Pervasive AI for IoT applications: A Survey on Resource-efficient Distributed Artificial Intelligence

Emna Baccour; Naram Mhaisen; Alaa Awad Abdellatif; Aiman Erbad; Amr Mohamed; Mounir Hamdi; Mohsen Guizani

Source-linked AI summary

Pervasive AI for IoT applications: A Survey on Resource-efficient Distributed Artificial Intelligence

Emna Baccour, Naram Mhaisen, Alaa Awad Abdellatif, Aiman Erbad, Amr Mohamed, Mounir Hamdi, Mohsen Guizani

arXiv:2105.01798v2May 4, 2021cs.DCcs.AI

TL;DR

Pervasive AI must coordinate limited IoT, edge, and cloud resources while meeting demanding application constraints. This survey synthesizes communication-efficient approaches for distributed inference, training, and online learning, highlighting resource trade-offs and open challenges.

Problem
Resource-limited pervasive AI systems need coordinated computation and communication, but dynamic, multi-request inference orchestration remains insufficiently studied.
Method
The paper surveys pervasive AI architectures, applications, metrics, and communication-efficient techniques for distributed inference, training, and online learning.
Results
The surveyed literature spans resource, latency, bandwidth, energy, computation, memory, and throughput trade-offs across distributed pervasive AI techniques.
Takeaways & Limitations
Pervasive AI offers a framework for supporting latency-sensitive IoT applications through distributed resource allocation and communication-aware computation.
Takeaways & Limitations
Communication-aware reinforcement learning for resource-constrained IoT environments remains scarce, with most testing conducted in artificial environments or video games.

Abstract

from arXiv · show

Artificial intelligence (AI) has witnessed a substantial breakthrough in a variety of Internet of Things (IoT) applications and services, spanning from recommendation systems to robotics control and military surveillance. This is driven by the easier access to sensory data and the enormous scale of pervasive/ubiquitous devices that generate zettabytes (ZB) of real-time data streams. Designing accurate models using such data streams, to predict future insights and revolutionize the decision-taking process, inaugurates pervasive systems as a worthy paradigm for a better quality-of-life. The confluence of pervasive computing and artificial intelligence, Pervasive AI, expanded the role of ubiquitous IoT systems from mainly data collection to executing distributed computations with a promising alternative to centralized learning, presenting various challenges. In this context, a wise cooperation and resource scheduling should be envisaged among IoT devices (e.g., smartphones, smart vehicles) and infrastructure (e.g. edge nodes, and base stations) to avoid communication and computation overheads and ensure maximum performance. In this paper, we conduct a comprehensive survey of the recent techniques developed to overcome these resource challenges in pervasive AI systems. Specifically, we first present an overview of the pervasive computing, its architecture, and its intersection with artificial intelligence. We then review the background, applications and performance metrics of AI, particularly Deep Learning (DL) and online learning, running in a ubiquitous system. Next, we provide a deep literature review of communication-efficient techniques, from both algorithmic and system perspectives, of distributed inference, training and online learning tasks across the combination of IoT devices, edge devices and cloud servers. Finally, we discuss our future vision and research challenges.

I. INTRODUCTION … V. Pervasive training

Pervasive AI combines pervasive computing with artificial intelligence to distribute learning and inference across resource-constrained devices, addressing the communication, computation, privacy, latency, and scalability limits of centralized cloud analytics. The survey reviews pervasive-computing and AI fundamentals, distributed training and inference, federated learning applications, comparisons, and future challenges.

I. INTRODUCTION: Centralized cloud analytics faces intensive training and inference demands, WAN pressure from bandwidth-intensive data, privacy risks, and scalability challenges.Edge computing reduces remote data exchange, but many pervasive devices still have limited power and memory.
I. INTRODUCTION: Federated learning keeps raw data at source entities while participants train local models and send results to a fusion server for global aggregation.Distributed inference can likewise use nearby ubiquitous devices to reduce cloud-transmission latency, while intensive tasks may still require cloud servers.
A. Our scope: Pervasive AI distributes learning and inference across ubiquitous participants to address constraints in memory, computation, bandwidth, energy, and communication overhead.The survey also considers real-time inference loads, data-collection pace, and environmental dynamics when selecting participants and designing parallelization strategies.
V. Pervasive training: For training and inference, the survey profiles communication and computation models, reviews state-of-the-art methods, compares existing works, and reports lessons learned and use cases.These reviews target resource-efficient distributed learning across pervasive systems.
B. Contributions and structure of the paper: The survey presents pervasive-computing architecture and participants, AI foundations and metrics, and IoT applications for pervasive AI.It covers deep learning and reinforcement learning alongside frameworks supporting AI tasks.
II. Fundamentals of pervasive computing: The paper introduces pervasive-computing fundamentals and architecture before presenting AI fundamentals, related surveys, and the scope of federated-learning studies.The supplied organization identifies Section II as covering pervasive-computing fundamentals and architecture.
B. Contributions and structure of the paper: The survey concludes with a future vision and open challenges intended to motivate further research on pervasive AI.It identifies resource management and communication overheads as major bottlenecks for pervasive AI.
V. Pervasive training: Federated learning is investigated in different domains, including UAV swarms for cooperative target recognition as a case study.The paper’s organization places this discussion in the section on pervasive training.

