Open research questions in IoT and Edge/Fog Computing

What the literature leaves open

• 6.1 Synthesis of the Evolutionary Trajectory From Device-Cloud to On-Device, the historical evolution of inference architectures has been a search for balance among competing objectives of computational power, latency, privacy, and cost. No single architecture is a silver bullet. The future lies in a hybrid, adaptive framework that treats the entire spectrum of resources—from the user's device to the centralized cloud—as a unified "compute continuum" [30]. In this paradigm, intelligent orchestration systems will dynamically route each inference request to the most appropriate execution venue based on factors such as query complexity, Service Level Objectives (SLOs), data privacy requirements, real-time system state (network bandwidth, device battery, server load), and economic cost. 6.2 Emerging Trends and Open Research Questions Purpose-Built Inference Hardware: The market is shifting towards developing hardware specifically engineered for inference, prioritizing latency, energy efficiency, and cost-per-query over raw training throughput. This includes novel chip designs and domain-specific architectures. Sustainable and Energy-Aware AI: The exponential growth of AI is creating a significant energy demand. Future research will increasingly focus on energy-aware inference, developing scheduling algorithms and hardware that optimize for performance-per-watt. Volume 8, Issue 5, Pp 1-5, 2026 4 ZeYi Luo Inference for Long-Context and Multi-Modal Models: The Key-Value (KV) cache size of the latest large models scales linearly with context length, posing a significant challenge for the limited memory of edge and on-device systems. Developing novel caching and attention mechanisms is a major open research area. Security and Trust in Decentralized Systems: While decentralization enhances user privacy, it also introduces a more complex attack surface. Securing collaborative inference systems against adversarial attacks (e.g., data poisoning, model inversion) and ensuring model integrity are critical challenges. In conclusion, the evolution of large model inference architectures is an ongoing process. The journey from centralization to decentralization reflects a maturing understanding of the complex interplay between computation, performance, privacy, and cost. The future is a continuum, where intelligence will be fluidly and adaptively distributed to deliver the benefits of AI efficiently, securely, and sustainably. COMPETING INTERESTS The authors have no relevant financial or non-financial interests to disclose. REFERENCES [1] Devlin J, Chang M, Lee K, et al. BERT: Pre-training of deep bidirectional transformers for language understanding. North American Chapter of the Association for Computational Linguistics, 2019. [2] Brown TB, Mann B, Ryder N, et al. Language models are few-shot learners. ArXiv, 2020, abs/2005.14165. [3] Touvron H, Lavril T, Izacard G, et al. LLaMA: Open and efficient foundation language models. ArXiv, 2023, abs/2302.

  THE EVOLUTION OF LARGE MODEL INFERENCE ARCHITECTURES: FROM CENTRALIZED CLOUDS TO DECENTRALIZED ON-DEVICE INTELLIGENCE · 2026 · DOI

• However, most existing task-offloading studies are still formulated for generic user equipment and primarily focus on uplink transmission, which is insufficient for practical sensing systems where sensor nodes continuously upload measurements while simultaneously receiving control commands, model updates, and feedback from the edge.

  A Sensor-Aware Multi-Agent Reinforcement Learning Framework for Joint Data Offloading and Power Control in Edge-Assisted Wireless Sensor Networks · 2026 · DOI

• accordingly, which may resources and affect system Future research should focus on mitigating these scalability challenges to ensure practical applicability in massive IoT deployments. Despite the privacy protection optimization, further research is needed on its sustainability as new technologies evolve.

  Security-Aware Scheduling Methods for Distributed Systems with Integrated Motion Data Privacy Protection · 2026 · DOI

• ML-assisted DEEC clustering Integrates K-NN and K-Means with DEEC for adaptive cluster-head selection ML-assisted PEGASIS routing K-means clustering for spatial partitioning; K-NN classification for adaptive routing Connectivity-enhanced…

  Size- and lifetime-aware disjoint dominating set formation and multi-armed-bandit-based scheduling for energy-efficient IoT-WSNs in sustainable smart city applications · 2026 · DOI

• Our study identified several key limitations: • Gap between theoretical advancements and practical industry applications: There is a disconnect between academic research and the real-world implementation of EdgeAI systems. Theoretical models often do not account for the practical challenges industry professionals face. Future work should bridge this gap by engaging directly with industry professionals through collaborations and partnerships. Incorporating case studies and empirical data from commercial deployments can enhance the practical viability and adoption of optimisation strategies. 163 Journal of Edge Computing, 2026, Vol. 5, Iss. 1, pp.

