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

Wireless networks need distributed machine learning because centralized training concentrates data and incurs computation, communication, congestion, and privacy costs, while existing literature lacks a systematic wireless-focused review. This paper surveys DML techniques, applications, architectures, efficiency, privacy, security, and open issues, reporting improvements in selected federated-learning and edge-cloud settings.

  • Problem

    Existing literature lacks a systematic review of distributed machine learning techniques and their applicability to wireless communication networks, despite centralized training's computation, communication, congestion, and privacy costs.

  • Method

    The paper provides a comprehensive survey of DML techniques, architectures, platforms, wireless applications, efficiency measures, privacy protections, security countermeasures, and future research directions.

  • Results

    Across reviewed applications, the survey analyzes optimality, scalability, convergence, computation cost, and communication overhead, including CE-FedAvg stabilizing with six times fewer rounds and transmitting three times less information than FedAvg.

  • Takeaways & Limitations

    DML offers a broad framework for wireless applications including power control, spectrum management, user association, and edge cloud computing, with practical gains reported in selected studies.

  • Takeaways & Limitations

    Existing works often assume identical training and testing feature spaces, although differing spaces can cause underfitting and undesirable performance.

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

Wireless networks motivate distributed machine learning because data, devices, and computation are geographically dispersed, while centralized training creates communication, computation, congestion, and privacy costs. This survey addresses the literature gap by systematically reviewing DML techniques, applications, architectures, efficiency, security, and open problems.

  • More capable mobile processors and memory make local distributed learning practically feasible on wireless devices.
  • Explosive data growth and congested wireless channels discourage terminals from sending raw data to centralized ML servers.
  • Centralized ML requires gathering wireless data in one training system, increasing computation and communication costs, channel congestion, and privacy risks.
  • Existing surveys largely cover centralized ML or federated learning, leaving DML techniques insufficiently studied and compared for wireless systems.
  • The survey reviews DML applications in power control, spectrum management, user association, and edge cloud computing, alongside optimality, scalability, convergence, cost, privacy, and security.
  • The paper identifies unresolved issues and future research topics, including growing interest in multi-agent reinforcement learning for distributed wireless applications.

II. ALGORITHMS AND MODELS

This section introduces deep learning and reinforcement learning models used in wireless DML. It explains how neural networks learn representations and how RL agents improve decisions through state, action, and reward interactions.

  • A. Deep Learning (DL): Deep learning automatically extracts features from raw data through layered nonlinear processing, commonly using neural networks.
  • A. Deep Learning (DL): Deep neural networks approximate complex functions by composing simple nonlinear operations performed by neurons.
  • B. Reinforcement Learning (RL): Reinforcement learning repeatedly maps system states to actions and uses transition rewards to maximize a long-term objective.
  • B. Reinforcement Learning (RL): Q-learning estimates action values through trial and error without a pre-stored dataset or explicit environment model.
  • B. Reinforcement Learning (RL): Deep reinforcement learning uses neural networks to approximate value or policy functions in high-dimensional environments.

C. Stochastic Gradient Descent (SGD)

SGD updates model variables using stochastic sample gradients, while variants and federated procedures distribute optimization across devices. These methods trade convergence behavior against computation and communication requirements.

  • C. Stochastic Gradient Descent (SGD): SGD randomly samples one data point and updates the optimization variable using its loss gradient as a stochastic estimate.
  • C. Stochastic Gradient Descent (SGD): Momentum-based SGD reduces gradient fluctuations by adding a weighted sum of previous gradients and can improve convergence.
  • C. Stochastic Gradient Descent (SGD): ADMM decomposes large-scale learning into parallel subproblems and can guarantee linear convergence under certain conditions.
  • C. Stochastic Gradient Descent (SGD): Federated learning trains local models on device-held datasets and exchanges model updates rather than raw data.
  • C. Stochastic Gradient Descent (SGD): FedAvg selects participants, performs local gradient updates, and aggregates their resulting model parameters at the server.

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

Vertical federated learning combines distinct feature sets from parties observing the same users, while federated transfer learning addresses differences in both samples and features. The section also situates these approaches alongside partitioned learning for decentralized computation.

