Program

Keynote 1: Energy Efficiency in Resource-Constrained Networks 
By:   Prof. Abdelmadjid Bouabdallah (University Of Technology Of Compiègne, France).

Keynote 3: In the Era of Generative AI and Machine Learning: Challenges and Opportunities 
By:   Prof. Mustapha Lebbah (Versailles University, France)

Keynote 4: Hybridization of Optimization and Artificial Intelligence: Toward Advanced Decision-Making Methods
By:   Prof. Issam Nouaouri (Artois University, France) Online

S1.1: Combined StyleGAN2-ADA and a Convolutional Neural Networks for Apple Leaf Disease Classification on a realistic dataset.

Mohamed Reda LAKEHAL (ENSB)*, Youcef FERDI (ENSB) 

Abstract. The size of datasets can significantly impact the performance of deep learning models used in computer vision, which is currently widely applied in precision agriculture. To address this issue, various data augmentation techniques based on different approaches have been developed. Among these, deep learning–based data augmentation methods have demonstrated superior performance compared to traditional approaches.  In this paper, we explore the impact of using a Generative Adversarial Networks with Adaptive Discriminator Augmentation (StyleGAN2-ADA) to improve the performance of deep learning models used for plant leaf disease classification. To this end, we used a StyleGAN2-ADA to generate synthetic images to increase the size of the original Plant Pathology dataset, which was collected from natural environments. The focus is on apple leaf images classified into three categories: healthy, rust, and scab. This study combines the use of StyleGAN2-ADA, implemented in PyTorch, to gradually augment the dataset size with the application of convolutional neural networks (CNNs) for classifying apple leaf diseases. Experiments were conducted using transfer learning models, including VGG16, InceptionV3, MobileNetV2, and Xception. Based on evaluation metrics such as accuracy, precision, recall, and F1-score, the results show slight overall improvements as the dataset size increases through the use of synthetic apple leaf images, although some performance fluctuations are observed even with this augmentation. While InceptionV3, MobileNetV2, and Xception all demonstrated high performance, MobileNetV2 achieved the best overall results, with accuracy rising from 88.19% to 93.53% and loss decreasing from 42.50% to 22.63% after dataset augmentation, all while requiring less computational time.

S1.2: Toward Smarter Traffic Signals: Swarm Intelligence-Based Reinforcement Learning Approaches.

Sarah BENHAFID (University Batna 2)*, Djalal HEDJAZI ( University Batna 2)

Abstract. Urban road networks are increasingly affected by congestion, which creates a demand for Traffic Signal Control (TSC) schemes that can adjust their behaviour in real time. A growing body of work addresses this problem with hybrid solutions that couple Swarm Intelligence (SI) metaheuristics and Reinforcement Learning (RL) in order to improve efficiency, adaptability, and robustness. 
This paper reviews and organizes existing hybrid SI–RL approaches for TSC. We analyse how different combinations of bio-inspired optimization and learning affect convergence properties and online decision making. The survey compares representative algorithmic frameworks, summarizes their reported performance, and proposes a three-category taxonomy of SI–RL integration mechanisms — covering policy initialization, exploration guidance, and meta-level parameter tuning — as a structured basis for understanding and designing next-generation traffic signal controllers.

S1.3 Traffic Sign Detection for Algerian ADAS: From Data Collection to Model Deployment

Oumaima Daif (ensia)*, hayet saadi (ensia), aicha boutorh (ensia)

Abstract. Traffic sign detection is fundamental for Advanced Driver Assistance Systems (ADAS), requiring high accuracy and real time performance. This paper presents a comprehensive methodology for developing a region-specific traffic sign detection system adapted to Algerian road conditions, with emphasis on data collection protocols, annotation workflows, and systematic model evaluation. Starting from a custom dataset of 2,000 images in Phase I, we progressively expanded to 9,484 images in Phase II through controlled frame extraction and extensive augmentation strategies. We conducted systematic experiments to determine optimal frame extraction rates (2.5–3 FPS), evaluated multiple annotation platforms, and compared modern YOLO architectures (YOLOv5 through YOLO12). The final system achieves 91.5% mAP@50 with 2.3ms inference time using YOLO11n, demonstrating the effectiveness of methodical data collection and progressive model refinement. This work contributes quantified best practices for dataset construction and provides empirical evidence for deployment oriented design decisions in real world Algerian ADAS applications

S2.1: Robust Machine Learning under Incomplete Data: A Comparative Study of Imputation Strategies.

Amira BENDRIMIA (University of 8 May 1945 Guelma)*; Ali KHEBIZI (University of 8 May 1945 Guelma) 

Abstract. Missing data (MD) is a common and critical issue in realworld datasets that can significantly degrade the performance and reliability of Machine Learning models. Incomplete observations may introduce bias, reduce predictive accuracy, and affect model generalization. Therefore, selecting appropriate MD management techniques is essential
for building robust learning systems. This study investigates the effectiveness of different MD handling strategies in supervised Machine Learning tasks. We compare simple statistical imputation with state-of-the-art imputation baselines under varying missingness rates. The impact of each strategy is evaluated using standard performance metrics to assess predictive accuracy and robustness.
The experimental analysis highlights that model performance is highly sensitive to both the MD mechanism and the chosen imputation method, especially as the proportion of missing values increases. The findings provide practical recommendations for selecting suitable MD treatment techniques to improve the robustness of Machine Learning models when dealing with incomplete datasets.
 

