. 9
  • For classification, minimizing cross-entropy against an empirical distribution minimizes KL divergence and maximizes log-likelihood, while the entropy term is model-parameter independent and negligible in the loss, [87]–[89].The empirical distribution assigns probability 1 to the observed class and 0 to other classes.Source: C. Divergence · p. 9
  • The LSTM-based sparse-vector approach minimized cross-entropy and substantially outperformed Simultaneous Orthogonal Matching Pursuit and model-based Bayesian schemes on two practical datasets, without adding encoder intricacy,.It modeled conditional probabilities of non-zero elements from preceding vectors and used a greedy decoder to re-establish sparse vectors.Source: C. Divergence · p. 9
  • FOCUS [99] uses server-held baseline samples and joint cross-entropy accuracy scores to weight agents, distinguishing noisy-label agents and diminishing their influence in synthetic- and practical-data experiments.The passage says FOCUS scored substantially higher, but its comparison target is not present in the supplied text.Source: C. Divergence · p. 9

V. COMPUTATION AND COMMUNICATION EFFICIENCY IN DML

  • Larger datasets and more communication rounds can improve training accuracy, but the resulting data volume raises dimensionality, complexity, and computation and communication costs, motivating efficiency approaches that preserve model performance.Source: §V · p. 10

A. Computation

  • A directed-acyclic-graph model was developed to reduce training time for decentralized synchronous SGD, where GPU workers repeatedly perform feed-forward and back-propagation while transmitting parameters or gradients [100]–.The passage identifies inter-GPU information transmission as a potential bottleneck to system-wide learning.Source: A. Computation · p. 10
  • On Amazon EC2, the coded coordination policy considerably reduced average execution duration versus current coded and uncoded approaches, with an insignificant gap to the established lower limit.The master completes each SGD round after receiving k distinguished calculations, while agents perform superfluous calculations to mitigate stragglers.Source: A. Computation · p. 10
  • Tests found that Caffe-MPI, CNTK, MXNet, and TensorFlow performed differently with the same GPUs and datasets; experiments identified throughput bottlenecks and remedies including preprocessing, autotuning, input layout, and cuDNN utilization –.The frameworks were compared for training speed on single and multiple GPUs; the paper emphasizes joint optimization of the DML model and implementation framework.Source: A. Computation · p. 10

B. Communication … A. MapReduce

The sections survey communication-efficient distributed learning, computation–communication trade-offs, and DML architectures, including MapReduce’s strengths, limitations, and applications.

B. Communication

  • Wireless-resource scarcity and lossy links can impair federated-learning accuracy and convergence, while sparse-data federated optimization reduces communication rounds and bandwidth.Federated optimization keeps parameters local to reduce uplink costs and protect device privacy; asynchronous uploading remains a potential improvement.Source: B. Communication · p. 10
  • Hierarchical gradient quantization substantially reduces information-exchange overhead versus sign-SGD while yielding similar training precision.The scheme decomposes gradients into normalized blocks, quantizes and adjusts them, and uses bit assignment to reduce measurement error; its convergence speed depends on uniform-quantizer bit count.Source: B. Communication · pp. 10–11, Figure 7 · p. 11
  • CMFL achieves 13.97 times vanilla FL’s transmission efficiency and improves FMTL transmission efficiency 5.7 times with 4% better estimation precision.It filters client renewals misaligned with the global model inclination; Gaia yields 1.26 times efficiency enhancement over vanilla FL.Source: B. Communication · p. 11

C. Trade-off Between Computation and Communication

  • Four decentralized deep-learning architectures do not scale well on Tesla V100 GPUs because 100 Gbps InfiniBand gradient transmission cannot keep pace with computation.Training speedup depends on input/output, computation, and transmission behavior.Source: C. Trade-off Between Computation and… · p. 11
  • GADMM converges to the optimum for convex losses and is faster and more transmission-efficient than LAG on linear and logistic regression tasks.Dynamic GADMM also achieves convergence under changing network topologies; a separate non-i.i.d. decentralized-gradient protocol behaves near-optimally in global loss across models and data distributions.Source: C. Trade-off Between Computation and… · pp. 11–12

VI. DML ARCHITECTURES

  • The DML architectures section surveys popular architectures used in wireless-network studies and points readers to a summary of their properties in Table IV.Source: §VI · p. 12

