PaperlayerView exact revision on arXiv

Source-linked AI summary

Edge-Cloud Collaborative Computing on Distributed Intelligence and Model Optimization: A Survey

Jing Liu, Yao Du, Kun Yang, Jiaqi Wu, Yan Wang, Xiping Hu, Zehua Wang, Yang Liu, Peng Sun, Azzedine Boukerche, Victor C. M. Leung

arXiv:2505.01821v5cs.DCcs.AIcs.LG

TL;DR

ECCC must reconcile computationally intensive intelligent applications with constrained edge devices, dynamic workloads, and strict performance requirements. This survey synthesizes model adaptation and optimization, resource management, security, applications, and emerging technologies within a unified edge-cloud perspective. It reports substantial compression, cost, latency, and application-performance gains while identifying persistent challenges in heterogeneity, security, benchmarking, and scalability.

  • Problem

    Constrained edge computation, limited battery life, dynamic workloads, and stringent latency and reliability requirements complicate deployment of intelligent applications.

  • Method

    The survey integrates model optimization, resource management, security, applications, and future technologies into a unified ECCC framework centered on design-time and run-time co-design.

  • Results

    Reported findings include 60-90% computational-overhead reductions from quantization and pruning, up to 55% cost reduction, 33.42× latency reduction, and practical low-latency, privacy-preserving deployments.

  • Takeaways & Limitations

    Effective ECCC systems require coordinated model efficiency and deployment orchestration across heterogeneous infrastructure, with LLM, 6G, neuromorphic, and quantum directions shaping future research.

  • Takeaways & Limitations

    Open challenges remain in decentralized security and privacy, heterogeneous hardware, standardized benchmarking, scalable performance, and hardware-software co-design.

Abstract

from arXiv · show

Edge-cloud collaborative computing (ECCC) has emerged as a pivotal paradigm for addressing the computational demands of modern intelligent applications, integrating cloud resources with edge devices to enable efficient, low-latency processing. Recent advancements in AI, particularly deep learning and large language models (LLMs), have dramatically enhanced the capabilities of these distributed systems, yet introduce significant challenges in model deployment and resource management. In this survey, we comprehensive examine the intersection of distributed intelligence and model optimization within edge-cloud environments, providing a structured tutorial on fundamental architectures, enabling technologies, and emerging applications. Additionally, we systematically analyze model optimization approaches, including compression, adaptation, and neural architecture search, alongside AI-driven resource management strategies that balance performance, energy efficiency, and latency requirements. We further explore critical aspects of privacy protection and security enhancement within ECCC systems and examines practical deployments through diverse applications, spanning autonomous driving, healthcare, and industrial automation. Performance analysis and benchmarking techniques are also thoroughly explored to establish evaluation standards for these complex systems. Furthermore, the review identifies critical research directions including LLMs deployment, 6G integration, neuromorphic computing, and quantum computing, offering a roadmap for addressing persistent challenges in heterogeneity management, real-time processing, and scalability. By bridging theoretical advancements and practical deployments, this survey offers researchers and practitioners a holistic perspective on leveraging AI to optimize distributed computing environments, fostering innovation in next-generation intelligent systems.

A. Background

ECCC combines edge and cloud resources to support demanding real-time intelligent applications despite device constraints, dynamic workloads, and communication requirements. Its enabling foundations span model optimization, adaptive networking, protocol selection, and resource-aware deployment.

  • ECCC integrates edge and cloud computing to address latency, bandwidth, and resource-utilization limitations in centralized and edge-only paradigms.
  • Resource-constrained devices require model compression and adaptive learning to deploy sophisticated AI models while maintaining efficiency and performance.Relevant techniques include pruning, quantization, knowledge distillation, transfer learning, federated learning, and continual learning.
  • Dynamic workloads and limited edge power motivate intelligent resource allocation, task offloading, low latency, and high reliability for applications such as autonomous driving and industrial automation.
  • SDN, NFV, NOMA, network slicing, and cross-layer optimization improve the flexibility, programmability, capacity, and coordination of edge-cloud connectivity.
  • Protocol choices span HTTP/HTTPS, MQTT, and CoAP, with lightweight IoT protocols addressing intermittent connectivity and resource constraints.

B. Related Works

Existing surveys address edge computing, cloud collaboration, reinforcement learning, optimization, and applications from different perspectives. This survey positions itself as a broader synthesis spanning the computing continuum, integrated technical domains, and diverse applications.