S2.2: Ensemble Fault Localization Using Voting over Suspiciousness Measures.

Abderrahmand Sidnas (University of Ain Temouchent)*; Fatima Zohra BELGRANA (University of Ain Témouchent); Nadjib Lazaar (Paris-Saclay University, France); Lakhdar Saïs (CRIL-CNRS, Artois University, France)

Abstract. Fault localization is a critical and time-consuming activity in software development. As software systems grow in size and complexity, the need for automated techniques that help developers efficiently identify faulty program statements becomes increasingly important. This need is particularly acute both during development and after deployment, especially for software embedded in safety-critical or high-risk systems. Spectrum-Based Fault Localization (SBFL) is one of the most widely adopted automated approaches. It exploits execution traces from passing and failing test cases to rank program statements according to their likelihood of being faulty, using suspiciousness measures. However, empirical studies show that no single measure consistently outperforms the others across different programs and fault types. In this paper, we propose an aggregation framework based on Majority Judgment voting to combine multiple fault localization measures. Each suspiciousness measure independently evaluates program statements, and the final ranking is obtained by applying Majority Judgment to aggregate these evaluations. By relying on ordinal judgments rather than raw suspiciousness scores, the proposed approach captures the consensus among measures while reducing sensitivity to outliers, resulting in a more robust and stable fault localization ranking.

S2.3 Leveraging Noise Injection to Mitigate Samples Limitations and Enhance Predictive Model Performance

Billel  Bouguedra  (University of Djelfa )*, Khaled  Sandjak  (University of Boumerdes), Mouloud  Ouanani  (University of Djelfa ), Hegazy  Rezk  (Prince Sattam bin Abdulaziz University, Al-Kharj, Riyadh)

Abstract. The samples of any database that may be assembled for AI-based prediction or classification research may be limited by complex tests that call for specialist equipment, a lack of resources, and difficult testing conditions. This lack can negatively impact ML-model training. Consequently, increasing the database's sample size using Gaussian noise injection approach may be a way to
address these issues. The database used in this study is based on geotechnical work results. There are only 238 samples total, which are made up of 11 inputs that represent the physical characteristics of the soil and the CBR output. With every batch of created data, the suggested variance shifts from 1% to 3%, 5%, 7%, and finally 9%. After the fourth extension, the statistical analysis (using Pvalue, K-S statistic, and comparing the statistical values between the original and synthetic data) gives a definitive picture of how similar the original and synthetic data are. To see how the model's predictive performance has improved, the LightGBM model is trained on both the original data and the four created datasets. In order to prevent overfitting, it is integrated with the 10-fold cross-validation to get best split between training (80%) and test (20%). By monitoring the values of R², RMSE, and MAE, performance is evaluated. After the fourth multiplication and with a sample size of 1190, the results showed that the LightGBM model performance improved greatly with the synthetic data (D4), yielding values of R² = 0.960, RMSE = 1.119, MAE = 0.758 in the test phase.

S2.4: Enhancing Session-Based Recommendations with Pre-trained GNNs for General User Preference Learning: A Novel Application.

Mouloud Amine Djenane (University of Tiaret)*, Boudjemaa Boudaa (University of Tiaret), Abdelhafid Abouaissa (University of Haute-Alsac)

Abstract. Session-based recommender systems (SBRSs) rely on graph neural networks (GNNs) to capture both local and global user preferences within short browsing sessions. Despite steady progress in GNN architectures, performance gains have begun to plateau, suggesting that architectural changes alone are insufficient. This paper presents a novel
application of pre-training and fine-tuning to GNN-based SBRSs, aimed at improving general user preference learning without architectural complexity. The core idea is to train a GNN on a large historical dataset to learn general item relationships and broad user behaviour patterns, then fine-tune it on a more recent subset to capture current trends. Two wellestablished GNN models are evaluated: SR-GNN and TAGNN. Experiments on the YOOCHOOSE dataset show that pre-training consistently improves both HR@20 and MRR@20 across both models. Pre-trained SR-GNN (P-SR-GNN) matches or outperforms more recent and architecturally complex models, including IGT and GTPAN. These results demonstrate that pre-training is a practical and effective way to enhance general user preference learning in GNN-based SBRSs.

S3.1: Evaluating the impact of swarm intelligence approaches for feature selection in software fault prediction.

Ahmed Taha Haouari (University of Science and Technology Houari Boumediene)*, Kahina Kessi (University of Science and Technology Houari Boumediene) 

Abstract. As software systems become more complex, predicting faults before deployment has emerged as a critical issue in software engineering. This research investigates the integration of machine learning models with swarm intelligence-based feature selection techniques to enhance software fault prediction. We systematically assess the performance of Particle Swarm Optimization (PSO), Ant Colony Optimization (ACO), and Grey Wolf Optimizer (GWO) across 15 open-source datasets from the PROMISE repository, which encompass object-oriented metrics. The results indicate that GWO consistently delivers superior predictive performance. Furthermore, a hybrid strategy combining GWO and a Genetic Algorithm achieves additional improvements. This work contributes an effective and interpretable prediction framework to assist developers in proactively identifying fault-prone components.

