Machine Learning and Edge Computing: A

Comprehensive Survey of Optimization, Resource

Management, and Security
Survey Paper
IEEE Conference Proceedings
2025

Abstract—Edge computing has emerged as a transformative computing include autonomous driving, industrial automation,
paradigm that brings computational resources closer to data real-time healthcare monitoring, smart cities, and intelligent
sources, addressing the limitations of traditional cloud-centric transportation systems.
architectures. The integration of machine learning (ML) with
edge computing presents unprecedented opportunities for real- Recent advancements have further emphasized the role of
time intelligent applications while introducing significant chal- machine learning (ML) as a key enabler of intelligence at
lenges related to resource constraints, optimization, security, and the edge. ML techniques allow systems to adaptively learn
deployment. This survey provides a comprehensive analysis of from data, make predictions, optimize resource usage, and
recent advances in ML-enabled edge computing systems by ex-
enhance system resilience. However, the deployment of ML
amining five key research contributions. We systematically review
quantum-inspired optimization algorithms for scalable ML in models at the edge is far from trivial: resource limitations
edge environments, task clustering techniques using unsupervised in terms of computation, energy, and memory, coupled with
learning for enhanced virtual machine utilization, caching place- heterogeneous network conditions, pose significant challenges
ment strategies leveraging deep reinforcement learning, network to efficient training and inference [1]–[5].
security detection using support vector machines, and distributed
ML solutions across the edge-cloud continuum. Our analysis
To address these issues, researchers have proposed in-
reveals critical trends including the adoption of lightweight novative approaches that optimize ML performance under
algorithms for resource-constrained devices, federated learning edge constraints. For example, Goyal et al. [1] explored
for privacy preservation, and hybrid edge-cloud architectures quantum-inspired optimization algorithms (QIOA) to enhance
for balancing performance and scalability. We identify existing energy efficiency and model scalability in resource-constrained
gaps in traditional ML algorithm adaptation, energy efficiency
optimization, and real-world deployment challenges. This survey
environments. Alnoman [2] leveraged unsupervised learning
concludes with recommendations for future research directions techniques, specifically K-means clustering, for efficient task
focusing on standardized frameworks, cross-layer optimization, clustering and VM resource allocation at the edge. Zhang and
and enhanced security mechanisms for next-generation edge Jabbari [3] utilized deep reinforcement learning (DDPG) to
computing systems. optimize caching placement strategies, significantly reducing
Index Terms—Machine learning, edge computing, optimization
algorithms, resource management, caching, security, Internet
response latency. Hou and Huang [4] proposed ML-based
of Things, distributed learning, quantum-inspired algorithms, network security mechanisms using SVMs to detect anomalies
federated learning in edge networks. Finally, Marozzo et al. [5] investigated
distributed ML solutions and federated architectures spanning
I. I NTRODUCTION the edge-cloud continuum, enabling scalable and privacy-
aware learning.
A. Background and Motivation
The exponential growth of Internet of Things (IoT) ecosys- B. Scope and Objectives
tems is reshaping the computational landscape, producing
massive streams of heterogeneous data that require real-time This survey focuses on synthesizing and analyzing these
processing and intelligent decision-making at the network five representative works to capture the state-of-the-art trends
edge. With billions of interconnected devices continuously in ML-enabled edge computing. Each of these papers tack-
generating sensory information, centralized cloud infrastruc- les a distinct technical dimension—optimization, clustering,
tures alone are insufficient to meet the stringent latency, caching, security, or distributed learning—providing a diverse
bandwidth, and privacy requirements of modern applications. yet complementary view of the research landscape.
Edge computing addresses these challenges by decentral- The primary objectives of this survey are:
izing computation, bringing processing capabilities closer to 1) To systematically review and analyze five key IEEE
the data source. This approach significantly reduces end-to- research papers that advance ML integration with edge
end latency, alleviates backbone bandwidth congestion, and computing.
enhances privacy by minimizing unnecessary data transmission 2) To identify recurring methodologies, algorithms, and
to remote clouds. Typical applications that benefit from edge architectural patterns across these works.
 3) To evaluate their effectiveness in terms of energy con- and decentralized training strategies, enabling scalability, pri-
sumption, latency reduction, accuracy, and resource uti- vacy preservation, and reduced network load. This represents
lization. an important shift from viewing edge and cloud as isolated
4) To highlight limitations, open research problems, and paradigms to treating them as complementary components of
future research opportunities. a unified intelligent infrastructure.
5) To provide a structured reference for researchers and Unlike many existing surveys that focus narrowly on deep
practitioners working at the intersection of edge com- learning or on cloud-centric architectures, the studies reviewed
puting and machine learning. in this work collectively cover a diverse range of machine
learning approaches, spanning both traditional and advanced
C. Organization techniques, and demonstrate their applicability to multiple
The remainder of this paper is organized as follows. Section facets of edge computing—including optimization, resource
II provides an overview of related work and technical back- management, caching, security, and distributed learning. These
ground. Section III outlines the methodology for paper selec- contributions underscore the growing importance of heteroge-
tion and analysis. Section IV presents structured summaries of neous ML strategies in building efficient, secure, and scalable
each of the five surveyed papers. Section V analyzes results, edge intelligence systems.
trends, and performance metrics in detail. Section VI discusses
open issues and future research directions. Finally, Section VII III. M ETHODOLOGY
concludes the survey with key findings and recommendations.
