GREYC

Communication overhead remains a primary bottleneck in federated learning (FL), particularly for applications involving mobile and IoT devices with constrained bandwidth. This work introduces FedSkipTwin, a novel client-skipping algorithm driven by lightweight, server-side digital twins. Each twin, implemented as a simple LSTM, observes a client’s historical sequence of gradient norms to forecast both the magnitude and the epistemic uncertainty of its next update. The server leverages these predictions, requesting communication only when either value exceeds a predefined threshold; otherwise, it instructs the client to skip the round, thereby saving bandwidth. Experiments are conducted on the UCI-HAR and MNIST datasets with 10 clients under a non-IID data distribution. The results demonstrate that FedSkipTwin reduces total communication by 12–15.5% across 20 rounds while simultaneously improving final model accuracy by up to 0.5 percentage points compared to the standard FedAvg algorithm. These findings establish that prediction-guided skipping is a practical and effective strategy for resource-aware FL in bandwidth-constrained edge environments.

With the continuous increase in satellite launches and the growing complexity of RF signals, traditional methods struggle to fully exploit the rich data derived from passive frequency scanning. In this paper, we propose an innovative approach based on Transformer architectures to detect and characterize satellites. Our method leverages the intrinsic capabilities of Transformers to model both local and global dependencies in high-dimensional spectro-spatial images. By alternating Transformer layers along the frequency and spatial axes, our model extracts robust, invariant representations even in highly complex signal environments. The design incorporates attention mechanisms that facilitate the simultaneous separation and reconstruction of spectral and spatial footprints, proving especially effective in scenarios with densely overlapping signals and interference. Experimental results on both synthetic and real-world datasets demonstrate that our approach significantly enhances detection and characterization performance compared to traditional architectures, paving the way for more precise and resilient space situational awareness systems.

Jamming attacks are among the most critical security threats to Wireless Sensor Networks (WSNs), as they can severely disrupt normal network operations, leading to data loss, network downtime, and reduced system performance. Intrusion Detection Systems (IDSs) have therefore become essential to protect WSNs. However, conventional IDSs often struggle to detect zero-day attacks, creating a significant security gap. To address this, Artificial Intelligence (AI)-based IDSs have been introduced, offering improved detection capabilities but frequently encountering high bias or variance issues, which reduce their reliability. Recently, ensemble learning (EL) has emerged as a promising approach to build more adaptable and data-resilient models by combining multiple learning algorithms. In this context, we propose AdaptiveBoost, an SDN-based Adaptive Ensemble Learning Framework, specifically designed for effective jamming attack detection in WSNs. The SDN integration allows AdaptiveBoost to optimize network traffic flow, identify anomalies in real-time, and adaptively fine-tune detection mechanisms based on current network conditions. We conduct several experiments to evaluate AdaptiveBoost using real-world WSN attacks; using the well-known public network security dataset, WSN-DS, show that AdaptiveBoost outperforms AI-based algorithms in terms of accuracy, precision, recall, and F1 score, while achieving a remarkable reduction in training time by a factor of 235, making it an efficient, scalable solution for securing WSNs against jamming attacks.

• Advantage of photometric stereo for 3D digitization of heritage. • Comparison between photometric stereo and conventional 3D reconstruction methods used in archaeology. • Description of a new method combining different approaches in multiview and multi-lighting context. There are mainly two families of photographic 3D reconstruction techniques. Photogrammetry techniques work according to the principle of triangulation, from the matching of different views, while photometric techniques link the appearance of a 3D point to the orientation of its normal, relative to that of the incident light. While photogrammetry allows to find the global shape of a 3D scene, if it is sufficiently textured, photometric techniques highlight the details of the relief, as long as the model linking the lighting to the shape and reflectance of the scene is sufficiently realistic. In this work, we compare these different approaches with some others in the context of reconstructing archaeological features. After discussing their advantages and disadvantages, we describe a promising new method combining both families in a multi-view, multi-lighting context.

The aim of this study was to assess the effectiveness of a video-based perceptual training programme designed to enhance anticipation skills in professional female football players. The video reference task was used to test participants in pre- and post-tests, where they predicted the ball's arrival location during corners. Participants were evenly divided into an experimental group, which received training in the task between the pre- and post-tests using a progressive temporal occlusion method, and a control group, which received no training. The rate of correct responses, response time, and confidence scores were analyzed in the reference task, as well as performance on corners in real matches, to assess the expected transfer of learning to the field. The results revealed that the experimental group, which underwent training, significantly improved their precision in predicting the ball's landing zone after the intervention, with their accuracy score increasing from 54% to 68% (p < 0.05, η2 = 0.60). Additionally, their response time decreased from 3.2 to 2.4 s (p < 0.05, η2 = 0.48), and their confidence score improved from 3 to 3.8 (p < 0.05, η2 = 0.76). This effect slightly diminished after a 6-month retention interval but remained significantly higher than at the pre-test. Furthermore, we observed that the performance of professional football players during corners in actual matches improved, suggesting a positive effect of the video-based perceptual-cognitive training.

Among brain tumors, glioblastoma (GBM) is the most common and the most aggressive type, and brain metastases (BMs) occur in 20%-40% of cancer patients. Even with intensive treatment involving radiotherapy and surgery, which frequently leads to cognitive decline due to doses on healthy brain tissue, the median survival is 15 months for GBM and about 6 to 9 months for BM. Despite these treatments, GBM patients respond heterogeneously as do patients with BM. Following standard of care, some patients will respond and have an overall survival of more than 30 months and others will not respond and will die within a few months. Differentiating non-responders from responders as early as possible in order to tailor treatment in a personalized medicine fashion to optimize tumor control and preserve healthy brain tissue is the most pressing unmet therapeutic challenge. Innovative computer solutions recently emerged and could provide help to this challenge. This review will focus on 52 published research studies between 2013 and 2024 on (1) the early characterization of treatment efficacy with biomarker imaging and radiomic-based solutions, (2) predictive solutions with radiomic and artificial intelligence-based solutions, (3) interest in other biomarkers, and (4) the importance of the prediction of new treatment modalities' efficacy.

The Psychomotor Vigilance Task (PVT) is widely recognized as the gold standard for measuring vigilance, providing a rapid and objective measure of this state. While driving simulations are also used, they typically require longer administration times. This study examines the sensitivity of driving simulation variables to sleep deprivation throughout the task. The aim is to determine the shorter duration at which performance declines can be observed. A secondary goal is to compare driving simulation and PVT variables' sensitivity in detecting sleep deprivation. Forty-three participants (22 males; aged 46.7 ± 17.8 years) completed a 90-minute driving simulation and a 10-minute PVT under two conditions (normal sleep and partial sleep deprivation of 3.5 hours). Signed-rank Wilcoxon tests and effect sizes were computed for variables from both tasks. Effect sizes were calculated for each 10-minute interval to assess sensitivity over time. All the variables showed sensitivity to sleep deprivation. The largest effect sizes were observed in the driving simulation and specifically for the standard deviation of lateral position (SDLP) (r = 0.73) and the standard deviation of steering wheel movement (r = 0.73). A large effect size for the SDLP (r = 0.71) was observed after only 20 minutes of driving. For the 10-minute PVT, the highest effect size was observed for the number of lapses (r = 0.52). Driving-related variables are highly sensitive to sleep deprivation while providing continuous performance measurements. The SDLP is a particularly sensitive variable even with a reduced driving time of 20 minutes, suggesting that driving simulation tasks can be effectively shortened to 20 minutes.