S3.2: A Method for Optimizing over the Integer Efficient Set of a Bi-Objective Integer Linear Programming Problem.

Ali zaidi (Center of Research in Amazigh Culture and Language(CRLCA))*

Abstract. In this paper, we propose an exact method to optimize a linear function Φ over the efficient set of a bi-objective integer linear programming problem. We propose a three-phase algorithm. In the first phase, we determine a pilot objective among the different objectives, and the initial solution is the optimal solution optimizing this pilot objective.
In the second phase, we generate a sequence of solutions optimizing the pilot objective under the constraint of not deteriorating the second objective and of improving the functionF ; this second phase ends when the problem becomes infeasible. The third phase determines the final optimal solution among the solutions ofS (where S is the set of solutions found in the preceding phases). A detailed didactic example is given to illustrate the different steps of our algorithm. It has been coded in Matlab using the CPLEX solver, and computational experiments have been undertaken to analyze the performance properties of the algorithm over different randomly generated problem instances.

S3.3 A Processing-Based GA Optimization platform of Fuzzy Controllers.

Mohammed Chaker Boutalbi (University of Kasdi Merbah Ouargla)*,  Amina  CHELGHOUM (National School of Advanced Technologies), Slimane Kerrouchi (University of Kasdi Merbah Ouargla), Bilal BENARABI  (University of Kasdi Merbah Ouargla), MOHAMMED SAID  ACHBI (University of Kasdi Merbah Ouargla), Boubakeur Rouabah (University of Kasdi Merbah Ouargla)

Abstract. The teaching of optimization techniques in control engineering often relies on static figures and numerical results, which conceal the dynamic behavior of evolutionary algorithms and limit students’ intuitive understanding of their internal mechanisms. This paper introduces an educational software platform developed in Processing-Java that animates an important number of optimization scenarios of fuzzy PD controller parameters. For optimization regards, students can observe convergence behavior, parameter interactions, and structural adaptation patterns, bridging the gap between theory and implementation. From control theory perspective, the student will be able to see the most impactful parameters on the controller behavior and which ones are meant to be flexible or rigid when it comes to the controller design. The framework is interactive and extensible, allowing learners and researchers to adapt it for custom case studies and nonlinear control applications. The results highlight the pedagogical value of animated optimization in improving conceptual clarity, engagement, and practical understanding of intelligent control design.

S4.1: Geodesic Quantum Learning: A Unified Framework for Unitary and Gradient Based Evolution in Variational Quantum Models.

Lokman Dallouli  (University of Batna 2)*, Hamouma  Moumen  (University of Batna 2 ), Rachid  Hedjam  (Bishop's University) 

Abstract. Variational quantum algorithms are central to near-term quantum computing, but their performance is often limited by difficult optimization landscapes, sampling noise, and sensitivity to hardware imperfections. In this work, we introduce theGeodesic Quantum Eigensolver(GQE), an optimization framework that interprets learning as a continuous, norm-preserving flow on the quantum state manifold. The method projects the energy gradient onto the tangent space induced by the Fubini-Study geometry, leading to an ideal dissipative dynamic that preserves physical normalization and guarantees monotonic energy descent in the continuous-time setting. For practical variational circuits, we implement this geometric flow through a stochastic parameter-space approximation based on simultaneous perturbation estimates. This avoids explicit construction and inversion of the quantum geometric tensor. For a fixed number of perturbations and shots, the resulting update uses a constant number of circuit evaluations per iteration with respect to the number of variational parameters, in contrast to the metric-estimation and inversion overhead required by Quantum Natural Gradient methods. Numerical experiments on transverse-field Ising and XXZ Hamiltonians show that GQE achieves competitive or improved final energies compared with SPSA, Adam, SGD, and QNG under the tested settings. Additional noisy simulations and a limited IBM hardware proof-of-concept suggest that the geometric update can act as a stabilizing regularizer under realistic perturbations. We also clarify the distinction between the ideal projected flow and its stochastic implementation, and identify broader noise benchmarking, hardware validation, and systematic barren-plateau analysis as future work.

S4.2: Leveraging Quantum Neural Networks for Forest Fire Detection: A Hybrid Approach to Early Detection and Inverse Problem Solving.

Nasreddine Guelfout (ESI)*, Leila Hamdad (ESI - Ecole nationale Supérieure d'Informatique), Hamid Haddadou (ESI - Ecole nationale Supérieure d'Informatique)