This section describes the systematic and reproducible
II. R ELATED W ORK procedure used to select, analyze, and synthesize research
contributions on the integration of machine learning (ML) with
Machine learning at the edge has gained increasing momen-
edge computing. The methodology emphasizes methodologi-
tum in recent years, driven by the proliferation of IoT devices
cal consistency and traceability: every extracted claim, metric,
and the need for low-latency, energy-efficient processing close
and design detail is explicitly mapped to its original source
to the data source. While early research primarily emphasized
(Goyal et al. [1], Alnoman [2], Zhang and Jabbari [3], Hou
task offloading and basic resource optimization to bridge the
and Huang [4], Marozzo et al. [5]) to enable transparent cross-
gap between edge devices and centralized clouds, more recent
checking and reproducibility.
studies explore sophisticated ML techniques and intelligent
architectures to fully exploit the potential of edge intelligence. A. Paper Selection Criteria
A number of key research directions can be identified
from the five surveyed IEEE papers. Goyal et al. [1] pro- The initial literature search was performed on the IEEE
posed a quantum-inspired optimization algorithm (QIOA) to Xplore digital library using keywords and Boolean combi-
enhance ML training performance on edge devices while nations such as “edge computing” AND “machine learn-
significantly improving energy efficiency. This work highlights ing”, “mobile edge computing” AND “caching”, “federated
how optimization algorithms can reduce both computational learning” AND “edge”, restricted to English-language journal
overhead and power consumption without sacrificing accuracy, and conference publications (2019–2024). From the candidate
demonstrating that lightweight hybrid ML approaches can be set we applied the following inclusion/exclusion filters and
effectively deployed on devices with constrained resources. selected five representative papers that collectively cover core
Alnoman [2] addressed the problem of resource alloca- technical themes in edge ML:
tion by introducing K-means clustering for intelligent task • Relevance: The paper must explicitly address ML tech-
grouping and VM utilization. Their approach improves overall niques applied within an edge/fog/edge-cloud continuum
resource utilization and system responsiveness while maintain- context (e.g., optimization, caching, resource mgmt, se-
ing manageable computation at the edge. Similarly, Zhang and curity, distributed training).
Jabbari [3] leveraged deep reinforcement learning (DDPG) to • Diversity: We prioritized papers that represent different
optimize caching placement, achieving substantial latency re- technical directions (optimization, clustering/VM alloca-
ductions and improved network responsiveness under dynamic tion, deep-RL caching, security detection, and distributed
workload scenarios. ML architectures).
In the domain of security, Hou and Huang [4] applied • Recency: Only papers published between 2019 and 2024
Support Vector Machines (SVM) with RBF kernels for intru- were considered to capture recent advances.
sion and anomaly detection in edge computing environments. • Technical Depth and Empiricism: Preference was given
Their work emphasizes that traditional ML algorithms remain to works with formal problem statements, algorithmic
practical and effective in real-world deployments, particularly descriptions (pseudo-code or equations), and experi-
for security applications where model interpretability and low mental/implementation evidence (simulations, hardware
inference cost are critical. testbeds, datasets).
Finally, Marozzo et al. [5] explored distributed machine • Impact and Peer Review: Only peer-reviewed con-
learning and federated learning architectures in the edge-cloud ference/journal papers indexed in IEEE Xplore were
continuum. Their work emphasizes hierarchical aggregation retained.
 The five selected works are referenced throughout this TABLE I
survey as: Goyal et al. [1], Alnoman [2], Zhang and Jabbari M APPING OF THE FIVE SELECTED PAPERS TO THE ANALYTICAL
FRAMEWORK ( PRIMARY FOCUS ).
[3], Hou and Huang [4], and Marozzo et al. [5].
Paper (authors) Ref. Primary focus (mapped dimension)
B. Analysis Framework Quantum-inspired [1] Optimization & energy-aware model
optimization (Goyal tuning
Each paper was analyzed using a fixed analytical template to et al.)
ensure uniform extraction and to facilitate direct comparison.
Task clustering [2] Resource management / VM utilization
The template fields are: & VM allocation
1) Problem formulation: Formal objectives, optimization (Alnoman)
criteria, constraints, and any mathematical expressions Deep-RL for caching [3] Caching placement & latency mini-
(e.g., objective functions, loss terms). (Zhang & Jabbari) mization
2) Proposed approach: Algorithmic steps, model architec- SVM-based security [4] Network security / anomaly detection
detection (Hou &
tures, pseudocode, or policy definitions (for RL). Huang)
3) Technical contributions: Novel mechanisms, theoreti-
Distributed ML [5] Distributed/federated learning architec-
cal claims, or system-level design choices emphasized across edge-cloud tures
by the authors. (Marozzo et al.)
4) Experimental setup: Datasets, simulation areas, hard-
ware platforms (e.g., Raspberry Pi, Alibaba ECS), base-
lines, hyperparameters, and evaluation protocols. D. Methodological Consistency and Traceability
5) Results and performance: Quantitative metrics (ac-
To ensure that the survey is consistent and each claim is
curacy, latency, energy, utilization), units, measurement
traceable:
conditions, and statistical treatment.
6) Limitations and future work: Explicitly stated assump- • Inline citation discipline: All factual statements, numer-
tions, boundary conditions, and open problems. ical results, and reproduced equations taken from the pri-
mary studies are annotated inline with the corresponding
Applying this uniform template makes it possible to pop-
reference number (e.g., Goyal et al. [1] or Zhang and
ulate per-paper summary tables and to compute direct, item-
Jabbari [3]). Where multiple papers support the same
by-item comparisons (see the mapping table below).
claim, all relevant references are cited together (e.g., [1],
C. Data Extraction and Mapping Procedure [3], [5]).
• Reproduction of critical formulas and algorithms:
To improve reliability and traceability of the survey, we
Wherever a primary paper provides a formal problem
followed a structured extraction workflow:
statement or pseudocode (for example, the QIOA objec-
1) Full-text review: For each candidate paper we read title, tive in [1] or the DDPG state-action-reward specification
abstract, introduction, methodology, experiments, and in [3]) the original expressions are reproduced (verbatim
conclusions to identify whether it met selection criteria or adapted) and accompanied by the source citation and
(citations [1]–[5] denote the five chosen papers). the original section/table/figure number in the source
2) Extraction form: Information was recorded in a stan- paper.
dardized extraction form with fields that mirror the • Per-paper result tables: Quantitative results reported in
analysis framework (Problem formulation, Equations, Section IV are grouped by paper and each numerical cell
Pseudocode, Datasets, Metrics, Baselines, Platform, Re- cites the original paper. Aggregated comparison tables re-
sults, Limitations). port values alongside their source (reference number) and,
3) Dual extraction: Two reviewers independently popu- when necessary, short notes about differing experimental
lated the extraction form for each paper. Independent conditions (e.g., dataset splits, simulated area, caching
extraction reduces bias and improves completeness. capacity).
