Traffic sign detection and classification are essential tasks for intelligent transportation systems and road infrastructure monitoring. However, most existing approaches rely on fully supervised object detection models that require large annotated datasets and often struggle with complex scenarios such as panoramic imagery where image sizes are large. In this work, we propose a hybrid pipeline that combines the Segment Anything Model (SAM3) for prompt-based object localization and YOLOv8m-cls for traffic sign classification. Unlike traditional methods, our approach does not require task-specific training for the detection stage, making it particularly suitable for low-data and real-world environments. The use of SAM enables flexible region extraction in high-resolution 360° images. To show the relevance of our proposal, we evaluate our method on a dataset of panoramic images in challenging conditions such as small and distant objects, scene complexity, and geometric distortions. Experimental results demonstrate strong performance, achieving a mAP@0.50 of 94.03%. Thus, this work highlights the potential of combining foundation models with specialized classifiers to develop scalable and practical solutions for real-world traffic sign analysis.

Real-world industrial assets exhibit diverse surface appearances due to multiple factors such as rust, wear, and dirt accumulation. However, simulation environments contain identical or clean object appearances, causing a significant gap between synthetic and real data. Moreover, manually designing unique and intricate realistic textures is inefficient. In this paper, we present a scalable texture synthesis pipeline to produce a multidimensional texture space featuring various industrial object lifespans. It consists of three stages: (1) state translation: to map textures to different states, (2) state interpolation: to generate smooth transitions across different statuses, and (3) space blending: to combine multiple factors affecting object appearances. We evaluate our pipeline with three common industrial assets: dollies, small load carrier boxes, and pallets. As a result, our pipeline generalizes to new shape variants, while preserving realistic and smooth transitions in extreme factors.

Today, the rise of simultaneously flying aircraft increases the constraints on air control systems. As air control systems are essential, they could be targeted by cyberattacks. The Automatic Dependent Surveillance-Broadcast (ADS-B) protocol enables the real-time tracking of aircraft. However, it is known for having serious cybersecurity issues. This is why finding a solution to secure ADS-B is important. This work focuses on the use of machine learning algorithms to detect or correct anomalies in ADS-B time series for low-altitude traffic. A set of four specific algorithms has been developed to face four realistic attack scenarios based on: spoofing, saturation, replay, and interpolated ghosts. The novelty of this work lies in the inclusion of realistic attack scenarios, where detection uses the context of ADS-B time series to enhance common machine learning algorithms.Spoofing attacks are detected using a Convolutional Neural Network (CNN) classifier that identifies incoherent trajectory shapes based on the aircraft type. Saturations are corrected with a Long-Short Term Memory model (LSTM) that verifies the normality of a trajectory by forecasting the future aircraft position. Replays are detected using a hash algorithm. Finally, interpolated ghosts are detected by a classifier trained to recognise overly smooth trajectories. This work has been packaged in a Python library easily usable to ensure anomaly detection in ADS-B data on any source.

The development of the Internet of Things (IoT) and communication technologies has led to the emergence of the multimedia IoT. The multimedia IoT contributes to a variety of multimedia applications, that can capture and transmit large amounts of images or video sequences over an IoT network. Due to restricted resources, multimedia IoT devices must transmit less multimedia data while retaining accuracy. This can save energy and memory use and increase multimedia IoT network performance. This paper proposes an efficient lightweight image compression approach (ELiCA) for Multimedia Internet of Things. The proposed ELiCA starts dividing the captured image into equal-sized blocks and the Discrete Cosine Transform (DCT) is applied to each of them to produce the DCT coefficients. Then, the DCT coefficients are quantized. The output of the quantization process is reordered using zigzag process. Finally, the output files of the zigzag are compressed using the error-bounded lossy compression technique known as SZ developed specifically for High-Performance Computing (HPC) systems. The compressed file will be transmitted by the MIoT device. The proposed ELiCA is evaluated using real sensed images and based on real IoT devices to show the efficiency of the proposed approach. The ELiCA presents better results compared with JPEG in terms of compression ratio, PSNR, and SSIM.

