
With the increasing number of real-time applications and internet of things (IoT) devices, the fog layer can help the cloud computing model in reducing workload and reducing delay. Along with the advantages of the fog computing model, we are faced with some challenges, including the issue of service placement, which can affect its performance. To address the service placement challenges, many works have been presented to improve the parameters such as resource management, load balancing, delay, and power consumption. Considering that each of these parameters is important in turn, we should look for a multi-criteria method to improve a number of these parameters together. In this work, a new service placement approach is proposed to simultaneously improve two important metrics of delay and power consumption using the multi-criteria parrot optimization algorithm (MCPOA). In our proposed approach, an attempt has been made to model a multi-criteria problem by designing several single-criteria problems together. Simulation results in the Edgecloudsim environment indicate that the proposed approach, compared to other methods, has reduced each of the delay parameters by at least 6
Anomaly detectors deployed on industrial monitoring systems produce false alarms that existing research reduces in count but does not explain. We formalize false-alarm attribution as a dual-granularity task (variable-level sensor faults vs. mode-level unseen modes) with alarm-resolving interventions, and introduce an injection-based ground-truth protocol that enables controlled evaluation of ten attribution methods across five detectors and five testbeds. Three findings emerge. First, attribution performance depends on the evaluated method–detector–fault-type combination: reconstruction is strongly favoured across 23 isolation-forest units, four Anomaly Transformer units provide suggestive, pairing-specific evidence (a counterfactual advantage on SWaT, 0.93 versus 0.04; near-parity on SMD), and within reconstruction, drift is localized (0.80) while stuck-at faults are missed (0.17). Second, attribution pays off through alarm resolution: mode-directed data addition reduces false-alarm rate by 22.1 percentage points over random controls, replicated in direction on a second detector family (+7.3 pp on PCA, Table S35). Third, the natural-alarm transfer gap is label-efficient: while injected calibration transfers imperfectly (best method 54
In recent years, machine learning has been widely adopted in domains such as healthcare and industrial IoT, while traditional centralized training paradigms rely on aggregated datasets, which can lead to challenges such as data silos and privacy leakage. Federated learning mitigates these issues by enabling decentralized collaborative training without requiring raw data sharing. However, two challenges still remain: (i) the potential risk of privacy leakage during model update sharing, and (ii) significant variations in data quality, which may degrade model performance. To address these challenges, we propose FedAR, a privacy-preserving federated learning framework. In FedAR, distributed Paillier encryption with threshold decryption is employed to protect client model updates and mitigate the risk of single-point key exposure. To reduce the impact of partially low-quality training data, a multidimensional statistical evaluation mechanism is designed to dynamically assess clients, and a progressive control strategy is introduced to effectively exploit useful information from low-quality data. Security analysis shows that FedAR protects individual client updates under the stated threshold-decryption and collusion assumptions. Experiments on CIFAR-10 and CIFAR-100 under IID and Non-IID settings show that FedAR achieves favorable classification accuracy and training stability compared with the evaluated baseline methods.
Multi-step forecasting of CPU utilisation supports cloud scheduling, but a short window of recent history can miss the daily pattern that is relevant to the forecast period. We investigate R-DCRNet, a recurrent forecasting framework that refines an LSTM’s direct 60-step forecast using a separately encoded, forecast-aligned daily sequence. Two proposed configurations share the backbone and correction network: R-DCRNet-S uses a learned static horizon-specific gate, whereas R-DCRNet-A uses a sample-conditioned adaptive gate. On a prospectively frozen, machine-disjoint evaluation of 100 previously unseen machines from Alibaba Cluster Trace v2018, the prespecified primary configuration R-DCRNet-S obtains mean absolute error (MAE) 4.6155 versus 4.6911 for BiLSTM (a 1.61 [-0.0953,-0.0577] ; Holm-adjusted p=2.3× 10^-12 ); it has lower error on 82 of 100 machines, with a favourable direction across all five training seeds. R-DCRNet-A obtains MAE 4.6173; the difference between gating configurations is not statistically resolved and does not establish equivalence. The gain is modest and limited to one workload trace, one machine split, and one 60-step forecasting window. Because R-DCRNet sees more history than the recent-history-only baselines, the improvement reflects the whole added-context pathway, not gating alone.
In the digital age, datasets used in machine learning increasingly exhibit complex characteristics. Traditional decision tree algorithms face several challenges, including difficulty in simultaneously capturing linear and nonlinear feature-class relationships, reliance on manual parameter adjustment, and limited robustness. To address these limitations, this paper introduces an adaptive hybrid feature selection decision tree (AHDT) algorithm with a hybrid feature-selection/splitting mechanism. Adaptive K-means discretization first transforms continuous features, with the number of clusters determined using the silhouette coefficient. DR-based global feature pre-filtering then removes features with low class relevance from the training data. During tree construction, node-specific DR scores and ReliefF-based feature weights are combined through an adaptively weighted divergence ratio criterion to select the splitting feature at each node. Experimental results demonstrate that AHDT achieves the best average rank among the compared single decision tree algorithms on 12 UCI benchmark datasets, providing a competitive and interpretable solution for medium-dimensional classification tasks. The source code is available at https://github.com/233580/AHDT.git. .