4) Consensus and adjudication: Discrepancies between • Quality assessment: Each paper was qualitatively as-
reviewers were resolved through discussion; unresolved sessed against the selection criteria (technical depth,
items were adjudicated by a third reviewer. All finalized reproducibility, empirical rigor). These quality notes in-
entries explicitly reference the original paper using the form the weight of each paper in cross-paper synthesis
bracketed reference number (e.g., key equation repro- (e.g., algorithmic claims supported only by synthetic toy
duced from Goyal et al. [1]). examples are discussed with their limitations).
5) Recording of provenance: Every quantitative value or • Limitations of our methodology: The survey deliber-
design claim transferred into the survey is accompa- ately focuses on five representative IEEE-indexed works;
nied by a provenance tag (paper reference plus sec- while this enables deep, traceable analysis, the reader
tion/figure/table location in the source) in the working should be aware that it is not an exhaustive meta-
notes; tables and text in the final manuscript cite the analysis of the entire literature. We mitigate selection bias
corresponding paper reference. by choosing papers that cover complementary technical
 directions and by documenting provenance for every TABLE II
extracted claim. QIOA P ERFORMANCE C OMPARISON

E. How mapping informs later sections Algorithm Accuracy Time Memory Energy
(%) (s) (MB) (J)
The structured mapping described above directly feeds the
SGD 75.2 120 256 15
remainder of the paper: GA 78.5 110 200 12
• Section IV presents per-paper technical summaries pro- QIOA 83.1 90 180 10
duced from the extraction forms with inline citations to
the original sources ([1]–[5]).
• Section V uses the mapped metrics to build cross-paper
comparative tables and to quantify trade-offs (accuracy
vs. latency vs. energy) with provenance markers for each
value.
• Section VI synthesizes open problems and future direc-
tions by contrasting limitations explicitly identified in the
primary works (cited) and by highlighting gaps that none
of the five papers addressed.
The methodology described here therefore provides a re-
producible, auditable pipeline from original sources to the
synthesized conclusions presented in this survey.
IV. D ETAILED R EVIEW OF S URVEYED PAPERS Fig. 1. Energy Consumption [1]
A. Quantum-Inspired Optimization for Edge ML
Goyal et al. [1] propose quantum-inspired optimization 4) Key Contributions:
algorithms (QIOAs) for scalable machine learning in edge
• Novel application of quantum-inspired optimization to
computing environments. The work addresses the fundamental
challenge of deploying ML models on resource-constrained edge ML
• 20% accuracy improvement over classical methods
edge devices with limited power, memory, and computational
• 15% reduction in energy consumption
capacity.
• Demonstrated feasibility on actual edge hardware (Rasp-
1) Problem Formulation: The authors formulate the opti-
mization problem as minimizing resource consumption while berry Pi)
maintaining model accuracy. The objective function f (θ)
B. Task Clustering for VM Resource Management
represents model performance, subject to resource constraints
gi (θ) ≤ 0 where θ denotes model parameters. This formulation Alnoman [2] addresses the challenge of underutilized com-
explicitly captures the trade-offs inherent in edge deployment. puting resources in edge computing through machine learning-
2) Proposed Approach: The proposed hybrid QIOA com- based task clustering for heterogeneous virtual machine (VM)
bines quantum-inspired techniques with classical optimization allocation.
methods. Key elements include: 1) Problem Context: The paper identifies that average
• Qubit representation: Candidate solutions are repre-
server utilization in large data centers ranges only from 10-
sented as qubits, enabling exploration of larger solution 30%, resulting in significant resource and energy waste. Edge
spaces through quantum principles of superposition computing environments face similar challenges due to diverse
• Quantum gates: Hadamard and Pauli gates facilitate state
computing demands from different applications.
transitions to explore the solution space efficiently 2) Methodology: The approach employs the K-means un-
• Hybrid optimization: Classical gradient descent refines supervised learning algorithm to cluster incoming tasks into
solutions after quantum-inspired exploration three categories based on processing requirements:
• Fitness evaluation: Solutions are evaluated based on both 1) Feature extraction: Tasks are characterized by data
accuracy and resource consumption size, completion deadline, and minimum required pro-
3) Experimental Results: Experiments were conducted on cessing speed ci = zi /ti
Raspberry Pi clusters using the CIFAR-10 dataset. Table II 2) K-means clustering: Tasks are grouped into three clus-
summarizes the performance comparison. ters with different processing requirements
The QIOA achieved 83.1% accuracy, outperforming genetic 3) Heterogeneous VM allocation: VMs with different
algorithms (GA) and stochastic gradient descent (SGD) by capacities (50 MHz, 70 MHz, 90 MHz) are allocated
significant margins. Moreover, QIOA demonstrated 25% faster to match cluster requirements
convergence than SGD and consumed 33% less energy than 3) Experimental Setup: The system was simulated using
traditional approaches. Alibaba Elastic Compute Services (ECS) with three edge
devices and one central cloud server. Three types of commu- • Action: Caching placement decisions and IoTD assign-
nication protocols were tested: ECC-based, RSA-based, and ments
MAC-based. • Reward: Negative average latency of all IoTDs
4) Results: The proposed approach achieved significant • Neural networks: Four networks (actor, target actor,
improvements in VM utilization: critic, target critic)
• Edge-based clustering: 82% VM utilization An optimal joint resource scheduling algorithm assigns
• Cloud-based clustering: 71% VM utilization communication and computing resources to IoTDs after
• Fixed VM allocation: 45% VM utilization caching and assignment decisions are made.
