The rise of connected and autonomous vehicles introduces significant cybersecurity challenges for embedded systems. One of the most vulnerable components is the Controller Area Network (CAN), which manages communication between a vehicle’s electronic units. Originally designed without built-in security mechanisms, the CAN bus is particularly susceptible to message injection attacks. This paper presents a hybrid anomaly detection framework that combines traditional models like Support Vector Machines (SVM) with recurrent neural networks such as Long Short-Term Memory (LSTM). We also evaluate additional algorithms, including Random Forest, Isolation Forest, and Autoencoders. The proposed architecture leverages the LSTM to extract temporal features from CAN traffic and uses the SVM for precise classification, balancing dynamic detection capabilities with real-time efficiency. Experimental results demonstrate that the hybrid model outperforms individual approaches in terms of precision, recall, F1-score, and Area Under the Curve (AUC), while also reducing false positive rates and increasing robustness. This makes the proposed framework a promising solution for enhancing the cybersecurity of modern in-vehicle networks.

Understanding driver behavior is essential for improving road safety and informing policy decisions. In this study, we use clustering techniques to classify drivers into behavioral groups based on personal attributes (such as date and place of birth), driving history (including traffic code violations), and license details (such as the age at which the license was obtained and the type of vehicle driven). Our approach applies KMeans clustering in a hierarchical manner to identify meaningful patterns within the data. We also used large language models to later facilitate interpretation of the clusters. And due to the large size and sensitive nature of the dataset, all processing is conducted within the organization’s secure data platform. The findings of this study could support the development of more effective traffic regulations, insurance models, and risk assessment strategies.

Multi-label classification presents a complex computational challenge with broad applications in text categorization, image annotation, and bioinformatics. In this paper, we introduce a pioneering approach that merges Deep Extra-Trees with three transformation methods to tackle this intricate task. Through comprehensive evaluations conducted on a range of benchmark datasets, we meticulously compare our method against established algorithms. The results not only validate our approach but also reveal its superiority, demonstrating enhanced performance and robustness. Our approach utilizes transformation methods of multi-class classification in conjunction with Deep Extra-Trees. Specifically, we implement Binary Relevance, Classifier Chains, and Label Powerset as our transformation methods, which effectively convert the multi-label problem into multiple single-label problems, thereby leveraging the power of Deep Extra-Trees for improved prediction accuracy. This substantiates the robustness and adaptability of our proposed methodology across diverse datasets. By offering a compelling solution to the multi-label classification problem, our research contributes significantly to the advancement of machine learning techniques in various domains. Moreover, we apply this approach not only to classification tasks but also to anomaly detection, further demonstrating its versatility and practical utility.

In this paper, we propose an innovative architecture for anomaly detection in videos, motivated by the need to answer quickly to danger in monitoring streams, without requiring expensive computational power. Drawing inspiration from human behavior our approach integrates spatial and temporal analyses. For the temporal analysis, which classifies video sequences, we associate a recurrent convolutional network combining Visual Geometry Group Net 19 (VGG19) and Gated Reccurrent Units (GRU), with a Multilayer Perceptron (MLP). Simultaneously, the spatial analysis of individual images is conducted through You Only Look Once version 7 (YOLOv7). Then, both predictions are combined to perform the final prediction, where an anomaly is signaled if a perceived suspicious object or unexpected action occurs on the screen. Our experimental results shows the integration of both approaches reduces the rate of false negatives, leading to improved identification of anomalous events within video streams for both binary and multi-class models. We also show that multi-class models are less suited for this task than binary models.

In this paper, we propose a novel approach for building and initializing deep neural networks based on extremely randomized trees (extra-trees) an ensemble learning method for both classification and regression and feature extraction techniques. We use convolutional neural networks (CNNs), a family of modern deep learning models, extensively used in the area of computer vision and image classification, to improve the accuracy and generalization performance of classifiers. First, a CNN model is built to automatically extract multi-level features from the data. Second, a random forest obtains the structures of the trees. Finally, the neural networks (MLP) are built. This hybrid method combines two standard adaptive methods: decision trees and artificial neural networks. In this article, we illustrate the structure of the hybrid method, the problems occurring during the building of the model, and the solutions for these problems. The experimental results indicate that the proposed approach achieves consistently high performance for a variety of regression and classification tasks. These results should motivate further studies seeking to develop accurate and efficient tree-based models.

This paper studies the problem of applying artificial intelligence (AI) methods to the detection of financial Suspicious Activity Reports (SARs). Financial institutions use semi-automated systems based on predetermined rules to identify suspicious activities, but most of the generated alerts are false positives and do not lead to a reporting to the authorities. From the initial investigation based on a predefined triggering scenario, to the decision of a case worth reporting to the authorities, critical parts of the process are carried out by human analysts. Part of a collaborative project with HSBC, the main objective of this research is to automate the identification of truly suspicious transactions by reducing the need for human intervention, as well as limiting the number of false alerts, using Neural Networks (NNs). We implement a prediction framework based on a combination of two unsupervised NNs used in an online learning mode and we compare the performances of our solution against other Machine Learning (ML) models.

This paper studies the problem of applying artificial intelligence (AI) methods to profiling financial crime behavior, more specifically money laundering in a context of environmental trafficking. The main objective is to provide a tool based on Neural Networks (NNs) for identification and analysis of suspicious transactions. This study is part of a collaborative research project with HSBC and we paid particular attention to real-world financial industry constraints, like real-time fast processing and explainability.

Banks are compelled by financial regulatory authorities to demonstrate whole-hearted commitment to finding ways of preventing suspicious activities. Can AI help monitor user behavior in order to detect fraudulent activity such as identity theft? In this paper, we propose a Machine Learning (ML) based fraud detection framework to capture fraudulent behavior patterns and we experiment on a real-world dataset of a major European bank. We gathered recent state-of-the-art techniques for identifying banking fraud using ML algorithms and tested them on an abnormal behavior detection use case.

This paper presents DyCorC, an extractor and cleaner of web forums contents. Its main points are that the process is entirely automatic, language-independent and adaptable to all kinds of forum architectures. The corpus is built accordingly to user queries using expressions or item keywords as in research engines, and then DyCorC minimizes the boilerplate for further feature-based opinion mining and sentiment analysis, gathering comments and scorings. Such noiseless corpora are usually hand made with the help of crawlers and scrapers, with specific containers devised for each type of forum, entailing lots of work and skills. Our aim is to cut down this preprocessing stage. Our algorithm is compared to state of the art models (Apache Nutch, BootCat, JusText), with a gold standard corpus we released. DyCorC offers a better quality of noiseless content extraction. Its algorithm is based on DOM trees with string distances, seven of which have been compared on the reference corpus, and feature-distance has been chosen as the best fit.

International audience

We present an experimentation line that encompasses various stages for research on graphemes distribution and unsupervised classification. We aim to help close the gap between recent research results showing the abilities of unsupervised learning and clustering algorithms to detect underlying properties of phonemes and the present possibilities of Unicode textual representation. Our procedures need to ensure repeatability and guarantee that no information is implicitely present in the preprocessing of data. Our approach is able to categorize potential graphemes correctly, thus showing that not only phonemic properties are indeed present in textual data, but that they can be automatically retrieved from raw-unicode text data and translated into phonemic representations. By the way, we observe that SOM algorithm copes well with very sparse vectors.