A. MapReduce

  • MapReduce parallelizes batch computation across machine clusters, but its failure to recognize iterative ML algorithms motivated iterative MapReduce systems such as Spark MLlib, Vowpal Wabbit, and Twister.The section introduces several popular DML architectures and summarizes their properties in Table IV.Source: A. MapReduce · p. 12, §VI · p. 12
  • K-means and PCA each detect 97% of actual anomalies; Spark DDT reduces learning time by 8 minutes on average for datasets up to gigabytes.Anomaly detection efficiency increases by 75% at high parallelization; residential electricity forecasting also shows computational-time and predictive-accuracy gains from decentralized systems.Source: A. MapReduce · p. 12

B. Parameter Server (PS) · C. Graph Processing Architecture

The section describes parameter-server coordination for distributed learning and graph-processing architectures that support model parallelism and locality. It also presents Cymbalo’s design and reported speedups, alongside a graph-based sensor-network application.

B. Parameter Server (PS)

  • Parameter servers maintain all or part of the parameters, aggregate weights from client groups, and exchange parameter copies among servers.Servers typically correspond in number to cluster hosts and synchronize only partially with one another, reducing server-client data transfer.Source: B. Parameter Server (PS) · p. 13
  • Clients perform convolution, gradient calculation, backpropagation, and weight refreshment using server-synchronized global gradients.The supplied passage says clients share parameters, but its description of with whom is truncated.Source: B. Parameter Server (PS) · p. 13

C. Graph Processing Architecture

  • Unlike iterative MapReduce and parameter-server frameworks, graph-based frameworks support model parallelism and distribute training data and parameters within one cluster for locality and lower communication overhead.The comparison attributes the locality and communication benefits to graph-based machine learning frameworks.Source: C. Graph Processing Architecture · p. 13, Figure 10 · p. 14
  • Graph-processing systems use vertex-centric computation: vertices respond to neighboring vertices or the system and halt, with the task ending when all vertices halt.Parameters are modeled as vertices, while iterative MapReduce and parameter-server designs store parameters separately from data or in a single node.Source: C. Graph Processing Architecture · p. 14
  • TuX2 combines graph engines with parameter-server concepts, offering edge-centric data exchange and vertex-centric updates, but translating vector evaluations into graph operations is complicated.Cymbalo’s vector-centric programming architecture was proposed to simplify ML implementation by avoiding low-level graph details.Source: C. Graph Processing Architecture · p. 14
  • On Spark 2.1.2, Cymbalo expedited ML operations by 1.6 to 5.8 times compared with other advanced DML systems.Cymbalo combines heterogeneous-data storage and a hybrid ML-operations framework with vector-centric programming, which the paper describes as more efficient than BiGraph and TuX2.Source: C. Graph Processing Architecture · p. 14
  • In a Boltzmann-machine sensor-network simulation with hundreds of trajectories, inference accuracy depended on observations and sensor positions.The network models sensors and inference units as graph vertices; inference units decide whether detected information should be transmitted onward.Source: C. Graph Processing Architecture · p. 14

VII. SOFTWARE PLATFORMS AND LIMITATIONS · A. MapReduce-Based Software Platform · B. Parameter Server-Based Software Platform

The section surveys MapReduce- and parameter-server-based software platforms for distributed machine learning, describing their mechanisms, applications, and limitations.

A. MapReduce-Based Software Platform

  • Hadoop partitions data and computation across hosts and can scale by adding servers; HDFS replicates file blocks across DataNodes, but its NameNode’s limited capacity restricts addressable blocks.The HDFS client asks the NameNode for DataNodes to handle block replicas, then pipelines data to those nodes; Fig. 11 depicts this architecture.Source: A. MapReduce-Based Software Platform · pp. 14–15, Figure 11 · p. 15
  • Spark keeps iterative ML state in memory and gains performance over Hadoop, but Spark ML can lag specialized implementations when fine-grained computation and communication scheduling is absent.Proper scheduling is important for fast and correct ML program implementation.Source: A. MapReduce-Based Software Platform · p. 15
  • S4 processes continuous keyed-event streams through processing elements that emit events and publish results, unlike MapReduce batch computing, but lacks robust live-PE migration.Its actor-based architecture supports encapsulation, location transparency, and concurrent applications, and is designed for large clusters using commodity hardware.Source: A. MapReduce-Based Software Platform · pp. 15–16
  • BAIPAS balances training data across workers based on data size and available memory and CPU/GPU, then shuffles data so each worker can learn from the full dataset and avoid overfitting.Source: A. MapReduce-Based Software Platform · p. 16
  • Coded techniques address stragglers and communication bottlenecks; coded computation speeds distributed matrix multiplication by log n, while coded shuffling multicasts a common coded message instead of separate unicasts.The encrypted erasure-code framework introduces redundant submissions and recovers results from a subset, dropping unfinished submissions.Source: A. MapReduce-Based Software Platform · p. 16