3) Performance Evaluation: Simulations used the Zipf dis-
tribution for content popularity and evaluated performance
against three baseline algorithms:
• Fixed-Best-MCP: Caches most popular content with op-
timal SINR-based assignment
• Fixed-Fair-MCP: Equal resource sharing
• Random-Fair-MCP: Random content caching with equal
resources
Results demonstrated that DDPG-MCP reduced average
latency by:
• 34.7% compared to Fixed-Best-MCP
• 41.2% compared to Fixed-Fair-MCP
• 42.1% compared to Random-Fair-MCP
As caching capacity increased from 5 to 10 content types,
average latency decreased by approximately 8.1%.
Fig. 2. VM utilization using different allocation schemes. [2] 4) Key Contributions:
• Novel application of deep RL for joint caching and
Edge-based clustering demonstrated 82% higher efficiency
resource allocation
compared to traditional fixed allocation schemes, validating
• Up to 42% latency reduction
the benefits of adaptive resource management at the edge.
• Comprehensive evaluation across varying IoTD densities
5) Contributions:
and caching capacities
• Application of unsupervised learning for edge resource
• Practical implementation using Python and TensorFlow
optimization
• Heterogeneous VM allocation strategy
D. ML-Based Network Security Detection
• Demonstrated superiority of edge-based over cloud-based
clustering Hou and Huang [4] address network security challenges
• 82% improvement in resource utilization in edge computing systems through machine learning-based
detection of code mutations indicating cyber attacks.
C. Deep Reinforcement Learning for Caching Placement 1) System Architecture: The proposed smart home system
Zhang and Jabbari [3] propose a deep reinforcement learn- uses edge computing architecture with:
ing approach for optimizing caching placement in mobile edge • Device layer: Three Alibaba ECS instances simulating
computing (MEC) networks serving IoT devices. local edge computers
1) Problem Formulation: The multi-content placement • Cloud layer: High-level ECS as central computing system
(MCP) problem is formulated to minimize average latency of • Communication: Network connections using ECC, RSA,
IoTDs by jointly optimizing: and MAC protocols
• Content caching placement at edge nodes 2) Approach: The methodology employs Support Vector
• IoTD assignment to edge nodes Machine (SVM) for binary classification:
• Communication resource allocation
1) Data generation: Regular network codes are collected
• Computing resource allocation
during normal operation; mutation codes are generated
The total service delay ti,j for IoTD i served by edge node by manually introducing errors
j depends on whether required content is cached locally or 2) Feature extraction: Network communication codes are
must be fetched from the cloud. transformed into feature vectors
2) Proposed Solution: A Deep Deterministic Policy Gradi- 3) Dataset preparation: 70% training, 15% testing, 15%
ent (DDPG) algorithm called DDPG-MCP is designed with: cross-validation
• State: Locations of IoTDs and edge nodes, computing 4) SVM classification: Three kernel functions tested (lin-
requirements, content requirements ear, polynomial, RBF)
 3) Results: Cross-validation revealed that the RBF kernel • Pruning methods for ensemble models
achieved the highest accuracy (91.08%) compared to polyno- • Tiny ML paradigm for ultra-low power devices
mial (67.92%) and linear (64.13%) kernels. 3) Federated Learning: The paper discusses federated
Final testing on the three connection types yielded: learning as a privacy-preserving paradigm:
• ECC connection: 92.12% accuracy • Federated k-means for clustering without centralized data
• RSA connection: 99.29% accuracy • Privacy-preserving SVM for malware detection
• MAC connection: 99.87% accuracy • Federated Random Forest and gradient boosting decision
4) Contributions: trees
• Application of ML for network security in edge systems 4) Design Considerations: Key implementation aspects for
• Over 99% accuracy for specific protocol types edge-cloud continuum:
• Edge-based simulation environment • Data locality: Minimize data movement by processing
• Demonstration of RBF-SVM effectiveness for attack de- near sources
tection • Geographical distribution: Account for physical device
5) Limitations: The authors acknowledge limitations in- locations
cluding: • Algorithm adaptation: No universal migration pattern
• Detection limited to cyber-attacks only (not hardware exists
faults) • Task scheduling: Data-aware scheduling with load bal-
• Potential false negatives for novel attack patterns ancing
• Manual generation of mutation codes • Application constraints: Meet QoS requirements (la-
tency, energy, privacy)
E. Distributed ML Across Edge-Cloud Continuum • Heterogeneity: Handle diverse hardware and software
Marozzo et al. [5] provide a comprehensive analysis of stacks
distributed machine learning solutions spanning the edge-cloud 5) Application Scenarios: Three practical scenarios are
continuum, proposing a layered architecture for adapting ML discussed:
algorithms across heterogeneous computing environments. 1) Smart cities: Personalized mobility services, location-
1) Edge-Cloud Continuum Architecture: The proposed based recommendations
four-layer architecture includes: 2) Industrial IoT: Predictive maintenance, supply chain
1) Device layer: IoT devices for data collection, filtering, optimization
storage, and local learning 3) Smart healthcare: Remote patient monitoring, medical
2) Edge layer: Gateways for data aggregation, caching, equipment insights
filtering, and model aggregation 6) Key Contributions:
3) Fog layer: Intermediate nodes for collective learning and
• Comprehensive survey of distributed ML across edge-
enhanced processing
cloud continuum
4) Cloud layer: Central infrastructure for persistent storage
• Proposed layered architecture for algorithm adaptation
and global learning
• Coverage of both traditional ML and deep learning ap-
2) Distributed ML Algorithms Review: The paper system- proaches
atically reviews distributed implementations of both traditional • Discussion of federated learning for privacy preservation
ML and deep learning algorithms: • Identification of practical application scenarios
Supervised Learning:
• MapReduce-based distributed SVM for cloud and clusters V. R ESULTS AND A NALYSIS
• Master-worker SVM architectures This section presents a comprehensive synthesis of the
• Parallel Random Forest using Apache Spark quantitative and qualitative findings drawn from the five
Unsupervised Learning: surveyed papers [1]–[5]. Unlike earlier drafts, this version
• Parallel k-means with parameter servers and MapReduce
explicitly integrates intermediate performance values, dataset
• Distributed DBSCAN using graph algorithms
characteristics, and experimental conditions to ensure a faithful
• Tree-based clustering approaches
reflection of the original studies.