  • 2) Vertical FL (or Feature-Based FL):: Vertical FL combines distinct feature spaces collected by parties observing the same users, such as a base station and a core-network gateway.
  • 2) Vertical FL (or Feature-Based FL):: Vertical FL aggregates private features to compute shared losses and gradients without exposing each party’s specific identity and state.
  • Federated transfer learning addresses datasets differing in both samples and features by learning a mutual representation from scarce overlapping public samples.
  • FTL can transfer a model between wireless operators when users roam, extending beyond horizontal and vertical FL settings.
  • Partitioned learning divides a supervised model into parameter blocks processed by different devices, whose results return to a server for model refresh.

A. Convergence of Deep Learning

DML convergence depends on algorithmic design and assumptions: momentum, parallelization, compression, and ADMM variants target faster or more scalable optimization, but communication can limit realized gains.

  • ADMM convergence analyses often require assumptions such as differentiability, positive gradient conditions, predefined accuracy, and sufficiently small step sizes.
  • O(N^3) to O(N^2): DLADMM reduces computational complexity through quadratic approximation and backtracking while proving global convergence for an ADMM-based DNN.
  • O(1/NT) convergence gives PMSGD an N-fold linear speedup over single-node SGD under N parallel workers.
  • Linear acceleration is difficult because classic PMSGD requires simultaneous updates from all participating agents, creating substantial inter-node communication cost.
  • Decentralized SGD variants use neighbor communication, compression, and sparsification to address high delay, poor bandwidth, and transmission volume.
  • Momentum methods can improve convergence by incorporating information from previous iterations and extending decentralized training to additional scenarios.

B. Convergence of Federated Learning

Federated-learning convergence is affected by non-i.i.d. data, communication costs, and noisy or heterogeneous clients. The surveyed methods use adaptive optimization, momentum, weighting, and distributed architectures to improve training behavior.

  • Adaptive FedAvg combines decentralized Adam optimization with compression to reduce convergence rounds and communication under non-i.i.d. client datasets.
  • Six times fewer rounds and three times less information: CE-FedAvg stabilizes at the required precision relative to FedAvg in numerical tests.
  • Momentum FL incorporates previous-iteration gradients, establishes convergence-speed bounds, and experimentally shows global convergence with enhancement over FL.
  • Cross-entropy and KL divergence quantify distributional gaps used in wireless-system optimization and model training.
  • FOCUS uses baseline samples and joint cross-entropy to score local data accuracy and calibrate client weights under label noise.
  • Distributed training architectures coordinate workers and masters, while GPU frameworks can differ in performance even with identical hardware and datasets.

B. Communication

Communication efficiency is central to wireless DML because model exchange can be costly and error-prone. The surveyed approaches reduce transmitted information, exploit distributed coordination, or expose computation–communication bottlenecks.

  • Communication constraints: Wireless bandwidth scarcity and lossy links can introduce symbol errors that reduce federated-learning precision and affect algorithm convergence.
  • Communication reduction: Federated optimization keeps training parameters local and uses sparsity to reduce communication rounds and network bandwidth during model training.
  • Quantization: Similar training precision with substantially lower information exchange: hierarchical gradient quantization reduces overhead relative to sign-SGD.
  • Communication reduction: Communication-Mitigated Federated Learning filters impertinent client updates using feedback about the global model-refreshing direction.
  • Computation–communication trade-off: In decentralized GPU training, calculation and transmission behaviors jointly affect speedup, and 100 Gbps InfiniBand may not keep pace with NVIDIA Tesla V100 computation.
  • Distributed coordination: GADMM separates agents into head and tail clusters, limits sharing to same-cluster neighbors, and is proved optimal under convex loss functions.
  • MapReduce: 97% anomaly detection and 30% faster execution: decentralized K-means and PCA preserve accuracy while benefiting from increased parallelization.
  • Distributed platforms: Distributed decision-tree and energy-forecasting systems use Spark and related platforms to reduce model-building time or gain computational efficiency on large datasets.

B. Parameter Server (PS)

Parameter-server architectures separate worker computation from centralized parameter maintenance, supporting distributed data-parallel learning. Related graph-based systems extend this design toward model parallelism and report faster ML operations.