VII. Privacy of pervasive AI systems … 1) Architecture

Pervasive computing embeds computational capabilities across heterogeneous devices and organizes them into a three-layer architecture for collecting, managing, analyzing, and applying data. Its infrastructure spans cloud, edge, cloudlet, fog, and IoT participants, with edge proximity reducing latency and transmission demands while supporting resource-aware AI services.

A. Definition: Pervasive computing embeds computation in devices that communicate and perform tasks while minimizing battery, memory, and CPU consumption.It can operate on devices ranging from resource-constrained nodes to highly performant servers, at any place and time.
B. Ubiquitous participants: Pervasive systems require scalable infrastructure to discover heterogeneous components, manage their interactions, interpret context, and rapidly adapt to new software and interfaces.The system must accommodate substantial differences among device capabilities and deployment contexts.
1) Data center and cloud servers: Cloud provides on-demand storage, management, computation, AI, and other services, allowing users and organizations to use provider infrastructure instead of owning servers.Cloud services may follow pay-as-you-go or subscription-based pricing models.
2) Mobile Edge Computing (MEC) servers: Edge computing brings cloud capabilities closer to users to reduce perceived latency, limit data transmission, and relieve cloud congestion.Mobile Edge Computing facilities equip cellular-network edges with services and computing capabilities.
3) Cloudlet devices: Cloudlets form the middle layer between mobile devices and cloud data centers, supporting low-latency edge–cloud task offloading.They connect cloud computing to mobile computing through components such as computer clusters.
4) Fog devices: Fog computing provides cloud-like storage, networking, computation, and AI services for geographically specific IoT applications requiring real-time responses.Examples include live streaming, interactive applications, online collective gaming, phones, and wearable health monitoring.
5) Edge devices: IoT devices are ubiquitous gadgets with processing, sensing, and software capabilities, whereas edge nodes include fog nodes, MEC servers, and cloudlets with higher-level resources.The distinction between end devices and edge nodes remains ambiguous in much of the literature, while future systems are expected to become more heterogeneous.
1) Architecture: The pervasive architecture has data source, data management, and application layers: sources generate information, management integrates and analyzes it, and applications deliver intelligent services.The paper focuses on AI-based data analytics in the data management layer, while examples of applications include health advisors and smart homes.

D. Intersection with AI … 2) Reinforcement Learning (RL)

The survey frames Pervasive AI as distributing training and inference across resource-constrained edge environments, then introduces deep-learning architectures and adaptive reinforcement-learning methods for pervasive systems.

D. Intersection with AI: AI distribution splits training and inference into smaller parts to address edge environments’ limited resources and emerging privacy challenges.The distributed elements may include data, models, or policies.
A. Background: The survey focuses on pervasive deep and reinforcement learning within the broader hierarchy connecting artificial intelligence, machine learning, deep learning, and reinforcement learning.Machine learning includes supervised, unsupervised, and reinforcement learning, while deep learning can be supervised or unsupervised.
1) Deep learning and Deep Neural Networks: Deep neural networks comprise input, hidden, and output layers whose sequential neurons transform inputs into classifications or features evaluated by a loss function.Networks with many sequential layers are termed deep neural networks.
1) Deep learning and Deep Neural Networks: The survey presents MLP, CNN, residual, and randomly wired networks as major deep-learning structures, while noting that many other state-of-the-art models are outside scope.CNNs add convolutional and pooling layers for scalable vision processing, whereas residual networks use shortcut connections to support very deep models.
2) Reinforcement Learning (RL): Reinforcement learning updates a model or policy at each time step, enabling adaptation to changing data without assuming its distribution.This adaptability allows policies or models to respond to data drift or morphing online.
2) Reinforcement Learning (RL): Bandit learning models discrete actions that produce reward feedback, assuming actions affect only the sampled reward and not the environment’s subsequent state.The survey illustrates this with article-selection decisions whose rewards are binary clicks or non-clicks.
2) Reinforcement Learning (RL): MDP-based reinforcement learning maps states to actions for long-term rewards, discovering effective choices through trials that produce immediate gains and penalties.Unlike bandits, the selected action can affect more than the direct reward by changing the environment’s state.
2) Reinforcement Learning (RL): Deep reinforcement-learning variants discussed include policy-gradient RL, DQN, DPPO, and asynchronous advantage actor-critic methods.These variants extend reinforcement learning toward deep model-based decision-making designs.