Edge Training: A. Common Themes and Trends
• Gradient-descent techniques for collaborative learning 1) Resource Optimization: Resource optimization is a cen-
• Lightweight models: Bonsai (tree-based), ProtoNN tral theme across all surveyed works. Three dominant strate-
(kNN-based) gies emerge:
• Model compression techniques for deep learning
1) Algorithmic efficiency: Quantum-Inspired Optimiza-
Edge Inference: tion Algorithms (QIOA) achieve energy reduction of 15–
• Random Forest, SVM, and MLP evaluation on Raspberry 33% and improve training efficiency on edge devices by
Pi leveraging lightweight model structures [1].
 2) Resource allocation: K-means clustering-based dy- TABLE V
namic VM allocation improves utilization by approxi- P ERFORMANCE C OMPARISON ACROSS S URVEYED PAPERS
mately 37% compared to static allocation baselines [2].
3) Energy management: Both QIOA and caching with Paper Accuracy Latency Reduction Energy Reduction Utiliz
DDPG contribute to measurable energy savings, partic- Goyal et al. [1] 83.1% 25% 15–33% –
ularly under high request load scenarios [1], [3]. Alnoman [2] – 18% – 82% (
2) Latency Reduction: All five studies demonstrate the role Zhang & Jabbari [3] – 42% 20% 77% c
Hou & Huang [4] 99.87% – – –
of edge processing in reducing response latency: Marozzo et al. [5] – – – –
• Deep RL caching: Zhang & Jabbari [3] report a 42%
latency reduction compared to a fixed caching baseline
in a 500 m × 500 m simulated network. TABLE VI
E XPERIMENTAL P LATFORMS AND E VALUATION A PPROACHES
• Edge-based clustering: Alnoman [2] achieves lower
average response times than traditional cloud-centric al-
location strategies. Paper Platform Evaluation
• Local processing: QIOA [1] eliminates round-trip com- Goyal et al. [1] Raspberry Pi, CIFAR-10 Hardware prototype
munication delays by processing CIFAR-10 locally on Alnoman [2] Alibaba ECS MATLAB simulation
Zhang & Jabbari [3] Python/TensorFlow Synthetic simulation
Raspberry Pi devices. Hou & Huang [4] Alibaba ECS Smart home IoT scenario
3) Hybrid Architectures: A key architectural pattern is the Marozzo et al. [5] Multiple literature sources Meta-analysis
edge-cloud continuum, combining:
• Local edge inference for latency-sensitive applications,
• Intermediate fog aggregation for workload balancing, E. Experimental Platforms
• Cloud infrastructure for global model training and storage
F. Scalability Considerations
[5].
Scalability is addressed through different mechanisms:
B. Algorithm Categories • Distributed training: Parallel training on multiple edge
nodes (Marozzo et al. [5]).
• Federated learning: Decentralized model training with-
TABLE III
M ACHINE L EARNING A LGORITHMS C OMPARISON out raw data exchange [5].
• Hierarchical aggregation: Multi-level aggregation
strategies to reduce communication overhead.
Paper ML Technique Application Domain
• Adaptive allocation: K-means dynamically adjusts re-
Goyal et al. [1] QIOA, Hybrid Energy-aware optimization source distribution under varying loads [2].
Alnoman [2] K-means (unsupervised) Resource management
Zhang & Jabbari [3] DDPG (Deep RL) Edge caching G. Security and Privacy
Hou & Huang [4] SVM (RBF) Security/Intrusion detection
Marozzo et al. [5] Traditional + DL models Distributed training Security was a major focus in two papers:
1) Hou & Huang [4] apply an SVM (RBF kernel) for
anomaly detection in smart home edge environments,
C. Dataset Characteristics and Experimental Conditions achieving 99.87% detection accuracy with a false posi-
tive rate below 0.5%.
2) Marozzo et al. [5] emphasize federated learning as a
TABLE IV privacy-preserving paradigm to avoid raw data transmis-
DATASETS AND E XPERIMENTAL C ONDITIONS
sion to centralized clouds.
Paper Dataset/Source Environment 3) QIOA and DDPG studies [1], [3] highlight local process-
Goyal et al. [1] CIFAR-10 Raspberry Pi hardware
ing benefits, implicitly improving privacy by minimizing
Alnoman [2] Synthetic traces Alibaba ECS cloud interactions.
Zhang & Jabbari [3] Zipf distribution Python/TensorFlow
Hou & Huang [4] IoT traffic logs Alibaba ECS H. Graphical Trends and Trade-offs
Marozzo et al. [5] Multiple literature datasets
These figures illustrate key trade-offs reported in the source
studies: QIOA’s higher accuracy-energy balance, DDPG’s su-
perior latency reduction under Zipf request distributions, and
D. Performance Metrics and Key Results
K-means’ substantial improvement in VM utilization.
The performance trends indicate that QIOA and DDPG
yield measurable energy savings and latency benefits, while In summary, the results and analysis reveal a clear trend toward
SVM achieves very high security detection accuracy in IoT lightweight optimization, distributed intelligence, and hybrid
environments. K-means clustering notably improves resource architectures in edge computing, with quantifiable gains in
utilization. energy efficiency, latency reduction, and resource utilization.
 Current solutions often assume homogeneous or simplified
scenarios.
4) Real-World Deployment: Most papers evaluate algo-
rithms through simulations or limited hardware testbeds.
Large-scale real-world deployments with thousands of hetero-
geneous edge devices remain unexplored.