The Internet of Things (IoT) is an essential platform for industrial applications since it enables massive systems connecting many IoT devices for analytical data collection. This attribute is responsible for the exponential development in the amount of data created by IoT devices. IoT devices can generate voluminous amounts of data, which may place extraordinary demands on their limited resources, data transfer bandwidths, and cloud storage. Using lightweight IoT data compression techniques is a practical way to deal with these problems. This paper presents adaptable lightweight SZ lossy compression algorithm for IoT devices (SZ4IoT), a lightweight and adjusted version of the SZ lossy compression method. The SZ4IoT is a local (non-distributed) and interpolation-based compressor that can accommodate any sensor data type and can be implemented on microcontrollers with low resources. It operates on univariate and multivariate time series. It was implemented and tested on various devices, including the ESP32, Teensy 4.0, and RP2040, and evaluated on multiple datasets. The experiments of this paper focus on the compression ratio, compression and decompression time, normalized root mean square error (NRMSE), and energy consumption and prove the effectiveness of the proposed approach. The compression ratio outperforms LTC, WQT RLE, and K RLE by two, three, and two times, respectively. The proposed SZ4IoT decreased the consumed energy for the data size 40 KB by 31.4, 29.4, and 27.3

A main R D pillar in the modern car manufacturing industry revolves around investigating the use of digital assets to train semantic segmentation models in the absence of sufficient real-world data. This study hypothesizes that the procedural generation capabilities of 3DGENie, combined with extensive randomization and realistic sensor simulation, will significantly enhance the effectiveness of synthetic point clouds in training robust semantic segmentation models. To test this, we propose a novel synthetic data generation pipeline called 3DGENie, designed to generate flexible and extensible 3D point clouds tailored for various scenarios and sensor configurations. Leveraging state-of-the-art procedural layout generation, 3DGENie produces region layout trees and applies 3D scene construction and asset randomization to create realistic virtual environments. Synthetic sensors are then used to simulate diverse data capture scenarios, such as colored point clouds and point clouds with reflectance, enabling a comprehensive study of environmental variables’ influence on model performance. Our experiments demonstrate that datasets generated by 3DGENie consistently enhance mean Intersection over Union (mIoU) and mean Accuracy (mAcc) for semantic segmentation models, particularly when real labeled data is limited. In addition to its performance, 3DGENie standouts out compared with existing 3D data generation methods is its flexibility and extensibility to different application scenarios: It is not strictly tied to any specific use case, and can be seamlessly primed to handle different scenarios according to the existing reference data and to the user’s needs. This is emphasized in this study through 3DGENie’s application to two real-world application scenarios using large-scale auto industry data: car assembly lines and autonomous driving. Built on Nvidia Omniverse and Pixar’s Universal Scene Description (USD), the pipeline ensures seamless interoperability across platforms. These results underscore the potential of 3DGENie to address data scarcity challenges, advancing applications from industrial automation to autonomous systems and beyond.

The Manufacturing Message Specification (MMS) protocol is frequently used to automate processes in IEC 61850-based substations and smart-grid systems. However, it may be susceptible to a variety of cyber-attacks. A frequently used protection strategy is to deploy intrusion detection systems to monitor network traffic for anomalies. Conventional approaches to detecting anomalies require a large number of labeled samples and are therefore incompatible with high-dimensional time series data. This work proposes an anomaly detection method for high-dimensional sequences based on a bidirectional LSTM autoencoder. Additionally, a text-mining strategy based on a TF-IDF vectorizer and truncated SVD is presented for data preparation and feature extraction. The proposed data representation approach outperformed word embeddings (Doc2Vec) by better preserving critical domain-specific keywords in MMS traffic while reducing the complexity of model training. Unlike embeddings, which attempt to capture semantic relationships that may not exist in structured network protocols, TF-IDF focuses on token frequency and importance, making it more suitable for anomaly detection in MMS communications. To address the limitations of existing approaches that rely on labeled samples, the proposed model learns the properties and patterns of a large number of normal samples in an unsupervised manner. The results demonstrate that the proposed approach can learn potential features from high-dimensional time series data while maintaining a high True Positive Rate.

This paper describes a new synthetic data generation pipeline called 3DGENie designed to generate 3D point clouds to train deep learning computer vision models. 3DGENie uses procedural layout generation to produce region layout trees. It then applies 3D scene construction and asset randomization to produce scenes populated with 3D assets. Synthetic sensors are placed in the virtual environment to simulate data capture from the 3D scenes as if monitored by real-world sensors. 3DGENie uses Nvidia Omniverse as its scene building platform and Pixar’s Universal Scene Description (USD) for 3D graphics representation to allow for seamless interchange across platforms. Our main application focuses on the generation of industrial car assembly lines, yet 3DGENie can be used across different applications. We conduct experiments to evaluate the generated 3D point clouds, using several deep learning semantic segmentation models. Results highlight the quality of our pipeline.