  • B. Parameter Server (PS): Parameter-server architectures use server and client nodes, with servers maintaining parameters and aggregating weights from client groups.Clients perform computations such as gradient calculation, backpropagation, and weight updates using synchronized global gradients.
  • B. Parameter Server (PS): Graph-based frameworks extend data parallelism with model parallelism by representing parameters as graph vertices and executing vertex functions concurrently.Unlike iterative MapReduce and parameter-server systems, graph-based frameworks distribute both training data and model parameters within the same cluster.
  • B. Parameter Server (PS): TuX2 combines graph and parameter-server ideas through edge-centric data exchange and vertex-centric update interfaces, but vector-to-graph translation can be complicated.Its programming model supports both edge-centric and vertex-centric operations, while vectorized ML algorithms may not map naturally onto graph vertices.
  • B. Parameter Server (PS): Cymbalo uses heterogeneous-data storage, a hybrid ML framework, and vector-centric programming to improve efficiency over other graph-processing systems.Simulations on Spark 2.1.2 report that Cymbalo expedites ML operations by 1.6 to 5.8 times compared with Spark, PowerGraph, and Angel.
  • B. Parameter Server (PS): Distributed graph processing also supports sensor-network applications by modeling sensor nodes as vertices in a graph-based framework.The cited sensor-network design uses a Boltzmann machine topology for distributed security applications.

VII. SOFTWARE PLATFORMS AND LIMITATIONS

DML software platforms distribute computation and data across clusters using systems such as Hadoop, Spark, S4, coded computation, and Petuum. Their trade-offs involve execution speed, streaming support, fault tolerance, consistency, and resource management.

  • VII. SOFTWARE PLATFORMS AND LIMITATIONS: Hadoop distributes data and computation across many hosts, while HDFS uses NameNodes and DataNodes to store and serve large datasets.HDFS organizes files and directories hierarchically, with the NameNode managing metadata and DataNodes backing file mappings.
  • VII. SOFTWARE PLATFORMS AND LIMITATIONS: Spark improves on Hadoop for many ML programs by keeping program state in memory rather than writing it to disk at every iteration.A Spark-based optimization framework can also apply the parameter-server architecture.
  • VII. SOFTWARE PLATFORMS AND LIMITATIONS: S4 provides scalable distributed stream processing for continuous unbounded data, unlike MapReduce’s batch-oriented operation.Its design uses keyed events and processing elements, with static routing and load balancing but no robust live processing-element migration.
  • VII. SOFTWARE PLATFORMS AND LIMITATIONS: Coded computation improves robustness to stragglers and communication bottlenecks by adding redundancy and recovering results from a subset of submissions.The cited encrypted framework uses erasure codes and can drop unfinished submissions.
  • VII. SOFTWARE PLATFORMS AND LIMITATIONS: Petuum supports data- and model-parallel computing through a parameter-server architecture and quantifies asynchronous inconsistency to support convergence and stability.Its scheduler lets users specify application consistency principles, while fault recovery and real-time resource allocation remain improvement areas.
  • VII. SOFTWARE PLATFORMS AND LIMITATIONS: A shared-memory DNN framework reduces image-recognition training time by 10 ∼50% compared with TensorFlow for CNN and MLP models.The framework accelerates parameter reading and updating by using remote shared memory across workers.

C. Graph-Based Platform

Graph-based platforms distribute graph models with schedulers and consistency controllers, but their correctness under asynchronous execution is not fully established. In wireless networks, distributed learning addresses computational and data-scale constraints while supporting wireless optimization schemes.

  • C. Graph-Based Platform: Graph-based platforms such as Pregel and GraphLab divide graph models while embedding scheduling and consistency control.The survey notes that analyses confirming correct ML execution under asynchronous consistency models and scheduling remain limited.
  • C. Graph-Based Platform: Large-scale wireless optimization can be computationally complex and slow to converge, while data-driven ML may exceed the computation and storage resources of mobile devices.These constraints motivate distributed approaches for wireless learning and computation.
  • C. Graph-Based Platform: The surveyed wireless applications include CFmMIMO schemes, federated learning, over-the-air aggregation, cooperative relaying, and edge learning.The passages identify power control, gradient collection, model updating, and edge-agent coordination as application settings.
  • C. Graph-Based Platform: Over-the-air federated learning represents device indices, model parameters, transmit power, learning rate, and device gradients within its communication-learning process.The cited formulation uses k for edge devices, w for model parameters, p for transmit power, γ for learning rate, and gk(t) for gradients.

B. Spectrum Management

Distributed learning methods support wireless spectrum management through federated, decentralized, and multi-agent approaches. The surveyed methods address energy, scheduling, convergence, and local decision-making under constrained wireless conditions.

