Equinix

Grid cybersecurity has a data-sharing problem. Training a good intrusion detection model needs measurements from multiple nodes — but those nodes belong to different operators who will not share raw operational data with each other. Federated learning is the standard answer: train locally, share only model weights, aggregate centrally. The question nobody has answered on real microgrid data is how much detection performance you actually give up by doing it that way. This paper runs that experiment on real measurements from the Mafate isolated microgrid on Réunion Island — 89,179 temporally aligned records split across three simulated federated client nodes. One node gets solar plant measurements. Two nodes get demand-side measurements. Four attack types — scaling, ramp, bias, and replay-bias — are injected at different rates per client, reflecting the non-IID conditions that make federated learning hard. Multi-Layer Perceptron models train locally for five epochs per round, then aggregate via Federated Averaging across ten communication rounds. The federated global model hit 83.79% accuracy, 72.37% F1-score, and 92.97% AUC-ROC. Global F1 grew from 59.76% at round one to 72.37% by round ten — slow, but sure people! The centralized trained on the same data scored an F1 of 99.43%. That gap is the cost of privacy on this dataset, under those conditions — a percentage-point difference of 27.06 when the maximal condition is satisfied. It is not concealed or rationalized — it is the number that this paper was created to quantify.

In light of rising cybersecurity threats, data center providers face growing pressure to protect their own management infrastructure from Distributed Denial-of-Service (DDoS) attacks. While tenant-managed cages generally fall outside the data center's direct security purview, a successful DDoS assault on core provider systems can indirectly disrupt network services. To address this availability assault, the authors developed a Graph Neural Network (GNN) based detection system which leverages Graph U-Nets to automatically classify and mitigate DDoS traffic. Although the model was developed using open-source network flows rather than proprietary data center logs, the model effectively identifies multi-layer DDoS attacks that resemble the malicious patterns threatening modern data centers. Adopting this system to data center environments requires minimal changes to existing operational workflows and processes. Specifically, the GNN based system can be integrated at critical areas within a data center's network infrastructure. Our model achieved an F1 score of over 95

Photovoltaic technology has emerged as a leading renewable energy solution, converting sunlight into electricity through solar cells. Traditional manual inspection and rulebased algorithms struggle with scalability of defects in large photovoltaic datasets. Electroluminescent images often contain thin cracks and subtle anomalies, which are difficult to classify using conventional convolutional networks. To address these limitations, scale-aware dilated lightweight convolutional network (SADL-Net) was proposed for defect classification. The methodology integrates lightweight convolution, dilated convolution and scale-aware attention modules to emphasize defect-relevant features across multiple scales. Experiments were conducted using ELPV and PVEL-AD datasets, containing diverse photovoltaic defects with annotated ground truth bounding boxes. SADL-Net achieved. The proposed SADL-Net model achieved an accuracy of 97.23 % on the ELPV dataset and $\mathbf{9 5. 3 2 \%}$ on the PVEL-AD dataset. Overall, SADL-Net provides a balanced, effective, and statistically validated solution for solar panel defect classification applications globally.

Smart grid controllers trust the sensors that are attached to their grids. By silently spoofing those sensor readings — keeping the corrupted values within normal-looking limits — the state estimator neither sees nor alerts danger, and the grid responds to a fake picture of itself. The False Data Injection problem, and standard chi-squared detectors are oblivious to it. This paper has a different method. We use the differences between what sensors report and what a rolling estimate based on the previous 30 minutes+ of readings says they should report to build features, rather than using raw measurements directly as inputs for classifiers. A 27-D feature space (raw readings, rolling residual and volatility metrics combined) obtained from the eight sensor channels of the real Mafate isolated microgrid on Réunion Island. We inject four types of attacks (scaling, bias, ramp, and replay) on real-measurement copies. Related Articles Here is how smart grid controllers trust sensors that connect to their grids. The static enemy is another high-tech tool: by pretending that the sensor readings are non-detectable — with corrupted values below an undetection, within normal-looking parameters — the state estimator sees no danger, and the grid reacts to a false representation of itself. So the False Data Injection problem that we are trained on (and also standard chi-squared detectors are blind to). This paper does it differently. Instead of feeding raw measurements to classifiers directly, we build features based on the differences between what sensors report and what an estimate calculated from previous readings (typically 30 minutes+) says they should be reporting. A 27-D feature space (raw readings, rolling residual and volatility metrics combined) extracted from the eight sensor channels of the real Mafate isolated microgrid on Réunion Island. We inject four attacks (scaling, bias, ramp and replay) on real-measurements copies.

To prevent the congestion of traffic in the urban areas is become crucial in now a day, because rapid growth of vehicles affect the quality of the urban life style. Traffic flow prediction (TFP) is the solution to overcome the problems in the transportation system. The deep learning (DL) models are used to develop the prediction of the traffic flow but the models are unable overcome the spatial and temporal dependencies. To address this problem, combination of Spatio-Temporal Graph Neural Network (ST-GNN) and Global Spatio-Temporal Network (GSTN) is proposed for prediction of the traffic flow. The GSTN is used to extract the spatio-temporal information from the data and fed into the ST-GNN. ST-GNN is used to learn the extracted spatio and temporal information's from the GSTN for the prediction of traffic flow. The proposed model utilizes two benchmark datasets, which achieved the high performance on PEMS04 and PEMS Bay respectively compared to the baseline models.