01 Lab · R&D

Our scientific publications

Our teams publish, prototype and drive innovation.

14 publication(s)

Springer Pattern Analysis and Applications
IA & Vision Revue
2025

FORT-RAJ: a fisheye-optimized deep learning model for real-time trajectory prediction

… 24 showcases the model’s performance in a previously unseen room within our Caplogy premises, providing insight into its generalization ability to generalize across different spatial …

S Bouzayane, M Kahouadji, B Magnier

Read the publication →
IEEE 2025 9th Cyber Security in Networking Conference (CSNet)
Cyber Conférence
2025

Multi-level Architecture for IoT-based Android Ransomware Detection and Family Attribution

En raison de la prolifération de l'Internet des objets (IoT), les appareils IoT basés sur Android sont de plus en plus ciblés par des menaces de ransomware sophistiquées. Pour répondre à ce problème, nous proposons une architecture hiérarchique pour la détection des ransomware et l'attribution de famille qui utilise uniquement des séquences d'appels d'API extraites d'applications Android. L'architecture est conçue pour analyser un flux continu d'applications. Elle se compose de trois niveaux séquentiels : (1) la détection de malware pour distinguer les applications bénéfiques des malware, (2) la détection de ransomware pour différencier les ransomware des autres types de malware, et (3) l'attribution de famille de ransomware pour attribuer les ransomware détectés à leur famille réelle. Cette architecture est novatrice car elle diffère des approches à un seul niveau proposées dans la littérature et prend en compte tous les types d'applications, y compris les applications bénéfiques, les ransomware et les autres types de malware. En outre, nous avons évalué de manière approfondie l'architecture sur le jeu de données CICAndMal2017 en combinant différents modèles et méthodes de prétraitement des données, ce qui permet d'évaluer un total de 135 flux de travail d'apprentissage automatique sur tous les niveaux. Les résultats de l'évaluation montrent que des précisions de 92 % et 100 % ont été obtenues pour la détection de ransomware et l'attribution de famille de ransomware, respectivement.

I Gharbi, A Agarwal, A Derhab

Read the publication →
IEEE 2025 9th Cyber Security in Networking Conference (CSNet)
Autre Conférence
2025

Hybrid LSTM-BiLSTM Approach for DDoS Attack Forecasting

Distributed Denial of Service (DDoS) attacks remain a critical threat to the stability and security of networked systems. To address the challenge of predicting such attacks in advance, …

Y Yakhlaf, R Bekkouche, H Garbouge

Read the publication →
IEEE
IA & Vision Revue
2024

Visual-HAR: Un modèle de reconnaissance d'activités humaines en temps réel basé sur des capteurs visuels is translated to Visual-HAR: A Real-Time Human Activity Recognition Model Based on Visual Sensors

Le document présente un modèle de reconnaissance d'activités humaines, Visual-HAR, qui utilise des capteurs visuels pour identifier les activités des personnes âgées. L'objectif de Visual-HAR est de détecter avec précision et rapidité si une personne âgée est en danger en reconnaissant ses activités en temps réel. Ce modèle utilise des réseaux de neurones convolutionnels (CNN) et se concentre sur l'amélioration du modèle ConViVit grâce à l'optimisation des hyperparamètres. Les performances des modèles Visual-HAR et ConViVit ont été évaluées à l'aide de trois jeux de données d'actions humaines : NTU RGB+D60, PKU-MMD et PKU-MMD2. Les critères d'évaluation comprenaient à la fois la précision et le temps d'exécution. Les résultats ont indiqué que Visual-HAR surpassait considérablement ConViVit avec une amélioration moyenne des performances de test de 7,41 % et une réduction du temps d'exécution d'au moins 36 %, ce qui est crucial pour la prise de décision en temps réel. Ces résultats mettent en évidence le potentiel de Visual-HAR en tant qu'outil fiable pour la surveillance des personnes âgées.

A Sabrina, B Sarra, M Khawla

Read the publication →
IEEE
Cyber Conférence
2024

Anticiper les menaces cybernétiques : Approches d'apprentissage profond pour la prévision des attaques DDoS

This paper examines the application of deep learning (DL) techniques for forecasting DDoS attacks (i.e., predicting future attacks in advance). We used a uniform modeling approach that enabled us to compare the performance of various DL algorithms, specifically Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRU). The results indicate that both short-term (a few seconds ahead) and mid-term (up to 20 seconds ahead) DDoS traffic prediction with very high accuracy are possible. Additionally, the results show that the LSTM model outperforms both RNN and GRU, as well as machine learning algorithms such as Support Vector Machines (SVM), Random Forests (RF), and k-Nearest Neighbors (KNN).