Abstract. Early detection of forest fires is crucial for minimizing environmental damage and ensuring public safety. While deep learning and CNN (Convolutional Neural Networks) models have shown promise in automating feature extraction and improving classification accuracy for small to medium-sized datasets, they often struggle with large-scale, high-dimensional data, such as those used in monitoring Algerian forests. Additionally, their performance can be hindered by complex environmental conditions like cloud cover and fog. QNNs (Quantum Neural Networks), which leverage quantum computing principles, offer potential advantages in processing complex data patterns and minimizing computing time. This paper proposes a hybrid approach combining classical layers and QNNs layers for early forest fire detection. The proposed model utilizes a quantum convolutional layer to extract quantum features from image patches, which are then processed by a classical classification layer. To address the challenges of noise and incomplete data in real-world scenarios, we formulate early forest fire detection as an inverse problem. Classical and quantum neural networks are interpreted as data driven regularized solvers of this inverse problem, with QNNs providing an alternative feature representation in a high-dimensional Hilbert space. Our novel framework effectively learns an implicit inverse operator that maps observed images to fire presence indicators, enabling robust detection even in the presence of noise and environmental interferences. The model is evaluated on a well-known dataset of forest fire, demonstrating an acceptable accuracy and robustness compared to traditional CNNs by identifying early fire indicators from noisy observations. The results highlight the potential of QNNs in enhancing image classification tasks for early forest fire detection.

S4.3 From Data to Diet: A K-Means and Linear Programming Approach to Cost-Effective Feeding in Small-to-Medium Dairy Herds.

Ferdaous Benrouba (Badji Mokhtar University )* 

Abstract. Despite feed expenses representing 50–60% of total operating expenses in dairy production, small-to-medium herds (50–200 cows) rarely have enough labor capacity to allow for individual-cow feeding. Total mixed rations from single groups consistently do not reflect the body weight and nutritional diversity in these populations. We built an open-source and modular framework that combines K-means clustering for animal segmentation with PuLP-based linear programming for least-cost ration formulation, applied to 226,840 filtered lactating cows (milk yield > 2 L/day) from a public database (Kaggle, 2025). Nutrient requirements were calculated per animal using simplified NRC (Committee on Nutrient Requirements of Dairy Cattle et al., 2021) equations for metabolic body weight, net energy for lactation and crude protein, with dry matter intake range limited between 2.5% and 3.5% of BW. Homogeneous feeding groups were then created using K-means clustering (n = 5), after which group-specific rations were optimized based on realistic feed prices for 2025–2026. With a conservative yield penalty (3%) assigned to those cows fed in groups, the five-group strategy reduced feed cost by USD 0.091/cow/day and increased net margin by 7.3%, due mainly to labor savings at USD 0.150/cow/day. The estimated annual savings varied between USD 4404 (50 cows) and USD 17,616 (200 cows). On a typical laptop, the entire pipeline runs in under five minutes.

S4.4: Combining Formal Verification and Reinforcement Learning for Program Synthesis.

Nour El Houda Tine (UBMA)*; Toufik Benouhiba (UBMA); Amina Boudjedir (UBMA)

Abstract. Program synthesis has advanced significantly with AI-based approaches. However, many learning-based approaches lack formal correctness guarantees—synthesized programs may contain logical errors or violate specifications. This limitation hinders adoption in safety-critical domains where correctness is essential. Furthermore, existing approaches typically target a single artifact, synthesizing either a formal model or an executable implementation, but not both. We present a framework combining reinforcement learning (RL) with formal property verification to synthesize both probabilistic models and their executable implementations simultaneously. Our approach uses formal properties to guide the RL agent, ensuring that both the model and its implementation satisfy all desired requirements. We evaluate our framework on an implementation of the Knuth-Yao algorithm, which models a fair dice using only fair coins, demonstrating that it synthesizes a model and its corresponding executable implementation that achieves the required properties. These results suggest that integrating formal methods with learning-based synthesis can bridge the gap between flexibility and correctness in automated programming.

P1: Minimum-time optimal control problem under simple exponential smoothing model constraints and extensions.

L'hadi Boughani (Université de BOUIRA)*, Raouia Saoudi (Université de Bouira); Mohammed Ahmed Boudref (Université de Bouira) 

P2: On Edge Lifting b-Critical Trees.

Samiha AGUEB (University Saad Dahleb, Blida)*, Mohamed Zamime (University Yahya Fares, Médéa, Algeria)

P3: AI-Driven High-Risk Event Prediction and Decision Optimization in Internet of Vehicles Networks.

Fayrouz Difallah (Higher National School of Renewable Energies, Environment & Sustainable Development)*, Guezouli Lyamine1 (Higher National School of Renewable Energies, Environment & Sustainable Development).

P4: Academic Risk Prediction with Machine Learning and a Decision Dashboard.

Larbi Omar (bechar university)*

P5: A Social Balance Lens on Cold-Start QoS Prediction: In-Context Learning Versus Gradient Boosting.

Ayoub Mokeddem (University of Batna 2)*, Saber Benharzallah (University of Batna 2), Chafik Arar (University of Batna 2); Tahar Dilekh (University of Batna 2), Moumen Hamouma (University of Batna 2)

P6: K-Nearest Neighbors-based Unified Model for Human Activity and Context Recognition.

Lydia Souici (University of Bejaia), Achour Achroufene (University of Bejaia)*, Hocine Attoumi (University of Bejaia), Redouane Saifi (University of Bejaia)

S5.1: A Scoping Review of Hybrid Spam Filtering and Federated Learning.