  Revisiting EdgeAI through the lens of communication, storage and computing optimisations · 2026 · DOI

• Future research should focus on the integration of model reinforcement learning and federated learning, the combination of meta learning and DRL, the optimization of multi-agent cooperation mechanism, the application of graph neural network and the lightweight deployment of algorithm.

  Research and Analysis on Task Unloading Strategy of Edge Computing Based on Deep Reinforcement Learning · 2026 · DOI

• The optimization ideas across model layer and layer level in this article can be directly applied, but the introduction of dynamic tensor shapes [86] makes the computational load of each operator highly uncertain, so more advanced operator scheduling within the stage should be studied [87], [88], which is our future work.

  Coconut: Multilevel Collaborative Deployment for Real-Time Deep Learning Tasks in Heterogeneous Edge GPU Cluster · 2026 · DOI

• GPU utilization analysis (Figure 8) compares four methods but does not investigate how FEKD performs on edge devices without GPU acceleration or with heterogeneous compute resources (TPU, NPU, CPU-only inference) prevalent in constrained IoT edge nodes, leaving a gap in understanding applicability to truly resource-limited deployments.

  Latency-efficient edge intelligence in IoT networks using knowledge distillation · 2026 · DOI

• The model accuracy improvement formula (Equation 15) assumes logarithmic convergence behavior, but the paper provides no theoretical justification for this functional form or empirical validation of its applicability across different knowledge distillation architectures, teacher-student model size ratios, or diverse IoT sensor modalities.

  Latency-efficient edge intelligence in IoT networks using knowledge distillation · 2026 · DOI

• Energy consumption analysis (Figure 7, Equation 18) shows FEKD remains below 18 mJ at 400 tasks, but this evaluation is limited to homogeneous task distributions and does not explore energy efficiency under dynamic task arrivals, variable task complexity, or device heterogeneity-induced workload imbalance in edge intelligence systems.

  Latency-efficient edge intelligence in IoT networks using knowledge distillation · 2026 · DOI

• Communication overhead modeling (Equation 16 and Figure 6) demonstrates FEKD maintains 6-7 MB transmission across 50-200 edge devices, but does not analyze how communication costs change with variable compression ratios, quantization schemes, or bandwidth-constrained scenarios typical in remote IoT deployments with intermittent connectivity.

  Latency-efficient edge intelligence in IoT networks using knowledge distillation · 2026 · DOI

• The privacy preservation score analysis (Equation 17) claims FEKD achieves >92% privacy scores through decentralized knowledge reduction, but provides no empirical evaluation against differential privacy attacks, membership inference attacks, or gradient inversion attacks specific to federated distillation in heterogeneous edge networks.

  Latency-efficient edge intelligence in IoT networks using knowledge distillation · 2026 · DOI

• The convergence time analysis (Table 5) shows FEKD convergence times only for four specific IoT scenarios (smart home, traffic prediction, healthcare sensors, industrial machinery), but does not evaluate how convergence scales with datasets larger than those tested or across different data distributions, federated edge topologies, or non-IID data partitioning schemes common in real federated learning deployments.

  Latency-efficient edge intelligence in IoT networks using knowledge distillation · 2026 · DOI

• The FEKD framework's performance across heterogeneous IoT devices shows varying scalability levels (89-95%), but the paper does not investigate how knowledge distillation effectiveness degrades when deployed on ultra-low-power MCU nodes (62-89% scalability gap) versus high-capacity FPGA nodes, nor does it explore hardware-specific distillation strategies to close this heterogeneity gap.

  Latency-efficient edge intelligence in IoT networks using knowledge distillation · 2026 · DOI

• The scalability of BRL-STOA is evaluated only for state dimensions ranging from 10 to 50; performance evaluation with significantly larger-scale edge networks and thousands of mobile devices is needed.