B. Performance metrics … E. Lessons learned

Pervasive AI applications require balancing application-specific priorities such as accuracy, latency, and energy with memory, computation, communication, and privacy constraints. Frameworks and near-source processing support distributed AI across vehicles, homes, cities, and immersive systems, while deployment choices depend on application requirements and pervasive resources.

1) Latency: Latency covers preprocessing, transmission, inference or training, and postprocessing, with real-time AI applications commonly constrained to around 100 ms.Latency is influenced by DNN size, host-device computational capacity, and transmission efficiency; near-processing can improve response speed.
2) Energy efficiency: Energy efficiency is especially important for battery-limited IoT devices because deep-model training, inference, communication, and computation can consume substantial energy.Energy consumption is typically measured in nanojoules and depends primarily on DNN size and computing-device capabilities.
3) Computation and memory footprint: Pervasive AI must account for substantial computation, memory, and communication demands, including VGG 16's 512 MB memory and 154.7 G MACC requirements and AlexNet's 217 MB and 7.27 G.Communication overhead depends on the input, model-layer configuration, output size, and communication technology, while privacy risks include white-box, poisoning, and membership attacks.
C. Pervasive frameworks for AI: Public frameworks support pervasive DNN training and inference, including TensorFlow for heterogeneous distributed systems and resource-constrained devices, MXNet for parameter synchronization, and Ganglia for resource monitoring.A hardware prototype has also been designed for distributed on-device prediction, while only a few frameworks specifically target distributed learning.
D. Pervasive AI for IoT Applications: Deep learning supports pervasive applications across intelligent vehicles, robots, drones, smart homes and cities, and VR/AR, including driving assistance, navigation, appliance services, waste classification, energy systems, and parking control.Smart-city data is described as large and heterogeneous, enabling high-quality analysis.
E. Lessons learned: Application requirements determine the key performance metric: VR/AR prioritizes delay, UAVs and robots prioritize energy, and surveillance prioritizes accuracy.These priorities create trade-offs because higher accuracy can require deeper networks with greater memory, computation, and communication overheads.
E. Lessons learned: Pervasive computation places AI near data sources to provide prompt responses and avoid the communication overheads that make centralized cloud analytics unsuitable for many real-time IoT applications.The paper identifies fast analytics and real-time data collection as common requirements across many IoT applications.
E. Lessons learned: Distribution mode should follow application requirements and the pervasive environment: privacy constraints and data size favor local federated learning, while low latency and limited resources favor partitioned inference.VR/AR systems use the motion-to-photons delay as a latency metric, underscoring their sensitivity to responsiveness.

IV. RELATED SURVEYS AND PAPER NOVELTY · V. PERVASIVE TRAINING · A. Federated Learning

The paper positions itself as a holistic survey of resource-efficient pervasive AI, covering distributed training, inference, and learning across pervasive devices. It then introduces federated learning as on-device collaborative training that preserves data locality while exposing communication, computation, energy, and latency trade-offs.

IV. RELATED SURVEYS AND PAPER NOVELTY: Unlike prior surveys, this paper jointly examines pervasive AI resource and communication challenges, inference distribution, model partitioning, and multi-agent reinforcement learning.It also covers federated learning, active learning, reinforcement learning, and DNN partitioning strategies, relating them to communication and computation complexity and collaboration scheduling.
V. PERVASIVE TRAINING: Pervasive training fits models or learning policies within distributed devices and includes federated learning, multi-agent reinforcement learning, and active learning.These techniques are distributed by design because their objective is to train learning models within pervasive devices.
A. Federated Learning: Federated learning broadcasts an initial model, trains personalized models on participants’ private data, and aggregates local parameters into a global model.This design processes data at its host, addressing privacy concerns associated with transferring acquired data to a central server.
A. Federated Learning: Each federated learning round initializes the task, selects participants, updates local models, uploads parameters, and aggregates them until convergence or satisfactory accuracy.The server broadcasts the initial model, participants train using local datasets, and the updated parameters return for global aggregation.
A. Federated Learning: The main federated learning challenge is exchanging model updates, whose communication and energy overhead depends on update size, participating users, local epochs, and communication rounds.The paper identifies communication frequency, participant selection, update compression, encoding, and device operating frequency as optimization targets.
A. Federated Learning: Increasing CPU frequency reduces local-model computation latency but raises energy consumption with the square of the operating frequency.For fixed data size and accuracy, faster computation therefore creates a latency–energy trade-off.
A. Federated Learning: Reducing local iterations can increase global communication rounds, creating a contradiction between computation cost and communication cost that increases convergence latency.The paper identifies local iterations E and global rounds T as key variables for jointly optimizing energy, computation, and communication costs.