  • Earlier surveys primarily emphasize architectural foundations, decentralized cloud architectures, or specific aspects of edge computing rather than AI-driven model optimization.
  • Contemporary surveys contribute specialized coverage of reinforcement learning, network dynamics, model optimization, or end-edge-cloud optimization.
  • Compared with related work, this survey covers edge, cloud, edge-cloud, and collaborative frameworks across the computing continuum.
  • Its technical scope integrates compression, adaptation, NAS, resource management, privacy, security, and emerging LLM paradigms.
  • Table I compares existing surveys by computing-continuum coverage, technical focus, and applications.

C. Scope and Contributions

The survey uses a five-layer framework to connect ECCC foundations, technical implementation, protection, deployment, and research vision. It contributes unified architectural, technical, application, and roadmap perspectives while recognizing optimization and scalability constraints.

  • Framework organization: The technical implementation layer covers model optimization and resource management, while later layers address security, applications, performance analysis, and future directions.
  • Technical integration: The survey systematically analyzes architectures, resource management, performance metrics, compression, adaptation, resource-aware optimization, and security mechanisms.
  • Framework organization: The survey organizes ECCC into five layers: overview and architecture, technical implementation, protection, practical deployment, and research vision.
  • Technical constraints: The survey identifies multi-objective complexity, uncertain system behavior, and large-scale scalability as limitations of AI-driven resource management.
  • Contributions: It claims to be the first comprehensive survey targeting distributed intelligence and model optimization within edge-cloud collaborative computing.
  • Contributions: It presents a unified architectural framework for distributing computational resources across the edge-cloud continuum.
  • Contributions: Its application analysis spans autonomous driving, smart healthcare, industrial automation, and smart cities, identifying domain-specific requirements and challenges.
  • Contributions: The survey provides a research roadmap covering advanced AI techniques, system improvements, and emerging applications.

D. Research Methodology and Organization

The survey follows a structured organization of ECCC research and situates the field within publication and citation trends from 2021 to 2025. Its sections progress from fundamentals through optimization, resource management, security, applications, and future research.

  • Figure 2 analyzes publication trends and citation patterns across distributed computing paradigms from 2021 to 2025.
  • Edge-cloud computing shows upward trajectories in both publications and citations, while edge-cloud collaboration exhibits exponential citation growth peaking in 2024.
  • The survey proceeds from edge-cloud fundamentals and distributed intelligence to model optimization, resource management, privacy and security, applications, performance analysis, and future directions.

II. ARCHITECTURE AND FUNDAMENTALS

ECCC combines cloud, edge, and end-device resources to distribute intelligence across heterogeneous environments. The survey organizes supporting techniques around distributed architectures, model optimization, adaptation, and resource-aware execution.

  • Architecture: A three-tier hierarchy places large-scale processing and training in the cloud, localized processing at edge nodes, and data generation and consumption on end devices.This arrangement positions computation and storage closer to data sources while retaining cloud scalability.
  • Distributed Intelligence: Distributed intelligence uses task, data, and model parallelism to coordinate computation, decentralized data, and partitioned models across the edge-cloud continuum.Assignments account for latency, computational complexity, data locality, device capabilities, network latency, and energy constraints.
  • Constraints: The main architecture challenge is balancing edge devices’ limited computation and low latency against cloud resources’ abundant computation and communication delays.Variable network conditions, battery constraints, and privacy requirements further require adaptive partitioning, workload distribution, and federated learning.
  • Model Optimization: Model optimization is organized around compression, adaptation, and resource-aware strategies that improve execution efficiency across distributed settings.Compression reduces model size and complexity; adaptation addresses new data or environments; resource-aware optimization responds dynamically to heterogeneous constraints.
  • Enabling Techniques: Compression methods reduce inference costs, adaptation methods address changing or distributed data, and resource-aware optimization selects, splits, or allocates models at runtime.Examples include pruning and quantization, transfer/federated/continual learning, dynamic model selection, model splitting, and reinforcement learning-based resource allocation.

D. Neural Architecture Search

Neural architecture search automates the design of efficient neural networks for ECCC constraints. The survey classifies NAS by search strategy.

  • D. Neural Architecture Search: NAS targets efficient neural networks tailored to the computational constraints of edge-cloud collaborative computing.The survey presents NAS as a critical design task for ECCC model deployment.
  • D. Neural Architecture Search: The survey categorizes NAS into reinforcement learning-based, evolutionary algorithm-based, and other search paradigms according to their underlying strategies.Reinforcement learning uses environmental feedback, while evolutionary search mimics natural selection to evolve architectures.