B. Parameter Server-Based Software Platform

  • Petuum uses error tolerance, dynamic dependencies, and non-uniform convergence to guide parameter synchronization and scheduling, but needs better partial-state recovery and real-time resource allocation.Its parameter-server APIs expose shared global models through a single-machine-like interface, while quantifying asynchronous inconsistency to support convergence without explicit network synchronization.Source: B. Parameter Server-Based Software… · p. 16
  • 10 ∼50% training-time savings over TensorFlow were reported for a shared-memory DNN framework training CNN and MLP models for image recognition.The framework uses remote shared memory across workers to maintain global shared parameters and accelerate parameter reads and updates.Source: B. Parameter Server-Based Software… · p. 16

C. Graph-Based Platform … B. Spectrum Management

These sections describe limitations of graph-based platforms and applications of distributed machine learning in wireless networks, including power control and spectrum management under computation, communication, and privacy constraints.

C. Graph-Based Platform

  • Graph-based platforms such as Pregel and GraphLab divide graph models using embedded schedulers and consistency controllers, but their asynchronous execution is not established as correct for all ML tasks.Pregel provides a scalable API for graph algorithms; limitations of popular software platforms are summarized in Table V.Source: C. Graph-Based Platform · p. 16

VIII. APPLICATIONS OF DML TO WIRELESS NETWORKS

  • Wireless DML frameworks such as federated learning and partitioned learning enable devices to acquire a global model with limited data exchange or partial models and datasets,, [151], [152].They address limits of centralized cloud or edge training, including bandwidth, channel conditions, latency, energy consumption, and privacy, as well as the computation and storage demands of data-driven ML on mobile devices.Source: §VIII · p. 16

A. Power Control

  • The hybrid CFmMIMO power-control optimization reduces FL training time by up to 55% versus three benchmark schemes, but assumes time-invariant channel state information.The benchmarks use equal downloading power with maximum uploading power and fixed local accuracy; equal power with maximum uploading power and local accuracy; or equal power with fixed local accuracy and downloading and uploading powers.Source: A. Power Control · p. 17
  • Gradient-statistics-based over-the-air FL power management outperforms full-power communication and threshold-based power management in model precision and convergence speed [153].It minimizes model-gathering mean square error by jointly selecting device transmit powers and an edge-server denoising coefficient; its closed-form optimum requires a priori gradient statistics, which can be estimated from historical gradients.Source: A. Power Control · p. 17
  • In cooperative FL, devices select relay nodes and transmit powers through a Stackelberg game, whose equilibrium is computed using the exterior point method [154].Devices transfer locally generated model updates via relays to a model master, which pays them for training; the game accounts for interference among devices using a shared relay.Source: A. Power Control · p. 18

B. Spectrum Management

  • FEEL bandwidth coordination assigns wider spectra to devices with poorer channels or weaker computing capacity, significantly saving energy while adapting to channel and computation conditions [64].The strategies target minimum mobile-device energy while maintaining the training effect, but may not generalize to other model-transmission cases such as asynchronous models.Source: B. Spectrum Management · p. 18
  • Hierarchical FL with gradient sparsification and periodic averaging reduces transmission delay without sacrificing model precision in CIFAR-10 image processing.Micro- or pico-BSs train with mobile users and send model updates to a macro-BS for network-wide agreement; sparsification transmits only part of the parameters each round.Source: B. Spectrum Management · p. 18, Figure 14 · p. 19
  • WoLF Q-learning for decentralized spectrum allocation improves file-transmission failure and stoppage probabilities versus a fixed learning rate of 0.1.The stadium-network simulations use αwin = 0.01 and αlose = 0.05, so the device learns more slowly after success and faster after failure.Source: B. Spectrum Management · pp. 18–19

C. QoS Provisioning … 2) Power allocation and latency:

The section surveys distributed learning for wireless resource and QoS problems, describing local decision-making and federated training alongside channel constraints, convergence trade-offs, and application-specific methods.