DB Ali, M Belaoued, S Dawaliby

Read the publication →
IEEE
IA & Vision Revue
2024

Fort-raj: a hybrid fisheye model for real-time pedestrian trajectory prediction

This paper introduces FORT-RAJ, a hybrid model designed for pedestrian trajectory prediction in the context of top-view fisheye images. To achieve this, FORT-RAJ merges the FORT (Fisheye Online Realtime Tracking) algorithm, which tracks people using fisheye cameras without prediction capabilities, with the GATraj model, known for trajectory prediction but not yet adapted for fisheye images. The proposed method, FORT-RAJ, is designed to detect pedestrians, track their trajectories, and predict their future positions. It leverages the wide field of view of fisheye cameras while addressing the distortions inherent in such images. The experiments demonstrated that the FORT-RAJ model performs satisfactorily on fisheye images, achieving an Average Displacement Error (ADE) of 0.38 meters and an Final Displacement Error (FDE) of 0.42 meters.

Y Amrouche, S Bouzayane

Read the publication →
Elsevier Procedia Computer Science
Cyber Conférence
2024

Parca: proactive anti-ransomware cybersecurity approach

Modern ransomware implement innovative techniques and tactics to bypass existing security measures. Hence, predicting and forecasting such kind of malware is crucial for enhancing the overall cybersecurity posture in an increasingly digital and interconnected world. In this paper, we propose a proactive approach for predicting ransomware attacks. We integrated a dynamic deep learning algorithm for analyzing memory-based features. This allows us to detect the existence of ransomware indicators regardless of the usage of obfuscation techniques, such as code encryption. Thus, offering advantages in terms of obfuscation resistance, realtime insights threat analysis and adaptability to evolving ransomware threats landscape. Experimental results using recent datasets demonstrate the effectiveness of the proposed approach in identifying modern ransomware samples with a weighted average of 99.98%, 99.96%, 100%, and 99.98% for accuracy, precision, recall and fl-score respectively.

A Djenna, M Belaoued, N Lifa, DE Moualdi

Read the publication →
IEEE
IA & Vision Revue
2024

Real-Time Human Activity Recognition Using Wi-Fi CSI: A Deep CNN Architecture Approach

Human activity recognition (HAR) using Wi-Fi signals has gained significant attention due to its non-invasive nature, ubiquity, and respect for privacy, in contrast to camera-based systems. This study aims to develop a robust deep learning model for real-time HAR using Channel State Information (CSI) from Wi-Fi signals. We propose a deep Convolutional Neural Network (CNN) architecture incorporating an attention mechanism to process CSI data and classify various human activities. Our model was trained and tested using a comprehensive dataset of Wi-Fi CSI, employing advanced data augmentation techniques to enhance model generalization. The proposed model achieved an impressive accuracy of 99.69%, significantly outperforming existing approaches in terms of precision, recall, and F1 score. These findings demonstrate the potential of using Wi-Fi CSI with deep CNNs and attention mechanisms for accurate and real-time HAR, paving the way for applications in smart homes, healthcare monitoring, and other domains requiring efficient and privacy-preserving activity recognition. Future work will explore adapting this methodology to a variety of datasets and integrating an emotion recognition system to further expand its capabilities.

M Karim, S Imène, G Aymen

Read the publication →
IEEE
IA & Vision Revue
2024

Comparative analysis of conventional object detection models on fisheye images

Fisheye cameras are increasingly used for object detection and tracking across various sectors. However, fish-eye images present distortions that complicate their processing compared to flat images. Two main approaches exist to tackle this challenge: calibrating fisheye images to convert them into flat images before applying traditional image processing models, or directly employing these models on fisheye images. In this study, we investigated both approaches using two fisheye image databases. Our findings indicate that calibration results in information loss and inadequate time savings. Furthermore, traditional models, despite optimization efforts, exhibit limited performance of approximately 20%, emphasizing the necessity for developing specific algorithms tailored to this technology.