E. Lessons Learned on Model Optimization

Model optimization in ECCC involves balancing model efficiency against performance, while accounting for deployment conditions and system heterogeneity. The review also situates runtime resource management and task offloading as complementary mechanisms for realizing optimized models.

  • Model optimization exhibits an inherent trade-off between performance and efficiency across different strategies.
  • Compression methods reduce model complexity but often degrade accuracy, with the best balance depending on the model and task.
  • Adaptation methods address privacy and model evolution but face communication bottlenecks and sensitivity to non-IID data.
  • Neural architecture search designs efficient models automatically, but its computational overhead motivates lightweight and one-shot approaches for edge devices.
  • Resource-aware optimization must dynamically consider device capabilities, network conditions, and application requirements through model selection and partitioning.

A. Task Offloading Mechanisms

Task offloading in ECCC uses latency-aware, energy-efficient, and dependency-aware mechanisms to place and schedule workloads across edge devices, cloud servers, and supporting systems. These strategies target distinct operational objectives while improving task execution in dynamic environments.

  • Task offloading mechanisms organize workload placement around latency, energy, and dependency awareness across edge-cloud systems.
  • Latency-aware offloading reduces response time for time-sensitive applications through intelligent task-placement decisions.
  • Comparative analyses identify iGATS and DRLIS among methods achieving substantial processing-delay reductions through intelligent placement.

1) Latency-Aware Offloading:

ECCC offloading and resource management address latency, energy, dependencies, allocation, privacy, and system adaptability through specialized and increasingly data-driven strategies. The review emphasizes multi-objective coordination while noting practical challenges for reinforcement learning and privacy-preserving methods.

  • Latency-Aware Offloading: Latency-aware offloading targets response-time reduction through intelligent placement, mobility management, and scheduling under stringent latency requirements.
  • Energy-Efficient Offloading: Energy-efficient offloading jointly considers computation and communication costs, with power control and transmission scheduling improving energy efficiency.
  • Dependency-Aware Offloading: Dependency-aware offloading models interconnected subtasks as dependency graphs or DAGs to avoid idle resources and increased latency.
  • Multi-Objective Optimization: Pareto and many-objective methods balance conflicting goals including latency, energy, cost, time, and load balance.
  • Resource Allocation: Dynamic allocation and collaborative management adapt resource distribution to changing workloads, mobility, communication loads, and latency constraints.
  • Energy Efficiency Optimization: Energy-aware scheduling reports DVFO energy savings of 40%, Co-HDRL power reductions of 50-70%, and green-computing reductions of 30-45% in data-center consumption.
  • Learning-Based Optimization: Reinforcement learning improves adaptive allocation and offloading policies, but deployment is constrained by training costs, slow convergence, and environmental-model dependence.
  • Data Privacy Protection: Privacy-preserving learning combines techniques such as federated learning, differential privacy, secure aggregation, and secure multi-party computation.

A. Data Privacy Protection

Privacy protection in ECCC uses collaborative learning and cryptographic techniques to protect sensitive information without direct data sharing. Differential privacy and secure aggregation provide complementary safeguards for federated learning updates.

  • Federated learning and secure aggregation enable collaborative model training without direct sharing of participant data.
  • Differential privacy adds calibrated noise to model updates before aggregation, making sensitive information harder to infer.

1) Privacy-preserving learning methods:

Privacy-preserving learning in ECCC combines cryptographic, statistical, anonymization, and access-protection techniques to support collaborative AI without exposing sensitive information. These mechanisms span data protection, model security, secure exchange, and healthcare-oriented deployments.

  • Privacy-preserving learning methods: Secure aggregation and secure multi-party computation enable joint model training or function computation without revealing individual inputs or raw data.Examples include hospitals training a joint model on patient data and aggregating updates without exposing individual contributions.
  • Privacy-preserving learning methods: Differential privacy protects individuals by adding calibrated noise, while zero-concentrated differential privacy provides tighter guarantees than traditional DP in federated learning.Differential privacy limits the impact of any one person’s data on analysis outputs.
  • Privacy-preserving learning methods: Identity protection methods such as StyleID anonymize faces through identity disentanglement while preserving other image characteristics.StyleID projects images into a GAN latent space and manipulates selected features to protect identities.
  • Privacy-preserving learning methods: Model security frameworks address unauthorized modification, adversarial attacks, poisoning, membership inference, and model extraction through layered integrity and detection mechanisms.Protection includes cryptographic hashing, trusted execution environments, adversarial training, and anomaly detection.
  • Privacy-preserving learning methods: Secure data exchange combines encryption, authentication, and access control to protect communication across heterogeneous edge-cloud infrastructures.Common components include symmetric encryption, asymmetric key exchange, TLS/VPN channels, multi-factor authentication, and role-based access control.