C. QoS Provisioning

  • Decentralized multi-agent Q-learning lets BSs make independent ON/OFF decisions from local battery states and receive rewards reflecting throughput and battery level,.In the RES-powered BS example, each agent locally updates its Q value from its agent-dependent reward.Source: C. QoS Provisioning · p. 19
  • FL is used to reduce computation and communication latency [155] and predict traffic for higher user throughput [156], but wireless bandwidth limits which users can renew models.Users with sufficiently reliable SINR may be selected for updates; wireless interference also affects convergence, which is analyzed across coordination strategies.Source: C. QoS Provisioning · p. 19
  • FL converges faster with proportional-fair than random or round-robin scheduling at high SINR targets, while round-robin is favored at low targets.Convergence drops sharply as the SINR target rises, and analysis identifies a trade-off between coordinated-user numbers and subchannel bandwidth.Source: C. QoS Provisioning · pp. 19–20
  • CA-DSGD converges faster and with better accuracy than D-DSGD in simulations for power-constrained agents training over a bandwidth-constrained fading MAC.CA-DSGD uses sparse gradient estimates, past-iteration errors, lower-dimensional projections, and a power-assignment policy to align gradients at the parameter server.Source: C. QoS Provisioning · p. 20

1) User association:

  • An ESN-based FL approach lets multiple BSs train locally and cooperatively estimate wireless VR users’ positions and poses from captured datasets.Source: 1) User association · p. 20

2) Power allocation and latency:

  • For vehicular URLLC, FL decentralizes joint power and resource allocation to minimize queueing delay by modeling queue tails [160].The cited FL studies, [160] do not account for wireless-channel lossiness or limited bandwidth, which may compromise model quality, performance, and convergence.Source: 2) Power allocation and latency · p. 20

3) Bandwidth allocation and user selection: · 4) Edge cloud computing: · IX. PRIVACY AND SECURITY

The sections describe wireless FL resource and user coordination, edge-cloud learning designs, and privacy and security threats and countermeasures for distributed learning.

3) Bandwidth allocation and user selection:

  • Wireless FL performance depends on packet errors and limited, time-varying channel capacity, so the base station selects a user subgroup and jointly assigns radio resources and user pairings.The best transmit power depends on user selection and channel assignment; the proposed approach uses a closed-form convergence-speed expression and convex search.Source: 3) Bandwidth allocation and user… · p. 20
  • Hybrid FL and communication reduces FL loss by at most 10% versus best user selection with stochastic resource assignment and 16% versus standard FL with stochastic user choosing and resource assignment.The framework jointly learns radio resource assignment and user pairing to minimize FL loss, using estimated convergence speed and convex search to select users and allocate channels.Source: 3) Bandwidth allocation and user… · p. 20
  • For bandwidth-constrained FL, the coordination strategy combines link states with local-update prominence; one agent per round is ideal for i.i.d. data, while multiple agents help with non-i.i.d. data.Each selected agent compresses its model renewal to fit channel capacity, and the number coordinated should rise for less diversified and impartial data distributions.Source: 3) Bandwidth allocation and user… · p. 20

4) Edge cloud computing:

  • FEEL jointly optimizes communication and computation to improve energy efficiency, but its full-model updates typically limit it to small or moderate training missions.Under a training-accuracy requirement and limited learning period, the design minimizes agents’ sum energy by optimizing transmit power, transmission speeds, and CPU configuration; it uses NOMA or TDMA.Source: 4) Edge cloud computing · pp. 20–21
  • PABA integration improves PARTEL latency by 46% and accuracy by 4%, while decentralized training matches centralized convergence rate within the same iterations.PABA uses successive bandwidth-aware or parameter-aware allocation schemes; the collaborative approach optimizes PABA using nesting bisection search and a convex program.Source: 4) Edge cloud computing · p. 21, Figure 15 · p. 21
  • A fairness-constrained real-time stream decomposition scheme beats the classic scheme on data rate and cost-effectiveness while evaluating only 24% of system connections in simulation.It decomposes continuous streams under dynamic system states, using a data-decomposition fairness metric and SGD to obtain real-time policies and asymptotically the largest decomposition gain.Source: 4) Edge cloud computing · pp. 21–22
  • DL caching schemes reduce RMSE by up to 33.7% and service latency by 47.4% versus other ML approaches; decentralized data exchange avoids releasing users’ individual data.The centralized scheme has the content server collect data from every mobile edge node, whereas the decentralized scheme lets nodes cooperate without sharing individual user data with the server.Source: 4) Edge cloud computing · p. 22

IX. PRIVACY AND SECURITY

  • Distributed learning faces privacy and security threats from malicious servers, eavesdroppers, and data poisoning, and the section reviews countermeasures for agents and servers.The discussion addresses protecting data against third parties and other devices within the distributed system.Source: §IX · p. 22

A. Privacy of Agent … 2) Secure Multi-Party Computation (SMC):

Distributed learning creates privacy risks for both agents, whose updates may expose information to the server, and servers, whose released parameters may be exposed to eavesdroppers. The section describes methods to protect both sides, including secure multi-party computation (SMC).