  An adaptive deep reinforcement learning framework enhanced by broad reinforcement learning–based state transition for efficient task offloading and resource allocation at the network edge · 2026 · DOI

• The paper does not discuss the scalability of the proposed ISSA algorithm to larger datasets or more complex healthcare IoT environments beyond the 10,000 records tested.

  ImmutableShield: A Fragmentation-Driven Security Algorithm for Real-Time Health Data Protection in IoT-Enabled Heat-Stroke Scenarios · 2026 · DOI

• Ethical and privacy issues dictated that real patient data would not be used; instead, synthetic heat-stroke healthcare data consisting of 10,000 controlled records obtained through anonymized randomization was applied.

  ImmutableShield: A Fragmentation-Driven Security Algorithm for Real-Time Health Data Protection in IoT-Enabled Heat-Stroke Scenarios · 2026 · DOI

• By synthesizing results, quantifying trends, and framing open challenges, this SLR provides researchers and practitioners with a comprehensive understanding of the strengths, limitations, and opportunities in MEC computation offloading.

  A Systematic Review of Optimization Techniques for Computation Offloading in Mobile Edge Computing: Lyapunov, Convex, Heuristic, Game-Theoretic, and Machine Learning Approaches · 2025 · DOI

• • Chatbots and Virtual Assistants: NLP and ML play a vital role in the development of chatbots and virtual assistants for healthcare. These conversational agents can understand and respond to natural language queries, provide basic medical advice, schedule appointments, and offer personalized health recommendations. By leveraging ML techniques, these systems can continuously improve their responses and adapt to individual user preferences. Downloaded for Anonymous User (n/a) at Kirikkale University from ClinicalKey.com by Elsevier on June 16, 2026. For personal use only. No other uses without permission. Copyright ©2026. Elsevier Inc. All rights reserved. 564 D.M. Mathkor, N. Mathkor, Z. Bassfar et al. Journal of Infection and Public Health 17 (2024) 559–572 Fig. 7. Significance of AI-based methodologies in diverse medical areas. Modern diagnoses and individualized treatment plans need to be delivered quickly in pharmacogenomics. AI provides real-time solutions in identifying novel paths for addressing specific ailments relying on historical and current data (Fig. 7). Using AI-based solutions, the many components of the healthcare ecosystem may be changed. These will include AI methods for building processors that automatically record healthcare data, schedule medical visits, decide on lab testing, drug treatments, drugs, and medical interventions, among other things. By receiving further training, these categories could aid decision-making processes. In the architecture for other classifiers, such disorganized data points that are unable to be digitized can be understood using techniques provided by NLP. These might be in the form of test results, notes from a physical exam, notes from a surgery, and other information pertaining to patient discharge. Machine learning also has the ability to forecast the future using data from the past. It makes the use of controlled, uncontrolled, or reinforcement learning to predict future events). Utilizing the medical data & patient histories, intelligence solutions are also tailoring the most efficient treatment strategy as well as specific prescriptions for people. Healthcare practitioners may simply check and receive information on patients’ heart rates and energy levels using monitoring devices sensor technologies. An AIbased technology is employed to analyze data and identify abnormalities in specific persons since the amounts of information is large and comes from several sources. Similar to this, in hospitals, information gathered from patient-specific health monitoring equipment can identify potential crises and notify medical personnel.

  Multirole of the internet of medical things (IoMT) in biomedical systems for managing smart healthcare systems: An overview of current and future innovative trends · 2024 · DOI

• OPEN CHALLENGES AND THE WAY FORWARD In this section, we discuss multi-faceted open research chal- lenges and point out strategies and candidate technologies as a remedy. Guo, “Semantic communications: Overview, open issues, and future research directions,” IEEE Wireless Commun. Verma, “Machine learning-based physical layer security: techniques, open challenges, and applica- tions,” Wireless Networks, vol.