2) Resource management for Federated learning

Resource management for federated learning addresses convergence, participation, communication, and heterogeneity constraints across one-layer and edge-assisted architectures. Key approaches include adaptive aggregation, participant selection, model compression, and convergence-oriented extensions.

One-layer FL: One-layer FL research optimizes convergence under limited computation and communication, balances partial participation against accuracy and cost, and reduces communication through sparsification and compression.
One-layer FL: Adaptive aggregation adjusts global-model frequency in real time to minimize learning loss in resource-limited systems affected by highly skewed non-IID data.
One-layer FL: Selecting subsets of users mitigates straggler-driven delays because including all participants can force the server to wait for poor wireless channels or large computation delays.
One-layer FL: Sparse-update and compression methods reduce exchanged-model entropy by transmitting only large-magnitude gradient entries or compressed updates instead of dense model updates.
Edge-assisted FL: Edge-assisted FL extends FedAvg with analytical convergence guarantees, probabilistic user selection to mitigate stragglers, and strategies addressing training bias from non-IID data.

3) Use case: Learning in the sky · B. Multi-agent reinforcement learning

The paper surveys resource-efficient learning in the sky, focusing on federated learning for UAV swarms and multi-agent reinforcement learning for collaborative, resource-limited pervasive systems. It emphasizes communication overhead, resource-management trade-offs, and taxonomies of distributed learning techniques.

3) Use case: Learning in the sky: In UAV swarms, followers train local models and send updates to a leader UAV, which aggregates them into a global model for subsequent inference.The framework supports tasks including trajectory planning and cooperative target recognition.
3) Use case: Learning in the sky: Federated learning enables collaborators to jointly train deep models using local data without revealing that data to a centralized server.It is presented as a response to the difficulty of transferring massive IoT-generated datasets to centralized servers.
3) Use case: Learning in the sky: Federated-learning research emphasizes aggregation mechanisms to address communication efficiency, system and model performance, reliability, heterogeneity, security, and scalability.One-layer approaches dominate prior work, while decentralized strategies are receiving increasing attention.
3) Use case: Learning in the sky: Full model updates can reach gigabytes and repeated communication rounds create a major federated-learning communication burden.The burden is especially problematic for densely connected deep-learning models and large datasets requiring many rounds for convergence.
3) Use case: Learning in the sky: Optimizing performance and wireless-resource usage for edge-assisted federated learning remains unresolved, with slow convergence and added difficulty in dynamic or resource-constrained networks.The cited challenges include vehicular networks and healthcare systems.
B. Multi-agent reinforcement learning: Multi-agent reinforcement learning studies collaborative agents that learn from past interactions to jointly optimize an objective while minimizing resource utilization.The section focuses on cooperative pervasive-AI systems rather than competitive or zero-sum games.
B. Multi-agent reinforcement learning: The MARL review covers distributed multi-arm bandit and Markov decision process algorithms and proposes a taxonomy organized by performance under resource budgets such as communication rounds.Collaboration is essential because pervasive systems contain numerous resource-limited devices whose collective experience can be leveraged.

1) Multi-agent multi-arm bandit learning

Multi-agent bandit learning extends single-agent exploration by enabling agents to collaborate when identifying high-reward actions. Distributed bandits address a shared instance, whereas federated bandits combine experience across related but distinct instances while managing communication overhead.

Overview: Bandit learning selects among actions with unknown reward distributions, using regret to measure rewards lost relative to the best expected action.Algorithms estimate arm quality from sampled rewards, and solution methods typically seek sub-linear regret growth.
Distributed Bandit Formulations: Collaborative multi-agent learning can accelerate discovery of the best action, but communication must remain low enough to satisfy latency constraints.A distributed recommender system illustrates how servers can reuse one another’s experience while limiting communication overhead.
Distributed Bandit Formulations: Research on distributed bandits studies the communication–performance trade-off through regret bounds, communication bits, network topology, and communication rounds.Reported approaches include logarithmic communication requirements and protocols that disseminate the best-performing arm across agents.
Distributed and Federated Bandit Formulations: Distributed bandits coordinate agents solving the same bandit instance, while federated bandits coordinate agents facing related but different instances.Distributed agents share the same action set and reward distributions; federated agents use one another’s experiences to identify a global action across distinct local instances.
Federated Bandit Formulations: Federated bandits require communication across related local environments so agents can agree on a globally best arm.One formulation has agents play the global arm at the beginning of each round after communicating, unlike distributed bandits’ shared-instance setting.