A. Privacy of Agent

  • In federated learning, agents may distrust the server because it can manipulate training or infer private information from their submitted parameters and weights.The section introduces privacy-preserving techniques as ways to address this risk.Source: A. Privacy of Agent · p. 22

1) Perturbation:

  • Differential privacy adds artificial noise to uploaded parameters, and Geyer et al. [173] report client-side privacy at little model-performance cost when enough agents participate.Their scheme aims to conceal agents’ training contributions while balancing privacy loss against learning performance.Source: 1) Perturbation · p. 22

2) Dummy:

  • Dummy data can hide true client information, but added redundancy costs bandwidth, buffer, and energy; Diyanat et al. minimize weighted communication cost and privacy degree while preserving data statistics.Kido et al. describe sending counterfeit locations alongside the true location; Diyanat et al.’s design guarantees a thresholded probability of server error in estimating the true data.Source: 2) Dummy · p. 22

3) Encoding:

  • The encoder in sends high-level data characteristics to a central server instead of raw data, preventing raw-data sharing while, for example, withholding the image itself.It extracts collective characteristics from admitted data and relays them to the server.Source: 3) Encoding · p. 22

B. Privacy of Server

  • Server-side privacy matters because released aggregated model parameters can leak to passive network eavesdroppers during synchronization.The server typically computes a data-size-weighted average of agent parameters before release.Source: B. Privacy of Server · p. 22

1) Aggregation:

  • Aggregation produces a global update or replacement model from agent updates, but FL collection mechanisms remain an open problem; hierarchical asynchronous learning trades exposure and convergence-control limitations against shorter convergence time.The hierarchy offers flexible parameter choices, but does not optimally prevent data exposure or select step sizes, potentially worsening classification performance.Source: 1) Aggregation · p. 23

2) Secure Multi-Party Computation (SMC):

  • SMC lets agents report an aggregate without exposing individual updates, using a four-round protocol that encrypts and masks model renewals before decryption.The protocol uses public-key encryption, homomorphic encryption, or Shamir’s secret sharing; a separate homomorphic-encryption design protects global model synchronization while maintaining accuracy.Source: 2) Secure Multi-Party Computation (SMC) · p. 23, Figure 16 · p. 23

C. Techniques and Implementation … 3) Blockchain-based solutions:

The section surveys distributed defenses for poisoning, privacy, and Byzantine failures, including differential privacy, encryption, blockchain, and Byzantine-resilient learning. It describes reported performance conditions and limitations alongside the proposed mechanisms.

1) Prevention of data poisoning:

  • DSVMs are less prone to poisoning attacks in networks with several nodes and higher degrees, but resistance depends on architecture and attack intensity.Poisoned samples can be typical but mislabeled, or atypical to the domain.Source: 1) Prevention of data poisoning · pp. 23–24
  • Zhang and Zhu [179], model poisoning as a zero-sum game and use a fully distributed ADMM algorithm to compute agents’ responses to hostile activity.Convergence is proved in without assumptions on sample data or network topology.Source: 1) Prevention of data poisoning · p. 24

2) Perturbation plus encryption:

  • The privacy-preserving FL scheme applies decentralized Gaussian perturbation, then nests BGV ciphertext inside A-LWE ciphertext for secure gathering.After gathering agents’ ciphertexts, the server can decode collected values without exposing individual agents’ privacy.Source: 2) Perturbation plus encryption · p. 24
  • Above a 0.5 collusion ratio, CNN accuracy drops dramatically, while privacy-budget changes have little influence under the tested scheme.The algorithm was tested on MNIST; its performance on high-dimensional datasets needs further qualification.Source: 2) Perturbation plus encryption · p. 24
  • The blockchain FL framework uses PoS or BFT consensus to decentralize training data; adjusting block-generation rate and PoW difficulty can reduce latency and minimize loss.Its latency may increase dramatically as SNR decreases.Source: 3) Blockchain-based solutions · p. 24
  • BYRDIE, avoids sharing primitive data and uses outer coordinate descent with an inner Byzantine-resilient solver for scalar-valued subproblems.The approach assumes the loss is the minimum of a convex combination of local empirical risk functions.Source: 2) Perturbation plus encryption · p. 24