Lamia Hamza (University of bejaia)*, Amine Abbas (University of Bejaia), Ferial Ameziane (University of bejaia) 

Abstract. Recently, the utilization of social networks has profoundly transformed spam strategies, resulting in the emergence of image spam in addition to text spam that is known as hybrid spam. While Deep Learning-based multimodal learning architectures have showcased promising performance gains over traditional unimodal counterparts, the overwhelming majority of these solutions are designed around a centralized data collection paradigm. This paper performs a scoping review of publications from 2019 to 2025 to understand and contrast the current state-of-the-art of multimodal spam filtering under both centralized and Federated Learning paradigms. Through this study, we find that although modern architectures that rely on centralized training using CNNs, LSTMs, and vision transformers coupled with state-of-the-art fusion techniques report excellent detection accuracies, little work has been done to extend Federated Learning to multimodal spam filtering. Further, several outstanding challenges remain for distributed learning with non-Independent and Identically Distributed (non-IID) client data, communication constraints, and a general lack of natively multimodal datasets for distributed learning. Finally, this study identifies research priorities and establishes the framework for future federated detection algorithms tailored to the detection of hybrid spam.

S5.2: Unifying Clustering and Client Selection for Heterogeneous Federated Learning.

Imad MAHAYA (University of Batna2)*, Ali BEHLOUL (University of Batna2), Khaled HAMOUID  (University of Batna2)

Abstract. Federated learning enables distributed clients to collaboratively train models while keeping their data private. However, non-IID data distributions cause client drift, degrade model performance, and slow convergence. Existing client selection methods evaluate clients as independent entities, ignoring the natural groupings formed by similar data distributions, while clustered methods group clients but lack strategic selection strategies. We propose Clustered Selective Federated Learning (CSFL), a hierarchical method that clusters clients by data similarity, ensures balanced participation, and trains specialized per-cluster models with dynamic adaptation. Extensive experiments on CIFAR-10, CIFAR100, and MNIST demonstrate that CSFL achieves accuracy gains of up to 23.44% over random selection under extreme heterogeneity (a = 0.01) and reduces the number of communication rounds to reach the target accuracy by up to 34.7%, while improving client fairness and maintaining minimal communication overhead.

S5.3 Optimizing Detection Placement in HFL: A Hybrid Multi-Level Similarity Framework.

Meriem Taleb (University of Sidi-Bel-Abbes)*, Sofiane Boukli Hacene (University of Sidi-Bel-Abbes), Samir Ouchani (CESI lineact)

Abstract. Federated Learning (FL) enables collaborative training without sharing raw data, but client-provided updates expose it to poisoning and Sybil attacks. Hierarchical FL (HFL) over Cloud–Fog–Edge improves scalability and communication efficiency, yet expands the attack surface across aggregation levels. A critical underexplored challenge is detection placement: deciding where in the hierarchy to deploy detection to balance robustness, computational overhead, and communication cost. We propose a unified hybrid multi-level detection framework for HFL that fuses complementary similarity and behavioral signals–distance-based and angular similarity, robust statistical screening, graph/spectral consistency, and temporal reputation–into a single adaptive scoring rule deployable at edge, fog, and cloud levels to detect anomalous and coordinated updates under non-IID heterogeneity. We further introduce a utility-driven evaluation jointly accounting for global accuracy, detection quality, and overhead. Extensive experiments under diverse poisoning and Sybil-style attacks show that detection effectiveness depends strongly on both placement and signal fusion: the proposed hybrid approach consistently outperforms single-metric and single-level baselines, particularly at high attack ratios, while maintaining favorable efficiency.

S6.1: A Neuro-Symbolic Approach for Automatic Certainty Annotation.

Samia  Lazib (Ummto)*; Nada Mimouni (CNAM), Yasmine Medjek (Ummto), Thanina Tamazirt (Ummto), Hassina Ait Issad (Ummto); Farida Bouarab (Ummto) 

Abstract. Automatic corpus annotation serves as a foundational element in natural language processing (NLP), facilitating the enhancement of raw textual data with linguistic metadata for applications such as sentiment analysis, misinformation detection and providing high-quality, structured datasets for the training, evaluation, and validation of artificial intelligence systems. The primary challenge involves creating scalable systems that precisely identify subtle features, like certainty levels in authorial expressions, amidst escalating data volumes. This paper introduces a neuro-symbolic framework designed to annotate movie reviews from the IMDb corpus, categorizing sentences into high, medium, and low certainty levels. Our methodology combines a neural element that calculates semantic similarities with predefined certainty seeds, alongside a symbolic component that applies morphosyntactic rules to examine modals and verb tenses. These separate classifications are merged via score averaging to yield a robust final annotation, capitalizing on the interpretability of rules and the flexibility of neural learning. Assessed on a manually annotated dataset comprising 500 sentences, the hybrid model attains an overall accuracy of 62%, surpassing standalone neural (48%) and symbolic (55%) variants, with notably strong F1-scores for low certainty (0.72). These outcomes illustrate the effectiveness of neuro-symbolic integration in managing ambiguity within opinion-laden texts. Limitations encompass dependence on older embedding techniques and a restricted corpus scale, pointing toward improvements using advanced contextual models. Ultimately, this framework propels hybrid AI forward in NLP, offering potential for practical certainty-sensitive systems.