  At the Dawn of Generative AI Era: A Tutorial-cum-Survey on New Frontiers in 6G Wireless Intelligence · 2024 · DOI

• There are still unresolved is- sues in several domains, lim- ited to specific areas and ap- plications, and may give un- readable results in other ar- eas. Lack of the necessary aware- ness of the security conse- quences until after a breach has happened, leading to sig- nificant losses including the loss of vital data. Limited description of secu- rity risks to the IoT network and not enough discussion of mitigation strategies. This study suffers from re- source restrictions for IoT devices.

  A Survey on Security, Privacy, Trust, and Architectural Challenges in IoT Systems · 2024 · DOI

• Das, “Emerging Access Technologies and Open Challenges in 5G IoT: From Physical Layer Perspective,” in International Symposium on Advanced Networks and Telecommunication Systems, ANTS, 2018, vol. Pierantonio, “Cloud-based modeling in IoT domain: A survey, open challenges and opportunities,” Companion Proc.

  Toward a Secure 5G-Enabled Internet of Things: A Survey on Requirements, Privacy, Security, Challenges, and Opportunities · 2024 · DOI

• Scalability and adaptability are two open issues that have not yet made full use of AI models as an economical way to boost the performance of computer applications. , “Transformative effects of IoT, blockchain and artificial intelligence on cloud computing: Evolution, vision, trends and open challenges,” Internet of Things, vol. , “Computation offloading in mobile cloud computing and mobile edge computing: survey, taxonomy, and open issues,” Mobile Information Systems, vol. , “A survey on the internet of things (iot) foren- sics: challenges, approaches, and open issues,” IEEE Communica- tions Surveys & Tutorials, vol. , “Advances and open problems in federated learn- ing,” Foundations and Trends® in Machine Learning, vol. , “Guaranteeing end-to-end qos provisioning in soa based sdn architecture: A survey and open issues,” Future Genera- tion Computer Systems, vol.

  Modern computing: Vision and challenges · 2024 · DOI

• the model DRLMOTS, tasks must be scheduled effectively onto the VMs. The scheduler to the execution queue must first send these priority jobs, and then they must be sent to the VMs. Our scheduler must monitor upcoming requests in our proposed architecture and notify the resource manager. Tasks are directly mapped to the cloud if they have a low processing time; otherwise, they go through the fog scheduler. Our suggested architecture keeps track of upcoming requests that will be executed, requests in virtual machines, and virtual resources in resource management for each time interval T scheduler. Therefore, based on scheduler, of 123 these circumstances, the scheduler will make a dynamic choice, such as assigning a job to an existing VM or a new VM. If an existing VM has the storage and processing power to accept new activities, you can add them to it or move on-going work to other VMs. Using a method called DRLMOTS, we were able to organize the jobs in an intelligent manner based on the predetermine circumstances for each time interval T. It will be updating the scheduler with its scheduling choices and take proper care of scheduling tasks into VMs. This scheduler’s main objective is to efficiently map jobs to virtual machines (VMs) depending on their priority while minimizing metrics like makespan, energy usage, and fault tolerance. In order to identify task dependencies, it is necessary to first assess work priorities. As a result, after determining task priorities, the overall load on the VMs must been calculated. Overall load balancing to fog VMs are can be identified the Eq. 10 X loVMq ¼ lon ð10Þ where lon represents the current state running n number of VMs and loVMq indicates the load on the VM queue. When both cloud and fog nodes have completed the workload, workload assigner must decide where to locate the workload. This situation arises when both the cloud and fog computing resources have processed a task, and a decision needs to be made about where the final processed data should reside. This results in the workload requesting access to the processed information in the cloud when more if a task resources are needed. This means that requires additional resources that are not available in the fog node, it can request to access the resources and information available in the cloud. Therefore, a threshold is established and adjusted according to the current set of workloads to optimize the number of effective workloads on the node. This threshold acts as a limit or guideline to determine the resource allocation between the cloud and fog nodes. It is fine-tuned based on the current tasks to ensure that the maximum number of tasks are effectively processed by the fog nodes. The workload assigner needs to set the choice threshold dj;p; for each time period, which is the resource threshold of node J within interval p. this threshold represents the maximum resources that node j can utilize during the time interval p.