A. Evaluation Metrics

ECCC evaluation requires multidimensional metrics that reflect distributed resources, heterogeneous hardware, dynamic networks, user-facing service quality, and trust. The survey organizes assessment around system performance, model efficiency, quality of service, and security and privacy while highlighting deployment constraints and energy trade-offs.

  • A. Evaluation Metrics: ECCC evaluation spans system performance, model efficiency, quality of service, and security and privacy.The dimensions cover latency, inference time, availability, and data breach frequency, among other measures.
  • A. Evaluation Metrics: System performance assessment includes throughput, latency, scalability, cross-tier response time, and resource utilization across edge-cloud components.Cross-tier latency includes network transmission, edge processing, and cloud computation overhead.
  • A. Evaluation Metrics: Model efficiency is evaluated through computational cost, memory usage, and inference time to inform model selection and deployment.Computational cost is typically measured in FLOPS, with higher values indicating greater processing requirements and potentially higher energy consumption.
  • A. Evaluation Metrics: QoS metrics such as availability, reliability, and user satisfaction measure whether ECCC services remain dependable and meet user expectations.Availability measures operational accessibility, while reliability concerns consistent delivery of expected service.
  • A. Evaluation Metrics: Heterogeneity management is essential because differing hardware, operating systems, bandwidths, data formats, and protocols create interoperability and scheduling challenges.Effective integration requires data transformation and device-specific scheduling across constrained edge nodes and cloud servers.
  • A. Evaluation Metrics: Real-world deployments face communication overhead, constrained edge resources, integration complexity, and economic barriers that complicate performance evaluation and scaling.Reported deployment challenges include 40-60% timeline overruns, 30-50% bandwidth-cost reductions, and 3-5 year payback periods.
  • A. Evaluation Metrics: Energy-aware scheduling can reduce consumption through dynamic voltage and frequency scaling, task offloading, sleep-wake control, and renewable-energy-based load balancing.DVFO achieves up to 40% energy savings during DNN inference, while Co-HDRL reports 50-70% power savings during low-utilization IoT periods.

V. BENCHMARKING METHODS

Benchmarking ECCC systems combines standardized comparison frameworks with simulation, emulation, and real-world testing. Comprehensive evaluation must cover multiple metric categories, representative workloads, controlled environments, and analysis methods while balancing realism, cost, and reproducibility.

  • V. BENCHMARKING METHODS: Performance comparison frameworks standardize assessment of heterogeneous ECCC architectures and optimization techniques while helping identify bottlenecks and best practices.Frameworks must capture interactions among edge devices, networks, and cloud resources.
  • V. BENCHMARKING METHODS: A comprehensive benchmark should include system performance, model efficiency, quality of service, and security and privacy metrics.The proposed metric set spans throughput, latency, scalability, model-level efficiency, service quality, and trust-related properties.
  • V. BENCHMARKING METHODS: Benchmark designs should use standardized datasets and real-world workloads, controlled testing environments, and statistical or visualization tools for reproducible analysis.These elements support bottleneck identification and interpretation of design trade-offs.
  • V. BENCHMARKING METHODS: Simulation offers cost-effective and flexible exploration of hardware, network, and workload configurations without physical deployment.Its main role is scenario exploration across varied system configurations.
  • V. BENCHMARKING METHODS: Emulation runs real software on virtualized infrastructure and reproduces network properties such as latency, bandwidth, and packet loss in controlled conditions.Tools such as NetEm and eBPF-based solutions support realistic network modeling.
  • V. BENCHMARKING METHODS: Real-world testing provides the most accurate assessment by capturing full system interactions but is expensive, time-consuming, and difficult to control or reproduce.It is essential for validating simulation and emulation findings and demonstrating practical deployment.

VI. SYSTEM DESIGN EVALUATION

System design evaluation compares ECCC architectures and optimization techniques across application scenarios while exposing trade-offs among scalability, latency, energy, accuracy, security, and usability. The survey links these evaluations to resource-aware design, security mechanisms, and future co-design priorities.