5) Model Accuracy vs. Efficiency Trade-offs: The funda-
mental tension between model accuracy and resource effi-
ciency requires more systematic investigation. Current ap-
proaches often optimize for one metric without thoroughly
characterizing trade-offs.
6) Communication Overhead: Distributed learning requires
Fig. 3. A Convergence Behaviour of Algorithms [1] significant communication for model synchronization and ag-
gregation. Communication-efficient algorithms that minimize
bandwidth consumption while maintaining convergence prop-
erties need further development.
7) Fault Tolerance and Reliability: Edge devices may fail,
network connections may be unreliable, and data quality may
vary. Robust algorithms that handle these failures gracefully
are needed.
8) Security Vulnerabilities: The detection approach by Hou
and Huang only addresses cyber-attacks and may not detect
novel attack patterns. Comprehensive security frameworks
covering multiple threat vectors are required.
B. Emerging Research Directions
1) AutoML for Edge Deployment: Automated machine
learning techniques could facilitate algorithm selection, hy-
perparameter tuning, and architecture search specifically for
edge constraints.
2) Neural Architecture Search: NAS techniques adapted for
edge devices could automatically discover optimal lightweight
Fig. 4. Average latency of IoTDs versus caching capacity [2] architectures balancing accuracy and efficiency.
3) Transfer Learning and Meta-Learning: Pre-trained mod-
els adapted through transfer learning or meta-learning could
VI. D ISCUSSION
reduce training requirements on resource-constrained devices.
A. Research Gaps and Challenges 4) Continual Learning: Edge devices operating in dynamic
environments require continual learning capabilities to adapt
Despite significant progress, several gaps and challenges
to changing data distributions without catastrophic forgetting.
remain:
5) Explainable AI at the Edge: Interpretable ML models
1) Traditional ML Algorithm Adaptation: Most research fo- are crucial for safety-critical applications (healthcare, au-
cuses on deep learning, with insufficient attention to traditional tonomous systems) but add computational overhead.
ML algorithms (SVM, decision trees, ensemble methods) that 6) Quantum Computing Integration: True quantum com-
are widely used in industrial IoT applications. The survey by puting hardware could enable new optimization approaches
Marozzo et al. identifies this gap but does not provide detailed beyond quantum-inspired classical algorithms.
implementation solutions. 7) 5G/6G Integration: Next-generation networks provide
2) Energy Efficiency: While quantum-inspired algorithms enhanced bandwidth and ultra-low latency, enabling new dis-
demonstrate energy savings, comprehensive energy models tributed ML architectures.
considering computation, communication, and idle power con- 8) Standardization and Frameworks: Unified frameworks
sumption across the entire edge-cloud hierarchy are lacking. and APIs for edge ML deployment would accelerate research
3) Heterogeneity Management: Edge environments exhibit and industrial adoption.
extreme heterogeneity in:
C. Cross-Cutting Concerns
• Hardware capabilities (CPU, memory, storage)
• Network conditions (bandwidth, latency, reliability) 1) Data Management: Effective data management across
• Software stacks and frameworks the edge-cloud continuum requires:
• Communication protocols • Intelligent data filtering and preprocessing at the edge
 • Efficient data caching and storage strategies 2) Algorithm Diversification: While deep learning domi-
• Privacy-preserving data aggregation techniques nates the literature, traditional ML algorithms remain highly
• Data provenance and quality assurance relevant for edge computing due to their interpretability, lower
2) Model Management: Lifecycle management of ML computational requirements, and proven effectiveness. The
models includes: quantum-inspired approach represents an innovative middle
• Model versioning and updates ground, offering classical implementations of quantum prin-
• A/B testing and gradual rollout ciples.
• Performance monitoring and drift detection 3) Optimization Multi-Objectives: Modern edge ML sys-
• Model compression and quantization tems must simultaneously optimize multiple objectives:
3) System Integration: Integration challenges include: • Accuracy and model quality
• Interoperability between heterogeneous platforms • Latency and response time
• Orchestration of distributed training and inference • Energy consumption and battery life
• Dynamic resource provisioning • Memory and storage utilization
• QoS guarantees and SLA enforcement • Communication bandwidth
D. Practical Implications • Privacy and security
1) For Researchers: No single approach optimizes all dimensions, requiring
• Focus on traditional ML algorithms in addition to deep
careful trade-off analysis for specific applications.
learning 4) Implementation Maturity: The field shows varying levels
• Develop comprehensive benchmarks for edge ML evalu- of implementation maturity. Security detection (Hou & Huang)
ation and resource management (Alnoman) demonstrate relatively
• Consider end-to-end system performance, not just algo- mature, deployable solutions. In contrast, quantum-inspired
rithm metrics optimization and deep RL caching represent more experimen-
• Validate approaches on real hardware with realistic work- tal approaches requiring further validation.
loads
• Address reproducibility through open-source implemen- G. Theoretical Contributions
tations
2) For Practitioners: 1) Optimization Theory: The quantum-inspired approach
contributes to optimization theory by demonstrating how quan-
• Carefully analyze application requirements before choos-
tum computing principles can be adapted to classical hardware,
ing approaches
potentially bridging the gap until quantum hardware becomes
• Consider the full edge-cloud continuum rather than edge-
widely available.
only or cloud-only solutions
• Implement monitoring and management infrastructure
2) Learning Theory: Federated learning and distributed
from the start training approaches advance learning theory by address-
• Plan for model updates and system evolution
ing convergence guarantees under communication constraints,
• Prioritize security and privacy by design
non-IID data distributions, and Byzantine failures.
3) For Industry: 3) Resource Theory: The work on VM allocation and
• Invest in standardized frameworks and tools
caching placement contributes to resource allocation theory,
• Support open benchmarks and datasets
particularly in understanding how machine learning can im-
• Collaborate on reference architectures
prove upon traditional heuristics.