2) Multi-agent Markov decision process learning

This section surveys multi-agent Markov decision process learning for pervasive computing, emphasizing communication–performance trade-offs and systems-oriented design. It covers MARL formulations, policy and value-based methods, CTDE, learned communication, and the scarcity of resource-aware evaluation in IoT environments.

Overview: The survey classifies pervasive-computing MARL research by how it manages communication–performance trade-offs rather than by the technical merits of learning algorithms.Its systems perspective focuses on communication aspects across the considered works.
Overview: MARL models pervasive multi-agent decision-making as a partially observable Markov game in which agents use policies to maximize individual or total expected rewards.The formulation includes agents, states, actions, observations, transitions, rewards, and a discount factor; differing rewards yield mixed decentralized POMDPs, while shared rewards yield collaborative POMGs.
Overview: Policy-gradient and value-based methods are the main solution approaches, with actor–critic methods using a state-value function as a baseline for policy-gradient optimization.The survey notes that methods tune these approaches according to the application.
Centralized training and Decentralized Execution (CTDE): CTDE eliminates or limits deployment communication by training centrally and executing decentralized policies, but this can reduce adaptability when the environment changes and renewed coordination is needed.Value factorization methods such as VDN, QMIX, and QTRAN discard the global function at execution, while CTDE actor–critic methods discard the centralized critic and retain decentralized actors.
Learned communication: Learned-communication MARL trains agents to decide what messages to send and how to interpret them over limited-bandwidth channels, but communication-aware reinforcement learning for resource-constrained IoT remains scarce.Existing work includes communication scheduling, message compression, and learned gating, while evaluations commonly use artificial environments such as Particle environments or StarCraft II.

C. Active Learning (AL)

Active learning (AL) progressively selects informative data to update both the training dataset and model, supporting continuous learning in pervasive systems. Its applications address noisy or limited labels in crowdsourcing, face identification, and connected vehicles through selective data acquisition and integration.

Overview: AL actively selects informative data over time, progressively updating the training dataset and learning model to improve learning quality.The training dataset is therefore dynamic rather than static.
Overview: AL must balance exchanged data type and volume against performance improvements and communication overhead among contiguous nodes.Nodes may exchange labels, features, or samples.
Overview: Unlike federated learning, AL and FL address orthogonal problems, with FL coordinating cooperative nodes and a central server to produce a global model.The distinction is framed around their different coordination and learning objectives.
Applications of AL: AL applications target settings where accurate ground-truth labeling is difficult, including crowdsourcing and automated vehicles with weak or noisy labelers.Crowdsourced labels can be affected by competence differences, biases, disingenuous behavior, low resolution, and stale information.
Applications of AL: For incremental face identification, AL progressively selects and labels informative samples using an active self-paced strategy before adding them to the training set.The strategy adds data from easier, higher-confidence samples toward more complex data.
Use case: AL for Connected Vehicles: In connected vehicles, AL operates over time-varying neighbor networks by combining offline learning, online labeling, and label integration for newly acquired information.Vehicles exchange labels, features, or samples with neighboring vehicles within communication range.

VI. PERVASIVE INFERENCE · A. Profiling computation and communication models · 1) Computation models

The paper frames pervasive inference as distributing trained-model segments across ubiquitous devices, then develops computation and communication models to analyze DNN partitioning, latency, and energy. It distinguishes binary and partial offloading, data and model parallelization, and per-layer versus per-segment splitting strategies.

VI. PERVASIVE INFERENCE: Pervasive inference partitions a trained model across ubiquitous devices, addressing challenges of centralized approaches through multiple distribution scenarios.Training the distributed model is outside this section’s scope.
A. Profiling computation and communication models: Computation and communication models formulate DNN operations for optimization by capturing task requirements, wireless latency, and participant energy consumption.
1) Computation models: Offloading is binary when indivisible tasks execute locally or remotely, whereas partial offloading divides DNN classification into segments executed across source and pervasive participants.
1) Computation models: Data parallelization independently distributes duplicated segments across participants, while model parallelization splits one task across devices and merges outputs under task dependencies.Model parallelization requires accounting for execution order because dependent tasks affect inference execution.
1) Computation models: DNN task-call graphs use DAGs with vertices for network segments and edges for dependencies, enabling per-layer or per-segment partitioning.Per-layer splitting allocates layer sets to participants, whereas per-segment splitting divides operations such as feature-map transformations, multiplication tasks, or neurons.
1) Computation models: Computation capacity, latency, memory, and energy constrain partitioning: higher emax can reduce computation latency but increases energy concerns for end devices and edge nodes.Offloaded inference energy includes data transfer, remote computation, and waiting or receiving costs.
1) Computation models: Per-layer splitting follows sequential layer dependencies and supports resource-constrained devices, but it does not enable model parallelism or latency minimization.
1) Computation models: Per-segment splitting enables model parallelism through output or input partitioning for fully connected layers and channel, spatial, or filter splitting for convolutional layers.These strategies distribute computations and, in some cases, filter weights or input channels, then merge outputs through concatenation or summation.