  • VI. SYSTEM DESIGN EVALUATION: System design evaluation compares architectural approaches and optimization techniques across diverse application scenarios.The analysis considers strengths, weaknesses, load balancing, resource allocation, and task scheduling.
  • VI. SYSTEM DESIGN EVALUATION: Cloud-centric architectures provide scalability but can add latency, whereas edge-centric architectures reduce latency while facing computational and synchronization limits.These alternatives expose the central trade-off between centralized capacity and local responsiveness.
  • VI. SYSTEM DESIGN EVALUATION: Hybrid architectures combine edge and cloud resources through predictive offloading, multi-agent resource management, and unified orchestration platforms.Examples include AppealNet, EdgeMatrix, and ACE, while federated learning supports collaborative training with privacy preservation.
  • VI. SYSTEM DESIGN EVALUATION: Load balancing, dynamic resource allocation, and task scheduling improve utilization and performance by coordinating workloads across edge nodes and cloud resources.Reinforcement learning is identified as a promising approach for adaptive resource management, while scheduling can account for latency, energy, and dependencies.
  • VI. SYSTEM DESIGN EVALUATION: Offloading intensive tasks to the edge can reduce latency but strain energy resources, whereas cloud offloading conserves edge energy while introducing network latency and bandwidth bottlenecks.Optimal design therefore requires resource allocation and task-offloading strategies that navigate the local-versus-central trade-off.
  • VI. SYSTEM DESIGN EVALUATION: ECCC design involves trade-offs between latency and accuracy and between security and usability because complex models and protections can increase computation, delay, or interaction overhead.Compression methods mitigate the latency-accuracy trade-off, while encryption and authentication can affect performance and login time.
  • VI. SYSTEM DESIGN EVALUATION: Design-time model optimization is presented as a critical enabler of run-time resource management in distributed intelligent systems.The survey connects model optimization, resource management, and security into a unified framework for resource-aware system design.
  • VI. SYSTEM DESIGN EVALUATION: Persistent challenges include standardized benchmarking, decentralized security, heterogeneous hardware, and scalable hardware-software co-design.The roadmap prioritizes communication protocols and federated learning, followed by LLMs, neuromorphic edge applications, 6G, quantum edge computing, and autonomous agents.

B. Key Technical Contributions and Findings

The survey identifies advances and trade-offs across model optimization, resource management, privacy, security, and practical ECCC deployments. It also highlights unresolved bottlenecks and proposes a staged roadmap spanning near-term standardization and benchmarking to longer-term 6G, quantum, and autonomous-agent integration.

  • Model Optimization: Federated learning and neural architecture search are foundational distributed-intelligence techniques, while quantization and pruning reduce computational overhead by 60-90% with acceptable accuracy.Non-IID federated data still creates convergence difficulties, and accuracy-compression trade-offs remain unresolved.
  • AI-Driven Resource Management: Reinforcement learning-based resource management shows promise, with DRLIS achieving up to 55% cost reduction and Zhai et al. reporting a 33.42× latency reduction.Multi-objective and Pareto optimization balance conflicting latency, energy, and application requirements, but complexity and scalability remain bottlenecks.
  • Privacy and Security: Privacy and security methods support regulatory compliance and include security frameworks, attack detection, blockchain-based trust, quantum-resistant cryptography, and zero-knowledge proofs.Privacy-preserving techniques often introduce 15-30% computational overhead in healthcare and financial applications.
  • Practical Deployments: ECCC deployments support sub-100ms perception processing in Level 4+ autonomous driving, 95%+ diagnostic accuracy with HIPAA compliance, and 20-40% lower industrial equipment downtime.These examples span autonomous driving, healthcare, and industrial automation.
  • Research Gaps: The survey identifies edge LLM inference, neuromorphic computing, quantum edge computing, and embodied AI as underexplored or frontier directions requiring substantial research investment.The stated challenges include model optimization, system integration, and performance bottlenecks.
  • Future Roadmap: A three-tier roadmap prioritizes communication standardization, non-IID federated learning, and benchmarking before progressing toward LLM integration, neuromorphic applications, quantum-resistant security, 6G, quantum edge computing, and autonomous AI agents.The roadmap assigns short-term, medium-term, and long-term horizons of 1-2, 3-5, and 5-10 years.
  • Priority Research Areas: Immediate priorities include AI-based resource management, ECCC-integrated NAS, benchmarking suites, energy efficiency, sustainability, and security and privacy enhancement.The survey frames these directions as routes toward more efficient, sustainable, and intelligent ECCC systems.
Loading 2505.01821v5…