The current Field-Programmable Gate Array (FPGA) implementation of cryptographic algorithms faces performance and security challenges because these algorithms were not originally designed to take FPGA features into account. One significant performance limitation arises from the iteration of a round function for a high round number, given the fixed structures like static substitution and diffusion primitives throughout the process. This paper introduces a new framework for a key-dependent, flexible one-round stream cipher scheme specifically designed to benefit from FPGA features. It is called RFCA. Security and performance analyses validate the effectiveness and robustness of the proposed solution, ensuring the desired cryptographic properties. In comparison with an AES implementation, RFCA is 34 times faster.

In the fog computing-based Internet of Things (IoT) architecture, the sensor devices represent the basic elements needed to sense the surrounding environment. They gather and send a huge amount of data to the fog gateway and then to the cloud due to their use in various real-world IoT applications. This would lead to high data traffic, increased energy consumption, and slow decisions at the fog gateway. Therefore, it is important to reduce the transmitted data to save energy and provide an accurate decision regarding the safety and health of the building’s environment. This paper suggests an energy-aware data transmission approach with decision-making (EDaTAD) for Fog Computing-based IoT applications. It works on two-level nodes in the fog computing-based TI architecture: sensor devices and fog gateways. The EDaTAD implements a Lightweight Redundant Data Removing (LiReDaR) algorithm at the sensor device level to lower the gathered data before sending it to the fog gateway. In the fog gateway, a decision-making model is proposed to provide suitable decisions to the monitoring staff in remote monitoring applications. Finally, it executes a Data Set Redundancy Elimination (DaSeRE) approach to discard the repetitive data sets before sending them to the cloud for archiving and further analysis. EDaTAD outperforms other methods in terms of transmitted data, energy consumption, and data accuracy. Furthermore, it assesses the risk efficiently and provides suitable decisions while decreasing the latency time.

Generating music-related notations offers assistance for musicians in the path of replicating the music using a specific instrument. In this paper, we evaluate the state-of-the-art guitar tablature transcription network named TabCNN against state-of-the-art computer vision networks. The evaluation is performed using the same dataset as well as the same evaluation metrics of TabCNN. Furthermore, we propose a new CNN-based network named TabInception to transcribe guitar-related notations, also called guitar tablatures. The network relies on a custom inception block converged by dense layers. The TabInception network outperforms the TabCNN in terms of multi-pitch precision (MP), tablature precision (TP), and tablature F-measure (TF). Moreover, the Swin Transformer achieves the best score in terms of multi-pitch recall (MR) and tablature recall (TR), while the Vision Transformer achieves the best score in terms of multi-pitch F-measure (MF). Motivated by the previous insights, we train the networks with more epochs and propose another network named Inception Transformer (InT) to surpass all the estimation metrics of TabCNN using a single network. The InT network relies on an inception block converged by a Transformer Encoder. The TabInception and the InT network outperformed all estimation metrics of TabCNN except the tablature disambiguation rate (TDR) when trained using a bigger epoch size.

With the advent of the Multimedia Internet of Things (MIoT), many image compression techniques have been proposed to address the network’s considerable challenges related to performance and security. However, many MIoT devices, such as the nRF52832 SoC with 64Kb RAM or even less, have significant memory constraints, making conventional methods unsuitable. MIoT networks face considerable challenges related to performance and security due to limitations in the power, computation, and memory of MIoT devices. These limitations result in difficulties in handling high image volumes. Multimedia compression is a potential solution to reduce data size. As MIoT devices often rely on wireless connections, they are also vulnerable to diverse security attacks (passive and active). This work introduces a secure and efficient image crypto-compression technique dedicated to devices having limited memory. It also proposes using denoising and a super-resolution deep learning model to reduce the overhead of the compression process and a lightweight cipher scheme that requires a single round of simple operations to reduce the overhead of the encryption process. The proposed approach effectively addresses the mentioned challenges with minimal overhead on the MIoT device, especially in terms of computational and communication delays, and extensive experimentation underscores its suitability in both effectiveness and robustness.