S6.2: A Hybrid NLP and Propagation Aware Framework for misinformation Detection and Mitigation in Online Social Networks.

Merini Hichem (Ecole militaire polythechnique)*, Adil Imad Eddin HOSNI ( Ecole militaire polythechnique), Kadda Beghdadbey (Ecole militaire polythechnique), Chakib AMRANI (Ecole militaire polythechnique), Islam BAIRA (Ecole militaire polythechnique), M'hamed Mataoui (Ecole Milaire Polytechnique), Abdellah Hamouda Sidhoum (Ecole Milaire Polytechnique)

Abstract. This paper tackles the dual challenge of detecting misinformation and mitigating its spread within online social networks (OSNs), where misinformation can rapidly distort public opinion and cause real-world harm. We propose a hybrid framework that unifies high accuracy NLP based detection with propagation aware mitigation grounded in human behavioral dynamics. The detection module benchmarks TF-IDF classical machine learning classifiers, a hybrid CNN–LSTM architecture, a fine-tuned BERT transformer, and lightweight ensemble fusion strategies to deliver both reliable predictions and confidence scores. To move beyond detection only solutions, we integrate these confidence scores into a Human Individual Social Behavior (HISB) rumor diffusion model, where a sigmoid based probabilistic adjustment reduces transmission likelihood for low credibility content, simulating user hesitation rather than enforcing rigid blocking. Experiments on the WELFake dataset show that BERT achieves the strongest standalone performance, while soft voting ensembles provide a competitive low-cost alternative. Largescale diffusion simulations further demonstrate that injecting detection confidence into HISB dynamics yields a substantial reduction in final infection size compared to baseline methods.

S6.3 Can Vision-Language Models Enforce Social Norms? A Replication of the Child Protest Paradigm.

Mohammed Issam Benmessaoud (University of Oum El Bouaghi)*, Mohamed Sedik Chebout (University of Oum El Bouaghi), Fabrice Mourlin (University of Paris-Est Créteil, Creteil)

Abstract. Social norms are essential for sustaining cooperation in human societies, and recent research increasingly explores how artificial agents might acquire similar normative abilities. With the growing prominence of generative agents, this work examines whether Vision-Language Models (VLMs) can infer and enforce simple social norms by reproducing a developmental psychology paradigm originally used with children. Three VLMs (gemini-2.5-flash, gemini-robotics-er-1.5-preview, and gemini-2.5-flash-lite) were shown videos in which an action is first demonstrated by a coordinator and then violated by a puppet. To prompt reflective responses, we use a Belief-Desire-Intention (BDI)-inspired prompting strategy intended to elicit spontaneous reactions without direct instruction to detect violations. We analyze protest rates across pedagogical, intentional, and accidental conditions, as well as across creativity levels. Results show significant variation in norm enforcement across models. High-performing models protest significantly more in pedagogical and intentional contexts than in accidental ones (χ = 112.77, p <0.001), indicating an emerging sensitivity to social cues and exhibiting the four protest forms seen in children, including normative and imperative protest, teaching, and tattling. However, unlike children, VLMs sometimes protest far more in intentional than pedagogical conditions. A logistic regression further shows that higher creativity significantly decreases the likelihood of protest (p <0.001). Overall, the findings suggest that, although prompting can guide VLMs toward a normative stance, they still lack the internalized, spontaneous grasp of social norms characteristic of humans, underscoring challenges for future work on human inspired normative learning in generative agents.

S7.1: A Hybrid Framework Integrating Machine Learning and Matrix Factorization for Cardiac Disease Classification.

Nourelhouda Groun (University Mohamed Khider Biskra)*, Eusebio Valero (Universidad Politecnica de Madrid), Soledad Le Clainche (Universidad Politecnica de Madrid), Jesus Garicano Mena (Universidad Politecnica de Madrid) 

Abstract. Medical imaging is a crucial component for accurately diagnosing cardiac diseases, yet machine learning techniques face challenges due to limited access to large, high-quality datasets. To address this issue, we propose a hybrid approach that joins matrix factorization methods with convolutional neural networks for cardiac disease classification.
This approach adopts a matrix factorization pre-processing step based on the Singular Value Decomposition method, creating principle modes that represent key cardiac features across five medical conditions: healthy, diabetic cardiomyopathy, myocardial infarction, obesity, and TAC-induced hypertension in mice. In the following step, our study compares two cases: classification using original echocardiography images and classification using images projected onto the principle modes. The results demonstrate a significant improvement in accuracy, showing an increase of approximately 50% when employing matrix factorization-based preprocessing.

S7.2: Deep Semantic Centroid Hashing with PCA for Multi-Label Image Retrieval.