  • B. Spectrum Management: Federated edge learning schedules local and server-side model learning while coordinating model synchronization among edge agents.The architecture is presented as a privacy-preserving edge-learning approach.
  • B. Spectrum Management: Bandwidth and coordination strategies can reduce mobile-device energy use by adapting to channel states and device computation capabilities.The optimized strategies assign wider spectra to devices with poorer channels or weaker computing capabilities, although generalization remains difficult.
  • B. Spectrum Management: Hierarchical federated learning uses small base stations to train with nearby users and sends renewed models to a macro base station for network-wide agreement.Gradient sparsification and periodic averaging reduce communication in the hierarchical system.
  • B. Spectrum Management: WoLF learning accelerates updates after failures and slows them after successful transmissions by using separate winning and losing learning rates.The rule requires αwin < αlose and is evaluated in a large-scale stadium-event network.
  • B. Spectrum Management: Decentralized multi-agent Q-learning lets base stations make local spectrum and resource-management decisions without acquiring one another’s explicit strategies.Each base station independently selects an ON or OFF action from its battery state and receives a reward capturing throughput and battery level.
  • B. Spectrum Management: Federated-learning efficiency depends on scheduling and SINR conditions: proportional fair performs better at high SINR targets, while round-robin is favorable at low targets.Higher SINR targets dramatically reduce convergence speed, and analysis identifies a trade-off between coordinated users and subchannel bandwidth.

2) Power allocation and latency:

Wireless DML resource allocation jointly considers learning configuration, computation, and communication constraints. The reviewed approaches address bandwidth-limited training, model partitioning, and edge-cloud data decomposition while reporting gains in convergence, latency, accuracy, or cost-effectiveness.

  • Bandwidth allocation and user selection: FL over wireless channels is affected by packet errors and limited, time-varying channel capacity, requiring subgroup user selection.Joint radio-resource assignment and user pairing are used to minimize the FL loss function.
  • Bandwidth allocation and user selection: 10% and 16%: the hybrid FL-communication framework reduces loss versus stochastic resource assignment and standard FL, respectively.The results support jointly considering learning configuration and radio-resource allocation.
  • Edge cloud computing: FEEL jointly optimizes communication and computation for energy efficiency using NOMA or TDMA transmission protocols.Each edge agent refreshes the entire model, making FEEL typically applicable to small or moderate training missions.
  • Edge cloud computing: PARTEL partitions a large-scale model into parameter blocks assigned to wireless agent clusters, whose renewals are managed to update the overall model.The framework targets numerous resource-limited workers and adaptively downloads blocks to refresh sample subsets.
  • Edge cloud computing: 46% latency and 4% accuracy improvements: integrating PABA significantly improves PARTEL on the News20 newsfiltering model.The decentralized training achieves the same convergence rate as centralized training within the same number of iterations.
  • Edge cloud computing: 24%: only this share of system connections needs evaluation while retaining asymptotic optimality in the reported dynamic data-decomposition simulation.The developed scheme also outperforms the classic scheme in data rate and cost-effectiveness.
  • Edge cloud computing: Distributed caching reduces content-demand estimation RMSE by at most 33.7% and service latency by 47.4% versus several other ML approaches.The decentralized scheme exchanges data among mobile edge nodes without releasing users’ individual data.

IX. PRIVACY AND SECURITY

DML systems face privacy risks from servers, eavesdroppers, and other participants because training updates can reveal sensitive information. The surveyed countermeasures include perturbation, dummy data, encoders, secure aggregation, and hierarchical architectures, while collection mechanisms remain an open problem.

  • Threats and privacy needs: DML privacy and security are increasingly important because data-hungry learning systems remain susceptible to attacks such as malicious servers and eavesdropping.The survey frames privacy protection as a requirement for distributed learning over large-scale datasets.
  • Threats and privacy needs: Untrusted FL servers may manipulate training or infer private agent information from uploaded parameter values and weights.Server-side model releases can also expose aggregated parameters to passive network eavesdroppers.
  • Perturbation and obfuscation: Differential privacy adds artificial noise to client parameters to conceal individual contributions, with little model-performance cost given enough participating agents.The surveyed scheme explicitly balances privacy loss against learning performance.
  • Perturbation and obfuscation: Dummy parameters or information can hide clients’ true contributions from the service provider.The cited location-aware scheme sends counterfeit locations alongside the true location so the provider cannot extract the true one.
  • Representation protection: Encoders protect data owners by transmitting encrypted high-level characteristics to a central server instead of raw user data.The encoder extracts functions from admitted data and serves as the relay to the server.
  • Aggregation and architecture: Secure aggregation remains an open collection-mechanism problem because servers must combine many client updates while selecting useful parameters or satisfying demands.The aggregation output may be a final update or a replacement model.
  • Aggregation and architecture: Hierarchical asynchronous learning offers flexible model-parameter choices and shorter computation time before convergence across layered data-owner and collection-node architectures.The architecture adapts solutions to different data-partitioning scenarios.