2) Communication models

Communication models for pervasive inference emphasize minimizing latency, data transmission, and energy through efficient DNN partitioning and resource-aware offloading. The review shows that partitioning and parallelism choices must match layer dependencies, network structure, and participating-device capabilities.

Overview: Communication latency and data transmission are central objectives in pervasive inference, motivating efficient DNN splitting and offloading across devices or remote servers when local resources are insufficient.
Overview: Communication cost models account for transmission latency across sequential, parallel, and general layer dependencies, as well as energy consumed during task offloading and result reception.
Partitioning strategies: Per-layer splitting uses simpler dependencies but can impose higher per-device transmission, while per-segment splitting trades greater total communication for lower computation and memory footprints per device.
Lessons learned: Model parallelism consistently achieves lower latency than data parallelism because it computes multiple subtasks simultaneously, whereas data parallelism incurs input-merging and transmission costs.
Lessons learned: The choice between parallelism modes depends on the partitioning strategy and dependency structure, with parallel or general dependencies enabling concurrent-segment distribution.
Lessons learned: Optimal partitioning depends on DNN properties such as channel and filter sizes, feature-map dimensions, neuron counts, and participant capacities; fully connected layers particularly depend on input and output sizes.

B. Resource management for distributed inference

Resource management for distributed inference focuses on jointly managing computational and transmission resources to reduce inference latency and energy consumption across pervasive systems.

B. Resource management for distributed inference: Joint computational and transmission resource management is central to achieving low inference latency and efficient energy consumption.The section reviews resource management for deep inference distribution and segment allocation in pervasive systems.

1) Remote collaboration · 2) Localized collaboration

The survey organizes collaborative DNN inference into remote and localized strategies that partition computation across devices, edge nodes, and cloud servers. These approaches trade latency, accuracy, energy, communication, privacy, and resource constraints through split selection, compression, early exits, and heterogeneous partitioning.

1) Remote collaboration: Remote collaboration partitions DNN inference between a source device and remote edge or cloud resources using binary, per-layer, or per-segment offloading.Partial offloading is the survey’s focus, including one or multiple layer split points and per-segment distribution.
1) Remote collaboration: Remote layer splitting can minimize inference latency but remains constrained when intermediate layers generate large data, as illustrated for VGG16.The wireless network is identified as a bottleneck for cloud inference, motivating dynamic partitioning based on network conditions, layer type, latency, and power.
1) Remote collaboration: Early exits and intermediate-feature compression reduce transmitted data and accelerate partitioned inference, with accuracy traded against communication savings.BranchyNet exits locally when accuracy is satisfactory, while encoding, quantization, and retraining address intermediate-data overhead.
1) Remote collaboration: Multiple cuts and hierarchical IoT-edge-cloud distribution balance transmission latency against computation delay while combining localized shallow inference with remote resources.These schemes can use early exits to provide fast and private inference at end or edge devices.
2) Localized collaboration: Localized collaboration distributes DNN layers or fine-grained segments across neighboring IoT devices or fog nodes, including per-layer and per-segment allocations.Per-segment methods accommodate lightweight devices such as Raspberry Pis and adapt partitioning to memory, computation, communication capability, and participant count.
2) Localized collaboration: Table VI compares distribution strategies with cloud-only, on-device-only, and edge-server-only baselines across latency, bandwidth, energy, computation, memory, throughput, and inference rate.Table VII compares distributed inference techniques introduced in the section.
2) Localized collaboration: The literature most often studies remote per-layer splitting, but localized per-segment inference increasingly enables resource-limited devices while avoiding remote-server transmission.Localized work emphasizes model parallelization and concurrent segment computation, whereas data parallelization and real-time adaptation remain less covered.
2) Localized collaboration: Split-point selection depends on device capacity, network characteristics, and DNN topology, while mixed partitioning suits heterogeneous devices and can reduce inter-device dependency and transmission.Compression is particularly useful when pooling or fully connected layers produce little variation in per-layer data size.

C. Use case: Distribution on moving robots … 1) Privacy

The paper surveys distributed AI for moving robots, emphasizing adaptive offloading trade-offs among accuracy, latency, energy, mobility, and image quality. It also identifies privacy and security threats from untrusted participants, servers, and compromised agents in pervasive AI systems.