Younes Bouali (University of Batna 2)*, Yassmina Saadna (University of Batna 2), Souheila Bouam (University of Batna 2)

Abstract. Deep supervised hashing has emerged as a fundamental technique for large-scale image retrieval, yet many existing methods suffer from high computational costs and optimization difficulties due to complex pairwise similarity learning. In this paper, we propose Deep Semantic Centroid Hashing (DSCH), a novel framework designed for multi-label image retrieval that reshapes the hashing process by pre-structuring the Hamming space using semantic anchors. Our approach bypasses traditional similarity-based constraints by computing class-wise centroids in the feature space and applying a PCA projection to generate discriminative binary targets that capture maximum inter-class variance. To handle multi-label environments, we introduce a composite target generation mechanism that maps images to the semantic barycenter of their constituent categories. A key strength of DSCH is that it operates on top of a frozen EfficientNet-B0 backbone without any fine-tuning, making it highly resource-efficient and straightforward to deploy. Empirical evaluations on MS-COCO and NUS-WIDE demonstrate that DSCH achieves a superior balance between retrieval precision and operational efficiency, maintaining high semantic consistency even at very short code lengths. The proposed method offers a scalable and energy-efficient solution for real-time, large-scale multi-label image retrieval.

S7.3 An LSTM-Based Decision Support Method Using Visual Risk Modeling in Surgical Video.

Samia NOUREDDINE (University of Health Sciences, Algiers)*, Nazim KAZAR (The Research Center in Computer Sciences from Paris 1 Panthéon-Sorbonne University), Abir  BETKA (Department of Electrical Engineering, University of El-Oued), Abida  TOUMI (Department of Electrical Engineering, University of Biskra ), Okba  KAZAR (Department of Computer Science, University of Kalba, Sharjah)

Abstract. Decision making  is  an  extremely  critical  process  that  takes  place  under  severe  time constraints  and dynamically changing visual  conditions.  Intermittent periods of  reduced visibility, such as fast-moving cameras,  obstructions,  and loss of contrast,  can increase cognitive burden and, ultimately,  undermine safety.  According to the World Health Organization (WHO), post-operative complications for hospitalized surgical patients can be as high as 25%. This paper aims to investigate a light and interpretable framework for measuring the visual instability of surgical scenes, with the hope of enabling future generative models for simulating or predicting future intraoperative conditions. The proposed method uses interpretable feature descriptors (i.e., mean brightness, global contrast, and inter-frame motion) and models their temporal behaviors with a well-established architecture called the Long Short-Term Memory (LSTM) network.  The proposed method also provides a continuous value of the Visual Risk Index (VRI) between 0 and 1, as well as  a  spatial  map of  instability. Observations obtained from experiments demonstrate a stable low-risk regime followed by a transient peak, reflecting short periods of vision-related disturbances picked up by the introduced indicators. The conducted investigation can still be treated as a proof of concept  despite lacking ground Truth information because it confirms that video feeds from surgical scenes can potentially be converted to quantified signals for use in upcoming predictive models.

S8.1: Adaptive Fuzzy Logic Control of Tramway Traction Motors under Variable Solar Power for Energy-Efficient Operation.

Fairouz Kendouli (Faculty of Science and Technology (FST), University Mentouri Constantine 1)*, Hemama Aboud (Faculty of Science and Technology (FST), University Mentouri Constantine 1), Khoudir Abed (Faculty of Science and Technology (FST), University Mentouri Constantine 1) 

Abstract. Tramway traction motors  are  vital  for  urban  transport,  requiring precise  torque  and  speed  control  to  ensure  passenger  comfort,  safety,  and energy  efficiency.  Conventional  control  strategies  often  cause  sharp  torque spikes and sudden speed changes, increasing mechanical stress and reducing reliability. This study introduces an adaptive Fuzzy Logic Controller (FLC) that dynamically adjusts system response based on real-time speed error and its rate of change.  The controller  ensures smooth,  realistic  responses under varying conditions, including acceleration, deceleration, and fluctuations in solar power supply.  MATLAB simulations  demonstrate  that  the  FLC maintains  DC-link stability  with a settling  time  of 1.5 seconds, and significantly  reduces peak deceleration (acceleration) by approximately 10% compared to PID (from -20 m/s 2  to -18 m/s 2 ). While PID showed a lower RMS error (3.92 vs 4.33), the FLC eliminated overshoot and provided much smoother torque transitions, reducing mechanical  stress  and  ensuring  realistic  motion.  Results confirm that the adaptive FLC enhances ride comfort, system stability, and energy efficiency, offering a practical solution for sustainable urban tramway operation.

S8.2: Improving energy and latency in WSNs using sleep scheduling algorithm.

BENGHENI ABDELMALEK (Ibn Khaldoun University, Tiaret)*

Abstract. In this article, we propose a Sleep Scheduling Algorithm to improve Wireless Sensor Networks lifetime (SSA-MAC) that aims to reduce latency and energy consumption by lowering the duty cycle of sensor nodes. SSA-MAC is based on a receiver-initiated scheme and uses the Low Power Probing (LPP) technique with beacon messages, allowing each node to adapt its sleep duration according to its remaining energy and predefined energy thresholds. The simulation results show that SSA-MAC achieves lower latency and better energy efficiency than EH2M and MTE-EHWSN.

S8.3 CNN-BiLSTM-Attention with ST-Maps for Real-World Arabic Sign Language Recognition.