3) Blockchain-based solutions:

  • BYRDIE scales with problem dimensionality, detects Byzantine failures, and improves the lowest statistical risk when training uses i.i.d. datasets.Its local scrutinizing process is incorporated into each iteration of distributed coordinate descent.Source: 3) Blockchain-based solutions · pp. 24–25

X. LESSONS LEARNED AND OPEN ISSUES … XI. CONCLUSION

The survey reviews DML techniques, applications, architectures, and tradeoffs in wireless networks, summarizes privacy and security countermeasures, and identifies open directions for integrating DML into wireless systems.

A. Lessons Learned

  • Federated and partitioned learning support global models with little or no data exchange, while SGD variants enable distributed training and ADMM offers an alternative for some learning problems.The survey says the techniques allow dataset partitioning and can reach global optimality, but may not converge quickly; partitioned learning supports model partitioning, while federated RL does not require stored training data. SGD variants can train large-scale models with fast convergence without sharing datasets. ADMM prevents gradient disappearance and tolerates poorly qualified inputs; it achieves global root-linear convergence when the objective combines a smooth, strongly convex function with a non-smooth ℓ1 regularizer.Source: A. Lessons Learned · p. 25
  • Privacy defenses include client-side perturbation, dummy data, encoding, and blockchain, plus server-side aggregation and secure multi-party computation; each entails performance, resource, or overhead tradeoffs.With sufficient participating agents, perturbation can preserve client-side differential privacy at a small model-performance cost [173]. Dummy methods require extra bandwidth and buffer and consume energy; perturbation plus encryption is effective against curious servers when compromised agents are fewer than 50%.Source: A. Lessons Learned · p. 25
  • MLC uses DML to improve wireless communications, whereas CML optimizes edge learning under device and link constraints; architecture choice depends on model and data scale and cluster characteristics.Iterative MapReduce supports only synchronous communication and can suffer stragglers; parameter-server architectures support synchronous or asynchronous communication but lack built-in ML algorithms, while graph-based architectures support asynchronous communication and data and model parallelism.Source: A. Lessons Learned · pp. 25–26

B. Open Issues •

  • Open issues include extending zero- or few-shot learning and meta-learning to wireless networks with limited CSI and data, and adapting models across non-i.i.d. BS domains and time-varying device data.Zero-shot learning distinguishes features using training data with relevant features, while meta-learning can learn and design algorithms with less human intervention –. Domain adaptation maps non-i.i.d. source and target data into a shared feature space to transfer objectives, but capturing massive time-varying data across devices remains challenging [211].Source: B. Open Issues • · p. 26
  • DML wireless networks still need anomaly detection for rapidly varying traffic and CPU loads, and explainability methods that account for distributed data provenance and preprocessing.Data partitioning, local cleaning, and value imputation can strongly affect the resulting model, especially in distributed settings; linking provenance with model explainability remains challenging.Source: B. Open Issues • · p. 26
  • Future DML architectures may combine AutoML with existing designs, while distributed DRL must decouple perception, information sharing, and training and achieve fast, stable adaptation across dynamic heterogeneous networks.IoT devices have limited computing capacity, and information exchange adds overhead. Multi-agent RL offers a game-theoretic approach for wireless resource allocation, protocol coexistence, and cooperative communications; complex cooperation can enlarge state spaces and slow convergence.Source: B. Open Issues • · pp. 26–27
  • Wireless DML needs better dataset-generation methods because artificially generated datasets may simplify practical systems and fail to generalize to real-world conditions.The survey calls for new ways to generate wireless datasets to narrow the gap between emerging DML frameworks and practical systems.Source: B. Open Issues • · p. 27

XI. CONCLUSION

  • The survey covers DML applications in power control, spectrum management, user association, and edge cloud computing, analyzes their tradeoffs, and outlines adversarial threats, countermeasures, and future directions.Its analysis considers optimality, scalability, convergence rate, computation cost, and communication overhead.Source: §XI · p. 27

Related papers

Loading 2012.01489v1…