C. Use case: Distribution on moving robots: Distributed inference across UAVs and MEC servers trades accuracy against latency and transmitted data, especially when poor image quality requires remote DNN layers.Low-level layers run on the UAV, while higher-level layers are assigned to remote servers; mobility affects transmission rate and target distance affects image quality.
C. Use case: Distribution on moving robots: UAV proximity to targets can improve image resolution and on-board accuracy while avoiding offloading, but hidden surveillance and harsh environments may prevent close positioning.The system therefore faces an accuracy–latency trade-off shaped by UAV position.
C. Use case: Distribution on moving robots: Reducing the processing coefficient saves battery life but increases inference latency, while selecting smaller intermediate representations can encourage offloading through reinforcement learning.The cited approach uses an online reinforcement-learning solution for the offloading trade-off.
C. Use case: Distribution on moving robots: When idle, a robot’s DNN computation and offloading increase device power consumption by 50%, while movement causes high consumption spikes that can limit performance.Moving-device variability also triggers frequent changes in CPU power-saving settings.
C. Use case: Distribution on moving robots: For ground robots, distributing DNN inference reduces energy consumed per device but increases total power consumption, motivating smaller model segments across multiple low-power robots.This strategy seeks an equilibrium between energy use and the number of executed tasks, although it depends on many highly coupled robots.
VII. PRIVACY OF PERVASIVE AI SYSTEMS: Pervasive AI introduces privacy risks because sensitive information can leak through shared intermediate data or updates, while untrusted participants may alter data or send incorrect parameters.These concerns are amplified when servers and participants are not controlled or owned by one operator.
A. Privacy for pervasive training: Federated learning faces vulnerabilities from malicious servers, insider adversaries, and outsider attackers, including honest-but-curious aggregators that inspect models.Participants may also randomly join or leave training, increasing the range of possible attack sources.
1) Privacy: Compromised agents can inject false data into bandit and MARL systems, breaking regret guarantees and convergence while potentially steering policies toward malicious objectives.The cited works characterize malicious agents as both disruptive and capable of actively influencing other agents.

2) Defense techniques and solutions … C. Lessons Learned •

Pervasive AI privacy defenses span data perturbation, encryption, blockchain, secure computation, poisoning prevention, and privacy-aware distributed inference, but they require trade-offs among privacy, accuracy, computation, and communication. Existing work remains limited for distributed-inference and multi-agent reinforcement-learning threats.

2) Defense techniques and solutions: Privacy defenses for pervasive AI include differential privacy, homomorphic encryption, blockchain-based communication, secure multi-party computation, and protection against data poisoning.These approaches respectively perturb data, compute on ciphertexts, secure update exchange, protect private inputs during joint computation, and defend against malicious training data.
2) Defense techniques and solutions: Differential privacy improves privacy by injecting noise, but increased privacy can cause significant performance loss, requiring calibration between privacy and learning performance.This trade-off was observed in federated learning on real-world healthcare datasets.
2) Defense techniques and solutions: Homomorphic encryption protects training integrity against outsider attackers and honest-but-curious servers, but fully homomorphic schemes remain inefficient for practical deep-learning use.Encrypted parameter aggregation is constrained by task degree and complexity.
B. Privacy for pervasive inference: Sensitive IoT data, including camera images, GPS coordinates, and patient vital signs, creates major security concerns when transmitted from end devices to remote servers.Edge computing reduces public-internet transfer, but additional privacy techniques are still needed against eavesdroppers.
2) Defense techniques and solutions a) Features extraction:: Feature extraction, noise addition, cryptography, and privacy-oriented model partitioning protect distributed inference, but each introduces accuracy, computation, communication, or security trade-offs.Feature sharing can enable input reconstruction, while encryption may impose prohibitive overhead on resource-constrained devices.
2) Defense techniques and solutions a) Features extraction:: Selecting a privacy-aware distributed-inference strategy depends on system requirements because security level, accuracy, computation load, and communication load must be jointly balanced.Table IX compares existing privacy-aware distribution strategies using high, medium, and low performance categories.
2) Defense techniques and solutions a) Features extraction:: DistPrivacy uses model partitioning to guarantee privacy without the additional feature-extraction, noise, or encryption tasks that overload pervasive devices.Its design differs from prior privacy-aware approaches that generally use per-layer DNN distribution.
C. Lessons Learned •: Privacy research has focused mainly on federated-learning attacks, while distributed-inference threats and multi-agent reinforcement-learning security remain underexplored amid substantial communication and computation costs.Existing studies often assume trusted agents and prioritize minimizing resource utilization.

VIII. FUTURE DIRECTIONS AND OPEN CHALLENGES … 5) Fault-tolerance of distributed inference

The paper identifies open challenges for pervasive AI spanning emerging-system deployment, trustworthy and explainable distributed intelligence, adaptive resource orchestration, privacy, robotic inference, non-sequential models, and fault tolerance. It proposes future research directions including 6G support, blockchain-based incentives, XAI, online coordination, privacy mechanisms, trajectory optimization, and dropout-inspired resilience.

