
Cloud–edge collaboration can keep time-sensitive predictions local while reserving cloud resources for later model updates, but the respective benefits of distillation and selective feedback under changing conditions remain unclear. We examine these questions using a reproducible synthetic flight-decision task with 36 inputs and three normalized outputs. A cloud-side teacher guides a 643-parameter edge student trained with noisy labels. An initial three-student ensemble scores states for asynchronous upload, while current predictions do not wait for cloud responses. We evaluate distillation across seven paired random seeds and compare uncertainty-triggered updating with random selection at similar upload volume under gradual shift. In the noisy-label setting, distillation reduces the test mean squared error (MSE), averaged across seeds, from 0.008364 to 0.004409. One update using clean labels supplied by the simulator reduces mean final-shift error from 0.008634 to 0.007050, although uncertainty-based selection shows no clear advantage over the upload-matched random policy. These findings provide a controlled simulation baseline for examining model-training and feedback effects and can guide subsequent evaluation with real flight data, measured device timing, and closed-loop tests.
The increasing adoption of the Internet of Things (IoT) is fuelling the expanded use of data pipelines (DPs) for data processing within Edge, Fog, and Cloud Computing Continuum infrastructures. Modern DPs, represented as directed acyclic graphs, consist of heterogeneous steps that process data to generate business value. In practice, DPs cannot be benchmarked in advance, which prevents the use of traditional scheduling approaches. To address this, we introduce the context-aware timeline scheduler (CATS), a novel approach that leverages pipeline context through dry runs with low-scale input volumes (as opposed to using benchmark data) to schedule data pipelines. CATS aims to optimize makespan and cost for large-scale deployments on Computing Continuum resources. Dry runs consist of consecutive testing deployments of pipeline steps on the target infrastructure, providing measurements of CPU, memory usage, and execution time. Based on these dry runs, data locality information and pipeline structure, CATS predicts the resource requirements, time of execution, and level of horizontal scaling for each step for larger total input volumes. Using the predictions, it computes a timeline of scheduling events to be executed when orchestrating the pipeline deployment on target infrastructure. We evaluated CATS using an industrial spot welding data preparation pipeline for downstream AI/ML-based predictive analytics, which was tested on real infrastructure that emulates a Fog/Cloud Continuum. The experimental results show accuracy for both predicting resource usage and makespan for target volumes that are up to two orders of magnitude higher than what was measured in the dry-run data.
Cloud computing is the paramount model of high-performance computing. Because of its contingent and dynamic nature, demand for computing resources varies but remains finite. Cloud computing platforms are widely used for executing large-scale scientific workflows. However, efficient workflow scheduling remains a major challenge due to task dependencies, heterogeneous resources, and unequal workload distribution. These problems frequently lead to higher makespan and ineffective resource utilization. Resource management in cloud platforms is a key area of research that consists of two core components: scheduling dependent or independent tasks and provisioning resources for these tasks. Therefore, managing the ratio of demand and resources can be controlled by scheduling. It is a method for efficiently provisioning resources to users according to their demands, which can enhance system performance. Here, the demands are the user’s requests that can be considered as a task, and multiple tasks refer to the load that directly affects the processing capability of the system. This scheduling is also known as workflow scheduling, and it is depicted by a Directed Acyclic Graph (DAG) and is based on dependent tasks. This paper is based on the dependent task of DAG and contributes new approaches for workflow scheduling using graph shortest distance heuristics to achieve two major objectives: makespan and load balancing. This paper considers a random workflow and two scientific workflows: Epigenomics and Montage, for the assessment and evaluation of the proposed algorithm. Experimental results demonstrate that the proposed method consistently outperforms existing state-of-the-art scheduling algorithms in terms of reduced makespan by ranging from 49
A core challenge in anomaly detection for Internet of Things (IoT) scenarios lies in the extreme scarcity of labeled samples in real-world deployments. Most existing methods rely on unsupervised paradigms to learn normal behavior patterns of IoT devices. However, such approaches often define overly strict boundaries for “normality,” making minor fluctuations prone to be misjudged as anomalies and resulting in high false positive rates, which severely undermines the reliability and operational efficiency of cloud-side monitoring systems. To this end, an improved self-supervised task learning framework is proposed specifically for IoT anomaly detection. In our work, an adaptive noise injection mechanism is designed for data augmentation, which automatically adjusts noise intensity according to the dynamic characteristics of input signals to simulate diverse anomaly patterns that closely resemble real-world scenarios. We then propose a joint learning enhanced paradigm that integrates contrastive learning with a parallel reconstruction task. Specifically, the InfoNCE loss is introduced to enhance the efficiency of contrastive learning in extracting discriminative features, while a novel loss function is devised to jointly optimize the two tasks. The resulting representation learning model maintains high sensitivity to anomaly deviations while preserving strong generalization capability to normal behavioral patterns, making it well-suited for the complex and dynamic IoT environment under cloud-edge collaboration. Finally, a classifier based on neighbor/non-neighbor representations is built to perform self-supervised anomaly discrimination. Comparative experimental results on five public IoT monitoring datasets, including MSL, SMAP, SMD, SWAT, and WADI, demonstrate that the proposed method achieves state-of-the-art F1-score and AU-PR on four datasets, with average improvements of 4.65
Software-Defined Networking (SDN) offers flexibility and programmability but also exposes the control plane to Distributed Denial-of-Service (DDoS) attacks that can degrade network performance. Existing intrusion detection frameworks often suffer from class imbalance, feature redundancy, and limited temporal modeling, which reduce their effectiveness in detecting diverse attack patterns. To address these challenges, we propose a novel pipeline that integrates Synthetic Minority Oversampling Technique (SMOTE) for balancing, Hybrid Crayfish–Gazelle Optimization (HCGO) for optimal feature selection, and a Temporal Hybrid Attention Recurrent Network (THARN) that combines LSTM, GRU, Temporal Convolutional Networks (TCN), and attention mechanism for robust classification. In addition, a reinforcement learning–based policy module, Q-Mitigate is specified to bridge detection and adaptive SDN mitigation. Experimental evaluation on the CIC-DDoS2017 dataset demonstrates that THARN achieves 99.38