With the rapid development of the Internet of Things (IoT) and artificial intelligence (AI) technologies, edge computing emerges as a crucial computing paradigm. By processing data near the source, edge computing enables faster and more efficient intelligent services. However, edge devices usually limit computational resources, and existing DNN inference latency profiling methods often rely on internal model details or large-scale latency measurements, making them costly and unsuitable for black-box deployment scenarios. This article proposes ProfEdge, a fast construction framework for deep neural network (DNN) latency profiling models based on Gaussian process regression and Bayesian optimization. ProfEdge adaptively samples real latency measurements under different batch-size states and tunes profiling-model hyperparameters to reduce construction cost while improving profiling accuracy. Specifically, ProfEdge builds an adaptive sampling module based on Gaussian process regression to locate high-error regions through coarse-grained sampling and dynamically refine the sampling process. It further designs a dynamic Bayesian optimization mechanism to improve the accuracy of the latency profiling model. Finally, ProfEdge constructs a cross-device performance mapping model to migrate an existing profiling model to a target device with lightweight stratified calibration, thereby avoiding full reconstruction of the target-device profiling model. Experiments on various edge devices and DNN models, including CNN-based and transformer-based workloads, show that ProfEdge reduces profiling errors by up to 80% and saves over 70% of profiling construction cost compared with existing methods. The cross-device migration results further demonstrate that ProfEdge can achieve competitive profiling accuracy on new devices with only a small number of target-device calibration samples.
In post-disaster or infrastructure-limited remote-area environments where local terrestrial networking infrastructure is unavailable or severely degraded, heterogeneous cloud–edge–ground (CEG) systems provide a viable platform for computation offloading. Ground devices such as IoT sensors and unmanned ground vehicles (UGVs) generate computation-intensive and latency-sensitive tasks, while fixed-wing unmanned aerial vehicles (UAVs) with edge servers offer intermediate processing and relay capabilities to the cloud. We aim to minimize system energy consumption and task latency by jointly considering offloading ratios, CPU frequencies and transmission power of the ground devices. However, efficiently partitioning tasks across these tiers is challenging due to dynamic wireless network conditions, limited onboard resources, and diverse task demands. We propose a temporal dynamics-aware multi-agent offloading and resource allocation framework for such CEG systems in this work. We introduce KAT-MAPPO, a Kolmogorov–Arnold Transformer-enhanced Multi-Agent Proximal Policy Optimization scheme. By embedding Kolmogorov–Arnold Network layers into the transformer encoder, the framework captures long-range temporal dependencies and improves feature representation, while MAPPO enables decentralized offloading under dynamic conditions. Resource allocation is further decoupled and solved analytically via convex optimization. Using a real-world setup, our simulation results show that KAT-MAPPO significantly outperforms existing baselines, achieving faster convergence, a 43% reduction in average task latency in the training evaluation setting, and up to 53.57% latency reduction for the case study scenarios, while also reducing energy consumption. These results highlight the promise of reinforcement learning considering temporal dynamics for efficient offloading in heterogeneous CEG systems.
Encrypted inference under homomorphic encryption (HE) imposes stringent constraints on computation, making the direct deployment of deep neural networks infeasible on resource-constrained edge devices. Existing approaches, such as FHEEDGE and UniHENN, primarily focus on improving the efficiency of encrypted computation through algorithmic and systems-level optimizations. In contrast, we adopt a complementary perspective by redesigning neural architectures to align with the inherent cost model of HE operations. In this work, we present a framework that reformulates HE-aware model design as a constrained neural architecture search problem, where the objective is to identify architectures that are both cryptographically feasible and computationally efficient while preserving task performance. Starting from a pretrained network, we perform a structured, iterative transformation process that includes HE-compatible operator substitutions, depth–width trade-offs, and rotation-efficient design strategies such as successive average pooling. These transformations are guided by constraints on ciphertext slots, multiplicative depth, and rotation complexity, and are evaluated using short-horizon training to efficiently explore the architecture space. We provide a theoretical analysis showing that the transformed network approximates the original model with bounded error under layer-wise Lipschitz propagation, ensuring controlled performance degradation. As a case study, we apply the proposed framework to encrypted crack detection for structural health monitoring. The optimized model achieves an execution time of 86.5 seconds for single-threaded CPU inference in the encrypted domain, compared to an estimated 3.5 hours required by the original encrypted convolutional pipeline. Furthermore, deployment on a Raspberry Pi demonstrates end-to-end encrypted inference within approximately 7.5 minutes. These results highlight that architecture-level optimization, when framed as a constrained search problem, complements existing algorithmic acceleration approaches and provides a practical pathway for enabling efficient and privacy-preserving inference on edge devices across a broad class of sensitive AI workloads.
Fog computing is an emerging computing paradigm suitable for time-sensitive and real-time Internet of Things (IoT) applications. The devices in the Fog environment are autonomous and not exclusively dedicated to Fog application processing. Due to that, the probability of device failure in the Fog environment is high compared with other distributed computing paradigms. Thus, handling failure in Fog is crucial to ensure successful application execution. Several techniques are available in the literature to handle failure, such as checkpointing and task migration, each of which works well in cloud based enterprise applications that mostly deal with static or transactional data. These failure-handling methods are not applicable to highly dynamic Fog environments. In contrast, this work focuses on solving the problem of managing application failure in the Fog environment by proposing a composite solution (combining fuzzy logic-based task checkpointing and task migration techniques with task replication) for failure handling and generating a robust schedule. We evaluated the proposed methods using real failure traces in terms of application execution time, delay and cost. Average delay and total processing time improved by 56% and 48%, respectively, on an average for the proposed solution, compared with the existing failure handling approaches.
Underwater Acoustic Sensor Network (UASN) comprises two types of nodes: anchored nodes and floating nodes. Due to the complex movement of floating nodes and the sparse deployment of anchored nodes, the message dissemination in hybrid UASN is achieved by utilizing the intermittent communication links between nodes (anchored nodes and floating nodes) to constitute the opportunistic communication paths to the sink node. In this work, the historical communication links are recorded and exploited to predict the future communication links, which can help the message holders to select the appropriate relay nodes. In the proposed Message Dissemination method with Lightweight Learning based Link Prediction for hybrid UASN (MD-LLLP), two kinds of learning models are deployed on floating nodes and anchored nodes to perform the local link prediction, respectively. Particularly, a lightweight learning model with low computational complexity is specially designed for floating nodes. Besides, MD-LLLP adopts a distributed message dissemination manner, and thus significantly reduces the communication overhead by avoiding the global exchanges of historical communication links which are locally recorded by nodes. Finally, extensive simulations developed by Python have demonstrated the superior performance of MD-LLLP. Specifically, compared to baseline methods, MD-LLLP achieves an average increase of 19% in delivery ratio, reduces the average delivery delay by 6.1 seconds, and reduces the number of message copies by 16%. By MD-LLLP, the data messages can be disseminated to the sink node quickly, making the delivery ratio of data messages enhanced and the delivery delay of data messages shortened.