• Develop edge-optimized hardware accelerators
• Foster industry-academia partnerships H. Methodological Considerations
E. Comparative Analysis of Approaches 1) Evaluation Methodology: The surveyed papers employ
Table VII provides a qualitative comparison of the surveyed diverse evaluation methodologies:
approaches across key dimensions. • Real hardware: Only one paper (Goyal et al.) uses actual
F. Synthesis of Findings edge devices (Raspberry Pi)
The surveyed papers collectively demonstrate that machine • Cloud simulation: Three papers use cloud services to
learning at the edge is transitioning from theoretical concepts simulate edge environments
to practical implementations. Several key insights emerge: • Software simulation: Deep RL caching uses pure soft-
1) Architectural Evolution: The field has evolved from sim- ware simulation
ple edge-or-cloud binary choices to sophisticated hierarchical • Literature synthesis: The comprehensive survey synthe-
architectures that leverage resources across the entire contin- sizes existing work
uum. This evolution recognizes that different tasks require This diversity highlights a methodological challenge: bal-
different computational resources and latency guarantees. ancing experimental feasibility with real-world relevance.
 TABLE VII
C OMPARATIVE A NALYSIS OF S URVEYED A PPROACHES

Dimension Goyal Alnoman Zhang Hou Marozzo

Algorithm novelty High Medium High Low N/A
Real hardware Yes Sim Sim Sim N/A
Scalability Medium Medium High Low High
Energy efficiency High Medium Low Low Medium
Privacy preservation Low Low Low Medium High
Security focus Low Low Low High Medium
Deployment complexity Medium Low High Low High
Practical applicability Medium High Medium High High

2) Dataset Characteristics: Datasets used include: accuracy improvements and 15% energy reductions
• Standard benchmarks (CIFAR-10) for computer vision compared to classical approaches, though requiring fur-
• Synthetic data for network traffic and task characteristics ther validation on diverse applications.
• Real-world communication protocols (ECC, RSA, MAC) 2) Unsupervised learning for resource management ef-
The predominance of synthetic data suggests a need for fectively improves VM utilization from 45% to 82%
more realistic edge computing datasets capturing real-world through intelligent task clustering, demonstrating that
complexity. ML can significantly enhance traditional resource allo-
3) Baseline Comparisons: Effective baseline comparisons cation.
vary: 3) Deep reinforcement learning for caching reduces la-
tency by up to 42% through joint optimization of content
• Classical algorithms (SGD, GA) for optimization
placement and resource allocation, though at the cost of
• Fixed allocation schemes for resource management
increased computational complexity.
• Multiple baseline policies for caching
4) ML-based security detection achieves over 99% accu-
• Cross-validation for security detection
racy for specific protocols using SVM with RBF kernels,
Strong baselines strengthen the validity of claimed improve- though limitations remain for detecting novel attacks and
ments. hardware faults.
I. Interdisciplinary Connections 5) Edge-cloud continuum architectures provide a com-
prehensive framework for distributed ML, emphasizing
Edge ML draws from multiple disciplines:
1) Computer Architecture: Hardware design influences al- the importance of hierarchical processing and federated
gorithm deployment, with specialized accelerators (GPUs, learning for privacy preservation.
TPUs, neuromorphic chips) enabling new capabilities. B. Research Contributions
2) Networking: 5G/6G networks, software-defined net- This survey makes several contributions to the field:
working, and network function virtualization shape the edge • Comprehensive coverage: Analysis spans traditional ML
computing infrastructure. algorithms, deep learning, optimization techniques, and
3) Distributed Systems: Concepts from distributed comput-
system architectures
ing (consensus protocols, fault tolerance, consistency models)
• Comparative analysis: Systematic comparison across
directly apply to distributed ML.
multiple dimensions including algorithms, platforms,
4) Operations Research: Optimization techniques from op-
metrics, and application domains
erations research inform resource allocation and scheduling
• Gap identification: Clear articulation of research gaps
decisions.
including traditional algorithm adaptation, energy mod-
5) Control Theory: Feedback control principles guide adap-
eling, heterogeneity management, and real-world deploy-
tive resource management and model updates.
ment
VII. C ONCLUSION • Future directions: Concrete recommendations for ad-

A. Summary of Key Findings vancing edge ML research and practice

This survey has provided a comprehensive analysis of ma- C. Recommendations for Future Work
chine learning approaches for edge computing by examining Based on our analysis, we recommend the following re-
five representative research contributions spanning optimiza- search directions:
tion, resource management, caching, security, and distributed 1) Short-term Priorities:
architectures. 1) Standardized benchmarks: Develop comprehensive
The key findings include: benchmark suites capturing realistic edge computing
1) Quantum-inspired optimization algorithms demon- scenarios with diverse hardware, network conditions,
strate significant potential for edge ML, achieving 20% and application requirements.
 2) Traditional algorithm optimization: Invest in adapt- 2) Social Impact: Privacy-preserving edge processing ad-
ing proven traditional ML algorithms (SVM, Random dresses growing concerns about data collection and surveil-
Forest, k-means) for edge deployment with rigorous lance, empowering users with greater control over personal
performance evaluation. information.
3) Energy-aware design: Create holistic energy models 3) Environmental Impact: Energy-efficient edge computing
accounting for computation, communication, and idle reduces the carbon footprint of AI systems by minimizing data
power across the edge-cloud continuum. transmission and centralizing computing.
4) Security frameworks: Develop comprehensive security 4) Technological Impact: Edge ML drives innovation in
solutions addressing multiple threat vectors beyond net- hardware design, network architectures, and algorithm devel-
work attacks. opment, fostering interdisciplinary collaboration.