2) Secure Multi-Party Computation (SMC):

Secure multi-party computation and related cryptographic methods protect distributed-learning updates from inspection while preserving aggregation functionality. The surveyed evidence also covers poisoning resistance, privacy-preserving FL, blockchain verification, and Byzantine-resilient distributed learning, with security depending on threat conditions.

  • Secure Multi-Party Computation: SMC prevents servers from inspecting individual updates by having devices report an aggregate after sufficiently many encrypted and masked renewals.The protocol uses four communication rounds, ending when devices release ciphers that enable decryption of collected renewals.
  • Secure Multi-Party Computation: Additively homomorphic encryption preserves algebraic relations between plaintexts and ciphertexts, enabling cryptographically protected global model synchronization.The surveyed application describes the resulting overhead as tolerable for a typical deep-learning system.
  • Attack resistance: Distributed SVMs are less prone to data-poisoning attacks in networks with several nodes and more degrees, but robustness depends on architecture and attack intensity.The attack inserts poisoned or mislabeled samples into training data.
  • Attack resistance: A zero-sum game and distributed ADMM algorithm model the conflict between a DSVM learner and an attacker changing samples and labels.The approach targets data poisoning and other system attacks without centralizing the learning process.
  • Perturbation plus encryption: A privacy-preserving FL architecture combines a key-generation center, encrypted participant gradients, and local training over private datasets.The cloud server encrypts gradients generated by participants before processing them.
  • Perturbation plus encryption: The nested BGV and A-LWE ciphertext design lets the server decode collected values without exposing an individual agent’s privacy.Agents first perturb local gradients with decentralized Gaussian noise, then apply internal and external encryption.
  • Perturbation plus encryption: More than 0.5 collusion ratio causes a dramatic CNN accuracy drop, while the privacy budget has little influence on precision in MNIST simulations.Performance on high-dimensional datasets requires further qualification.
  • Decentralized trust and resilience: Blockchain-based FL verifies and transmits learning models to support trusted decentralization, while BYRDIE avoids sharing primitive data under adversarial conditions.BYRDIE is described as Byzantine-resilient distributed learning.

3) Blockchain-based solutions:

The survey reviews distributed learning techniques for wireless networks, emphasizing their applications, costs, privacy risks, countermeasures, and unresolved research directions.

  • Distributed learning methods: BYRDIE scales with the dimensionality of distributed learning problems and detects Byzantine failures through local scrutinizing within distributed coordinate descent.Its statistical-risk improvement assumes training on i.i.d. datasets.
  • DML frameworks: Federated learning and partitioned learning can obtain global models with little or no data exchange while reducing communication cost and protecting privacy.Partitioned learning supports model partitioning, whereas federated reinforcement learning does not require previously stored datasets for training.
  • Privacy protection: Privacy threats include data poisoning and leakage, with perturbation, dummy data, encoding, blockchain-based methods, aggregation, and secure multi-party computation proposed as countermeasures.Perturbation plus encryption is reported as effective against curious servers when compromised agents are fewer than 50%.
  • Edge computing: Edge DML must operate under constrained computing, memory, energy, communication, and latency conditions, motivating optimization of edge architectures and operations.These constraints are identified as a research direction for applying DML at the network edge.
  • Open issues: The survey identifies limited-data learning, domain adaptation, adaptive architectures, and distributed DRL as open directions for wireless networks.These directions address few-sample learning, differing feature spaces, straggler problems, and the separation of perception, information sharing, and training under constrained devices.
  • Survey scope: The survey covers DML applications in power control, spectrum management, user association, and edge cloud computing while analyzing optimality, scalability, convergence, computation cost, and communication overhead.It also discusses adversarial attacks, countermeasures, and future research topics.
Loading 2012.01489v1…