The electricity market is witnessing an increasingly digital transition and market liberalisation. To support this transition and promote market liberalisation, digital platforms employed in the power energy market must incorporate smart services and enable seamless deployment of these services in proximity to smart meters. The emergence of Compute Continuum and Edge-to-Cloud solutions provide a promising avenue in this regard. This paper discusses such a distributed computing continuum architecture and accompanying services for predicting/planning local household as well as a whole microgrid electric consumption/production. We present two compute continuum strategies for load forecasting in electrical grids: (1) a centralised approach, which involves training a model on a centralised server, and (2) a decentralised approach using Federated Learning (FL). The former approach involves centralising data from multiple sources onto a single server, while the latter distributes the training process across edge devices and preserves data privacy and security. In both cases the inference model is deployed on edge devices close to the collected data. Results show that our suggested FL forecasting model offers privacy-preserving advantages compared to non-private centralised models, with a slight trade-off in prediction accuracy.

Obfuscation techniques (e.g., blurring) are employed to protect sensitive information (SI) in images such as individuals’ faces. Recent works demonstrated that adversaries can perform deep learning-assisted (DL) attacks to re-identify obfuscated face images. Adversaries are modeled by their goals, knowledge (e.g., background knowledge), and capabilities (e.g., DL-assisted attacks). Nevertheless, enhancing the evaluation methodology of obfuscation techniques and improving the defense strategies against adversaries requires considering more "pessimistic” attacking scenario, i.e., stronger adversaries. According to a 2019 article published by the European Union Agency for Cybersecurity (ENISA), adversaries tend to perform more sophisticated and dangerous attacks when collaborating together. To address these concerns, our paper investigates a novel privacy challenge in the context of image obfuscation. Specifically, we examine whether adversaries, when collaborating together, can amplify their DL-assisted attacks and cause additional privacy breaches against a target dataset of obfuscated images. We empirically demonstrate that federated learning (FL) can be used as a collaborative attack/adversarial strategy to (i) leverage the attacking capabilities of an adversary, (ii) increase the privacy breaches, and (iii) remedy the lack of background knowledge and data shortage without the need to share/disclose the local training datasets in a centralized location. To the best of our knowledge, we are the first to consider collaborative and more specifically FL-based attacks in the context of face obfuscation.

As industries increasingly rely on computer vision applications, the demand for robust object detection models grows. Synthetic datasets offer a promising solution by providing diverse and accurate data scenarios, enhancing the performance of these models. In this study, we tackle one aspect of the ongoing challenge to bridge the reality gap between synthetic and real images in industrial applications. We incorporated textures that accurately replicate signs of usage on industrial assets. Further-more, we conducted a comprehensive experiment featuring three distinct datasets from the same simulated environment to evaluate the impact of texture variations on model performance. Our approach introduces a combination of aging texture layers to six industrial assets for object detection. To assess performance, we trained three Deep Learning architectures with these datasets and evaluated them by inferring on real images captured in industrial settings. The results of our experiments demonstrate the effectiveness of the aging textures in real-world scenarios, affirming the success of our proposed approach in enhancing object detection models for industrial applications.

The application of advanced microscopy imaging techniques to 3D specimens or motions faces the problem of the limited depth of field of optical lenses [1]. Thanks to numerical focus computations, digital holography (DH) releases these limitations and extends significantly the allowed axial range of imaging. However, the digital cost of focus distance determination and of object reconstruction makes real-time 3D imaging hardly possible, especially when both in-plane and out-of-plane metrics must be extracted simultaneously.

Success in supply-chain relies, in large part, on good stock management. It is quite simple to guess that there will be an increase in demand for a type of product, or rather reluctance over a period of time, but it becomes complicated to know in advance the exact or optimal number of products to order to avoid stock-outs and at the same time overstocking. This article shows how transactional data can be used with Machine Learning to forecast demand in the retail industry. To train the machine learning models, a sample of 5,115,472 records of receipt data was obtained from the French branch of one of the largest Belgian supermarket chains's data warehouse. The results revealed that the machine learning models manage to learn the seasonality effects and allow to make better predictions.

Deep neural network training on a single machine has become increasingly difficult due to a lack of computational power. Fortunately, distributed training of neural networks can be performed with model and data parallelism and sub-network training. This paper introduces a mathematical framework to study the convergence of distributed asynchronous training of deep neural networks with a focus on sub-network training. This article also studies the convergence conditions in synchronous and asynchronous modes. An asynchronous and lock-free training version of the sub-network training is proposed to validate the theoretical study. Experiments were conducted on two well-known public datasets, namely Google Speech and MaFaulDa, using the Jean Zay supercomputer of GENCI. The results indicate that the proposed asynchronous sub-network training approach, with 64 GPUs, achieves faster convergence time and better generalization than the synchronous approach.