VIII. FUTURE DIRECTIONS AND OPEN CHALLENGES: Pervasive AI research should address emerging-system integration, efficient distributed inference, enhanced federated learning, and multi-agent reinforcement learning.The section frames these as promising opportunities and future directions for pervasive AI systems.
1) Pervasive AI-as-a-service: 6G could support pervasive AI through greater network configurability and softwarization, but its design, implementation, and use cases remain at an early research stage.The passage calls for academia and industry to advance beyond initial theoretical conceptualization.
2) Incentive and trusty mechanism for distributed AI using blockchain: Distributed AI requires incentive mechanisms for nodes contributing memory, energy, communication, and computation, while protecting participants’ private information through approaches such as blockchain.Relevant participant information includes locations, identifiers, and capacities.
3) Explainable AI (XAI): Explainable AI is important because opaque decisions in critical applications limit trust, while distributed inference and federated learning can improve transparency by exposing model components and influential inputs.The passage identifies decomposability, feature extraction, clustering, and local-model training as mechanisms supporting interpretation.
1) Online resource orchestration: Online resource orchestration must adapt participant selection and distributed inference to volatile resources, changing request loads, heterogeneous devices, data quality, and high-dimensional DNN parameters.Existing work emphasizes model partitioning, but the passage identifies limited long-term evaluation of inference bottlenecks and gains under dynamic loads and participants.
2) Privacy-aware distributed inference: Privacy-aware distributed inference must defend shared data and trained models against untrusted participants and black-box or white-box attacks, despite limitations of noise addition and cryptography.These attacks can jeopardize original inputs, motivating stronger privacy mechanisms for heterogeneous ubiquitous devices.
3) Trajectory optimization of moving robots for latency-aware distributed inference: Moving robots such as UAVs need latency-aware distributed inference strategies that account for energy consumption and unstable communication with remote servers during real-time analysis.Applications include highway inspection, search and rescue, and border surveillance.
5) Fault-tolerance of distributed inference: Fault-tolerant distributed inference must mitigate failures across segmented DNNs without excessive retransmission bandwidth, potentially by training without skip connections and activating them only during failures.The proposed direction draws on dropout and calls for studying transmission cuts across splitting strategies.

6) Data-locality-aware algorithms … IX. CONCLUSION

The survey identifies resource, communication, computation, memory, adaptability, and networking challenges across pervasive AI, federated learning, bandit, MARL, nanotechnology, and data-locality-aware systems. It concludes by outlining future technical challenges and opportunities for improved performance, resource management, privacy, and applications.

6) Data-locality-aware algorithms: Distributed inference assigns DNN tasks according to computation and energy resources, but typically ignores memory required to store model structures such as VGG-16’s weights.The section highlights a gap in data-locality-aware scheduling: input-data memory is considered, whereas DNN-structure storage is not.
7) Pervasive inference for nanotechnology applications: Nanotechnology applications require AI-driven analysis under severe particle energy and computation limits, motivating new partitioning strategies and per-multiplication dependency studies.Existing neuron, spatial, and filter splitting can remain too complex for nanoscale particles and their numerous participants.
1) Active Federated Learning: Active federated learning should combine active learning with federated schemes to reduce communication overhead and slow convergence in dynamic, heterogeneous environments.The proposed direction exploits federated learning’s distributed nature while seeking adequately fast convergence.
2) Blending inter and intra data parallelism for federated learning: Federated learning distributes deep-model training across many CPUs and GPUs through inter-data parallelism because larger networks exceed the resources of individual devices.The passage motivates blending inter- and intra-data parallelism for pervasive-node training.
1) Demonstrated applications: Practical multi-agent bandit studies should quantify communication and energy effects, congestion, regret-bound improvements, and quality-of-experience outcomes in real applications.Wearable-device energy and edge-node congestion are identified as concrete evaluation targets.
2) More general forms of MABs: Multi-agent bandit research should extend beyond finite-action stochastic settings to adversarial, linear, pure-exploration, and non-stationary bandits while examining regret and communication use.These broader problem classes remain insufficiently studied under multi-agent settings.
3) Heterogeneity of Bandit agents: Bandit agents may differ in computational capability, causing different interaction rates, sample counts, and estimate quality; this heterogeneity remains underexplored in distributed and federated bandits.Its effects are heavily studied in supervised federated learning but not yet in the corresponding bandit settings.
4) MARL performance/communication trade-off; 5) MARL under networking constraints: MARL faces a trade-off between communication-efficient centralized training and decentralized adaptability, while delay, noise, failures, time-varying topologies, bandwidth, and multiple access remain important networking constraints.Centralized-training methods may require retraining in non-stationary environments, whereas many MARL studies omit practical network factors.