B Sarra, A Arezki, M Houssam

Read the publication →
IEEE
IA & Vision Revue
2023

ConViViT-A Deep Neural Network Combining Convolutions and Factorized Self-Attention for Human Activity Recognition

The Transformer architecture has gained significant popularity in computer vision tasks due to its capacity to generalize and capture long-range dependencies. This characteristic makes it well-suited for generating spatiotemporal tokens from videos. On the other hand, convolutions serve as the fundamental backbone for processing images and videos, as they efficiently aggregate information within small local neighborhoods to create spatial tokens that describe the spatial dimension of a video. While both CNN-based architectures and pure transformer architectures are extensively studied and utilized by researchers, the effective combination of these two backbones has not received comparable attention in the field of activity recognition. In this research, we propose a novel approach that leverages the strengths of both CNNs and Transformers in an hybrid architecture for performing activity recognition using RGB videos. Specifically, we suggest employing a CNN network to enhance the video representation by generating a 128-channel video that effectively separates the human performing the activity from the background. Subsequently, the output of the CNN module is fed into a transformer to extract spatiotemporal tokens, which are then used for classification purposes. Our architecture has achieved new SOTA results with 90.05 %, 99.6%, and 95.09% on HMDB51. UCF101. and ETRI-Activity3D respectively.

DR Reda, F Chaieb, H Drira

Read the publication →
Publication
Data & Systèmes Thèse
2023

Système de détection de la fraude financière à l'aide d'approches et de techniques d'intelligence artificielle

La fraude financière est devenue un défi majeur à l'ère numérique moderne, posant des menaces à la stabilité économique, à la confiance des clients et à l'intégrité des systèmes financiers. Les approches traditionnelles basées sur des règles et guidées par des experts pour la détection de la fraude ont montré des limites dans la résolution de la nature dynamique et évolutive des activités frauduleuses. Cette thèse présente une exploration approfondie d'un nouveau paradigme qui combine la puissance de l'Intelligence Artificielle (IA) basée sur la Connaissance et des techniques d'Apprentissage Automatique (ML) pour une détection efficace et robuste des fraudes financières.L'objectif principal de cette recherche est double : d'une part, examiner la viabilité de l'utilisation des méthodes ML et DL pour détecter la fraude financière et d'autre part, évaluer l'efficacité d'une approche basée sur l'IA basée sur la Connaissance pour traiter les subtilités du paysage frauduleux en évolution. La recherche commence par proposer deux hypothèses : premièrement, que les méthodes ML et DL peuvent identifier efficacement les schémas frauduleux et deuxièmement, qu'une approche basée sur l'IA basée sur la Connaissance offre un cadre flexible pour intégrer des connaissances spécifiques au domaine pour améliorer la détection de la fraude.Pour tester ces hypothèses, un cadre expérimental approfondi est utilisé, impliquant divers algorithmes ML tels que XGBoost, Random Forest, les réseaux neuronau

A Hussaini

Read the publication →
Springer Security Informatics and Law Enforcement
IA & Vision Revue
2023

Deep learning for windows malware analysis

Malwares, such as ransomware, Trojans, spyware, and botnets, are the most common cyber-threats that can cause significant damages for organizations, governments, and individuals. …

M Belaoued, A Derhab, N Chekkai, C Ramdane

Read the publication →
Springer Lecture notes in computer science
IA & Vision Conférence
2023

Cyber-attack Proactive Defense Using Multivariate Time Series and Machine Learning with Fuzzy Inference-based Decision System

… belaoued@ caplogy. com 4 Caplogy, Poissy, France … The authors would like to thank Caplogy for supporting this work, which is the result of their collaboration in 2022. …

M Belaoued

Read the publication →
AIRCC · CS&IT Artificial Intelligence and Applications
Autre Revue
2022

The NP-completeness of quay crane scheduling problem

This paper discusses the computational complexity of the quay crane scheduling problem (QCSP) in a maritime port. To prove that a problem is NP-complete, there should be no polynomial time algorithm for the exact solution, and only heuristic approaches are used to obtain near-optimal solutions but in reasonable time complexity. To address this, first we formulate the QCSP as a mixed integer linear programming to solve it to optimal, and next we theoretically prove that the examined problem is NP-complete.

A Skaf, S Dawaliby, A Aberkane

Read the publication →