Federated Learning (FL) has emerged as a machine learning approach able to preserve the privacy of user’s data. Applying FL, clients train machine learning models on a local dataset and a central server aggregates the learned parameters coming from the clients. The training of the global machine learning model runs without sharing user’s data. However, the state-of-the-art shows several approaches to promote attacks on FL systems. For instance, inverting or leaking gradient attacks can find, with high precision, the local dataset used during the training phase of the FL. This paper presents an approach, called FAST Improved Deep Leakage from Gradients (FAST-iDLG), which is able to improve the inverting gradient attack, considering the spatial correlation that typically exists in a set of images stored in the local FL dataset. Instead of initiating the attack of an image using dummy data, FAST-iDLG blends the previous reconstructed images using a sliding window and an exponential function to weight the blend. The performed evaluation shows an improvement of 66% - 26% in terms of attack success rate and reduces by 55% - 25% the number of iterations per attacked image.
The rapid growth of electric vehicle (EV) adoption has intensified the need for secure, reliable, and interoperable charging infrastructures. The Open Charge Point Protocol (OCPP) 2.0.1 introduces advanced features to support secure and intelligent EV charging. However, critical fields like the state of charge (SOC) still remain highly unexplored and optional in practice. The selective adoption of the SOC within the OCPP protocol introduces risks related to the fair scheduling, grid stability, and billing accuracy. In this paper, we introduce a case study on SOC spoofing to demonstrate the implications of treating this field as optional within the OCPP protocol. The priority hack and the overstay hack are studied as two different types of attacks, where the adversaries manipulate the SOC values in order to gain unfair charging priority or extend the station occupancy, respectively. An autoencoder-based anomaly detection mechanism is designed to identify the spoofing attacks, achieving an F1 score of 87%. Our testing and validation demonstrate that even when only the initial and final SOC values are transmitted, the spoofing can still be detected. Our analysis concludes to the proposal of making the SOC reporting mandatory, at least at session start and end, in order to significantly strengthen the reliability and security of the OCPP 2.0.1-enabled infrastructures.
Recent advances in Machine Learning (ML) are revolutionizing industrial processes, particularly in the context of Industry 5.0, which promotes sustainable and intelligent manufacturing practices such as Zero Waste Manufacturing. However, the application of ML in industrial settings is often constrained by limited labeled data and the high dimensionality of sensor inputs, which hinders both model performance and interpretability. This study addresses the challenges of high-dimensional industrial data by evaluating conventional Feature Selection and Feature Extraction methods, such as ANOVA, Mutual Information, and PCA, as well as more advanced techniques based on Explainable AI (XAI) and Causal AI. In addition to benchmarking traditional methods, we experiment with a novel hybrid feature selection approach that combines SHAP-based feature importance analysis with causal discovery techniques. Experimentation was conducted using a real-world dataset from Carpigiani, a leading manufacturer of ice cream production machines. We aim to predict mixture filling levels to prevent critical freezing events during production. Our results demonstrate that integrating XAI and Causal AI enables the development of more interpretable and reliable ML models, thereby enhancing computational efficiency and decision transparency, a key requirement in Industry 5.0 applications.
In recent years, deep learning models for computer vision applications like YOLO (You Only Look Once) have become widely used for object detection applications in real time. Mobile GPUs such as the NVIDIA Jetson series are highly promising for deploying these models in resource-constrained edge computing platforms. The objective of this work is to examine and compare the performance for object detection (including detection accuracy, model size, execution time, and power consumption) of different versions of YOLO models deployed on different NVIDIA Jetson mobile GPUs using the COCO8 dataset. This allows for evaluation of the four performance metrics, emphasizing very different aspects, across combinations of various YOLO versions, edge devices, and datasets. The experimental results offer important insights into the trade-off among the choices of different YOLO versions and mobile GPUs and their practical potential. Our findings demonstrate that the YOLOv8 models maintain similar accuracy levels across all Jetson devices, where model accuracy achieves up to 0.779 in the COCO8 dataset, underscoring their effectiveness for real-time applications in edge computing. Additionally, our analysis reveals significant variations in power consumption, with the Jetson AXG Orin showing an efficient balance between performance in inference time and energy usage, where the power consumption is 11.05 W and the inference time is 78.1 ms even while running the most complex YOLOv8x model. This study will help researchers and practitioners to gain insight and select the best combination of mobile GPU and YOLO version for future object detection tasks at the edge.
The rapid adoption of Internet of Medical Things (IoMT) devices integrated with the proposed novel idea of Ceiling-Mounted Systems (CMS) significantly enhances healthcare data analytics and patient monitoring. However, this integration presents critical challenges related to data security, privacy, and efficient data transmission. Therefore, this study introduces an innovative cryptographic framework specifically designed to secure data exchange between IoMT devices and CMS infrastructures. The proposed system combines advanced encryption algorithms, circular matrix-based cryptographic methodologies, and deep reinforcement learning-based resource allocation to ensure data confidentiality, integrity, and timely transmission. Simulation results demonstrate that the proposed approach significantly outperforms traditional methods in reducing latency, optimizing energy consumption, and enhancing overall system security.