  An efficient deep reinforcement learning based task scheduler in cloud-fog environment · 2024 · DOI

• Task scheduling in cloud computing poses a significant challenge due to the diverse range of tasks with varying lengths and runtime capacities. Precisely allocating these tasks to suitable virtual resources is particularly difficult, especially when dealing with computationally intensive and time-sensitive tasks. Thus, we choose for Fog computing as an expansion to schedule heterogeneous jobs that are both time-sensitive and need significant computational resources, depending on the priorities determined at the task manager level. This study introduces a multi-objective task scheduler that prioritizes workloads and virtual machines based on energy efficiency and fault tolerance. The scheduler is coupled with a DQN model, which utilizes proposed reinforcement techniques.

  An efficient deep reinforcement learning based task scheduler in cloud-fog environment · 2024 · DOI

Most-cited papers in IoT and Edge/Fog Computing

• Deep Reinforcement Learning for Task Offloading in Mobile Edge Computing Systems · IEEE Transactions on Mobile Computing · 2020 · 520 citations
• Joint Task Offloading and Resource Allocation for Energy-Constrained Mobile Edge Computing · IEEE Transactions on Mobile Computing · 2022 · 376 citations
• Lyapunov-Guided Deep Reinforcement Learning for Stable Online Computation Offloading in Mobile-Edge Computing Networks · IEEE Transactions on Wireless Communications · 2021 · 365 citations
• Joint Optimization of Service Caching Placement and Computation Offloading in Mobile Edge Computing Systems · IEEE Transactions on Wireless Communications · 2020 · 321 citations
• An Application Placement Technique for Concurrent IoT Applications in Edge and Fog Computing Environments · IEEE Transactions on Mobile Computing · 2020 · 308 citations
• UAV-Assisted Task Offloading in Vehicular Edge Computing Networks · IEEE Transactions on Mobile Computing · 2023 · 269 citations
• Online Distributed Offloading and Computing Resource Management With Energy Harvesting for Heterogeneous MEC-Enabled IoT · IEEE Transactions on Wireless Communications · 2021 · 267 citations
• Edge Computing for Industry 5.0: Fundamental, Applications, and Research Challenges · IEEE Internet of Things Journal · 2024 · 266 citations
• Computation Efficiency Maximization in Wireless-Powered Mobile Edge Computing Networks · IEEE Transactions on Wireless Communications · 2020 · 236 citations
• Artificial Intelligence-Based Sensors for Next Generation IoT Applications: A Review · IEEE Sensors Journal · 2021 · 227 citations

Most recent work

• IoT Service Orchestration in Edge–Cloud Continuum With 6G: A Review · IEEE Internet of Things Journal · 2026
• Adaptive Federated Learning for Future IoV-Oriented IoT End-to-End Network Planning · IEEE Internet of Things Journal · 2026
• D3QN-LMA: A memory-augmented deep reinforcement learning framework for energy-latency tradeoff optimization in mobile edge computing · Advanced Engineering Informatics · 2026
• Time-Dependent Path Selection and Online Learning for Efficient DAG Task Offloading in In-Network Computing · IEEE Internet of Things Journal · 2026
• Revisiting EdgeAI through the lens of communication, storage and computing optimisations · Journal of Edge Computing · 2026
• Unlocking Apple's Private Cloud Compute: An Analysis of Privacy-Preserving Artificial Intelligence · 2026
• Bi-Objective Optimization for Scalable Resource Scheduling in Dense IoT Deployments via 5G Network Slicing Using NSGA-II · Telecom · 2026
• ImmutableShield: A Fragmentation-Driven Security Algorithm for Real-Time Health Data Protection in IoT-Enabled Heat-Stroke Scenarios · International Journal of Drug Delivery Technology · 2026
• Design and Performance Analysis of a Dynamic Channel Allocation Method for IoT Systems with Finite Number of Nodes and Heterogeneous Data Traffics · Sakarya University Journal of Computer and Information Sciences · 2026
• Exploring the Future of IoT and its Integration with Cloud Computing · International Journal on Science and Technology · 2026

Related topics in Computer Science