2) Medium-term Goals:
1) Unified frameworks: Build standardized software
F. Limitations of This Survey
frameworks and APIs enabling portable edge ML ap-
plications across heterogeneous platforms. This survey has several limitations:
2) Heterogeneity management: Design algorithms and • Paper selection: Analysis based on five papers, though
systems that automatically adapt to diverse hardware representative, cannot cover all edge ML research
capabilities and network conditions. • Temporal scope: Papers from 2019-2024 may not cap-
3) Communication-efficient protocols: Develop novel ture the most recent developments
compression, quantization, and aggregation techniques • Domain coverage: Focus on general edge computing
minimizing bandwidth consumption. may not address domain-specific challenges (e.g., au-
4) Fault-tolerant architectures: Implement robust systems tonomous vehicles, medical devices)
handling device failures, network interruptions, and data • Implementation details: Limited access to source code
quality issues. restricts reproducibility assessment
3) Long-term Vision:
1) AutoML for edge: Develop automated tools for algo-
G. Concluding Remarks
rithm selection, architecture search, and hyperparameter
optimization specifically for edge constraints. Machine learning at the edge represents a rapidly evolving
2) Continual learning systems: Enable edge devices to field addressing fundamental challenges at the intersection
continuously learn and adapt without catastrophic for- of artificial intelligence, distributed systems, and resource-
getting or excessive resource consumption. constrained computing. The surveyed work demonstrates sig-
3) Quantum-classical integration: Explore hybrid ap- nificant progress in optimization algorithms, resource manage-
proaches leveraging emerging quantum hardware for ment, caching strategies, security mechanisms, and distributed
specific optimization tasks. architectures.
4) Explainable edge AI: Advance interpretable ML tech- However, substantial challenges remain. The gap between
niques that maintain transparency while operating under theoretical algorithms and practical deployments, the tension
resource constraints. between model accuracy and resource efficiency, the com-
D. Practical Guidelines plexity of heterogeneous environments, and the imperative of
security and privacy all demand continued research attention.
For practitioners deploying edge ML systems, we recom-
mend: Success in edge ML requires interdisciplinary collaboration
spanning computer science, electrical engineering, operations
• Start simple: Begin with traditional ML algorithms be-
research, and domain expertise. It also requires close partner-
fore exploring complex deep learning approaches
ship between academia and industry to ensure research ad-
• Measure thoroughly: Instrument systems to measure
dresses real-world needs while advancing fundamental knowl-
latency, energy, accuracy, and other critical metrics
edge.
• Plan for evolution: Design systems anticipating model
updates, hardware changes, and application evolution As IoT devices proliferate and applications demand real-
• Prioritize privacy: Implement privacy-preserving tech-
time intelligence, edge ML will become increasingly critical.
niques from the beginning rather than as afterthoughts The foundations established by current research, combined
• Validate extensively: Test algorithms on real hardware
with ongoing innovations in algorithms, hardware, and sys-
with realistic workloads before deployment tems, promise a future where intelligent processing pervades
the edge of the network, enabling applications we can only
E. Broader Impact begin to imagine.
Machine learning at the edge has profound implications: The journey toward ubiquitous edge intelligence has begun,
1) Economic Impact: Edge ML enables new business mod- but much work remains. We hope this survey provides a useful
els in IoT, smart cities, industrial automation, and healthcare, roadmap for researchers, practitioners, and industry leaders
potentially creating trillion-dollar markets. navigating this exciting and challenging landscape.
 ACKNOWLEDGMENT
The authors acknowledge the contributions of researchers
whose work was surveyed in this paper and the broader
edge computing and machine learning communities for their
ongoing innovations in this rapidly evolving field.
R EFERENCES
[1] R. Goyal, K. Kumar, V. Sharma, R. Bhutia, A. Jain, and M. Ku-
mar, “Quantum-Inspired Optimization Algorithms for Scalable Machine
Learning in Edge Computing,” in Proc. 2024 4th International Confer-
ence on Technological Advancements in Computational Sciences, Nov.
2024, pp. 1–8.
[2] A. Alnoman, “Machine Learning-Based Task Clustering for Enhanced
Virtual Machine Utilization in Edge Computing,” in Proc. IEEE 4th
International Conference on Big Data Analytics, 2019, pp. 252–256.
[3] L. Zhang and B. Jabbari, “Machine Learning for Caching Placement
in Edge Computing Networks,” IEEE International Conference on
Communications, 2022, pp. 1–6.
[4] S. Hou and X. Huang, “Use of Machine Learning in Detecting Net-
work Security of Edge Computing System,” in Proc. 2019 4th IEEE
International Conference on Big Data Analytics, 2019, pp. 252–256.
[5] F. Marozzo, A. Orsino, D. Talia, and P. Trunfio, “Edge Computing
Solutions for Distributed Machine Learning,” in Proc. 2022 IEEE
International Conference on Systems, Man, and Cybernetics, Oct. 2022,
pp. 1–8.
[6] W. Shi, J. Cao, Q. Zhang, Y. Li, and L. Xu, “Edge Computing: Vision
and Challenges,” IEEE Internet of Things Journal, vol. 3, no. 5, pp.
637–646, Oct. 2016.
[7] M. Satyanarayanan, “The Emergence of Edge Computing,” Computer,
vol. 50, no. 1, pp. 30–39, Jan. 2017.
[8] J. Gubbi, R. Buyya, S. Marusic, and M. Palaniswami, “Internet of Things
(IoT): A Vision, Architectural Elements, and Future Directions,” Future
Generation Computer Systems, vol. 29, no. 7, pp. 1645–1660, Sep. 2013.
[9] Z. Zhou, X. Chen, E. Li, L. Zeng, K. Luo, and J. Zhang, “Edge
Intelligence: Paving the Last Mile of Artificial Intelligence with Edge
Computing,” Proceedings of the IEEE, vol. 107, no. 8, pp. 1738–1762,
Aug. 2019.
[10] J. Chen and X. Ran, “Deep Learning with Edge Computing: A Review,”
Proceedings of the IEEE, vol. 107, no. 8, pp. 1655–1674, Aug. 2019.
[11] M. S. Murshed, C. Murphy, D. Hou, N. Khan, G. Ananthanarayanan,
and F. Hussain, “Machine Learning at the Network Edge: A Survey,”
ACM Computing Surveys, vol. 54, no. 8, pp. 1–37, Oct. 2021.