Rym Ines Bouchama (University Abdelhamid Mehri Constantine 2); Feriel Belaribi (Université Abdelhamid Mehri Constantine 2); Houda Benaliouche (University Abdelhamid Mehri Constantine 2)*

Abstract. Bridging the gap between gestures and spoken language remains a fundamental challenge for AI in a world where 466 million people with hearing impairments rely on sign language as their primary communication tool. This work presents a deep learning pipeline that uses smartphone-captured videos in real-world scenarios to recognize 20 words of Arabic Sign Language (ArSL) without specialized sensors. The pipeline begins with MediaPipe Hands to extract 3D landmarks, followed by spatio-temporal normalization and data augmentation. These steps produce 4-channel Spatio-Temporal Maps (ST-Maps) encoding hand positions and velocities to capture gesture dynamics. Classification relies on a CNN-BiLSTM-Attention architecture: CNN blocks augmented with Convolutional Block Attention Module (CBAM) extract spatial features, dual BiLSTM layers model bidirectional temporal dependencies, and a temporal self-attention mechanism weights discriminative timesteps prior to classification over 20 classes. On the public ASL 20-Words dataset (8,467 videos, 72 signers), the model achieves 95.63% accuracy in an 80/20 validation split, surpassing the state-of-the-art benchmark of 92% by 3.63%. Ablation studies confirm contributions of velocity channels (+3.6%), CBAM (+0.18%), and temporal attention (+31.23%).Our experiments confirm strong generalization under signer-independent evaluation: 94.76% on ASL 20-Words and 94.38% on the 50-class KASRL dataset, outperforming Swin-T (Tiny) by 28.71 and 46.88 percentage points respectively. On the KASRL dependent protocol, our model also surpasses Swin-T by 11.88 percentage points (98.55% vs. 86.67%). 

S9.1: ZK-Organic: Verifiable Organic Compliance Without Data Disclosure.

Lina Moumni (University of Batna2)*, Hamouma Moumen (University of Batna2), Guezouli Lahcene (University of Batna2) 

Abstract. Organic certification requires farms to disclose extensive operational records to certification authorities, including pesticide usage, fertiliser quantities, and supplier information. While integrating such data into blockchain systems enhances traceability and verification through immutable ledgers, it also introduces a significant data security paradox, as the ledger’s transparency may expose sensitive agricultural and proprietary information. This paper proposes a privacy-preserving framework for verifying compliance with organic standards using zero-knowledge proofs (ZKPs). Instead of revealing raw farm records, producers generate cryptographic proofs demonstrating adherence to certification policies. To efficiently represent dynamic certification rules, we introduce Merkletree policy commitments, which enable certification authorities to publish verifiable rule sets and allow farms to prove membership in allowed policy lists. The proposed architecture integrates blockchain-based verification with privacy-preserving compliance proofs. We describe the formal system design and cryptographic model utilising zk-SNARK circuits and smart contracts to enable verifiable certification audits while protecting sensitive agricultural data. The framework is evaluated through computational complexity analysis and formal security definitions, demonstrating its capacity for constant-time verification and scalability within decentralised execution environments.

S9.2: Multi-Source Power Systems (MSPS) Modeling, Simulation, and Optimization.

OUAREM Mohamed (University Mohamed El Bachir El Ibrahimi)*, Toufik Madani Layadi (University Mohamed El Bachir El Ibrahimi)

Abstract. Renewable energy sources have known large exploitation worldwide. Integration of these sources allows the obtaining of sustainable clean energy. Modeling, simulation, and optimization of multi-source power systems combining different renewable energy sources represents the goal of this research paper. In this work, a new Python-programmed simulator/optimizer for multisource power systems is presented and tested with a case study. This software allows to calculate the optimal sizing of different components of the multisource power system through a real meteorological database.  The results obtained with this simulator demonstrate important performance. Furthermore, the simulator is characterized by its various useful features.

S9.3 Knowledge Distillation for Lightweight IoT Security: Implications for IoMT Environments.

Khouloud Benammar (University of batna2)*, hamouma moumen (university of batna 2 ), chafik arar (university of batna 2), mohand tahar kechadi (University College Dublin)

Abstract. The Internet of Medical Things (IoMT) enables advanced healthcare services through interconnected wearable sensors, implantable devices, and cloud-based analytics. However, the expanded attack surface exposes sensitive medical data and patient safety to significant cyber threats. AI-based intrusion detection systems (AI-IDS) offer promising
solutions, yet their deployment on resource-constrained IoMT devices remains challenging. In practice, AI-IDS may also be deployed on edge gateways or backend servers, here latency, privacy, and regulatory requirements still apply. Knowledge Distillation (KD) enables lightweight, resource-efficient IDS while retaining predictive performance. Existing KD-based approaches, primarily designed for generic IoT, overlook IoMT specific challenges such as latency-sensitive clinical applications, heterogeneous patient data, regulatory compliance (HIPAA, GDPR), and safety-critical requirements. This paper surveys 16 representative KD based IDS studies, identifies critical deployment gaps, and proposes a hierarchical security framework that combines lightweight on-device detection with edge- or cloud-based analytics to enhance practical IoMT security architectures.

S10.1: A Lightweight Hybrid CNN-Transformer Model for Single Lead Atrial Fibrillation Detection.

Abdelfettah HAMID-SIDI-YKRELEF (Yahia Fares University of Medea)*, Ahmed KHELDOUN (Yahia Fares University of Medea)
 

Abstract. Atrial fibrillation (AFib) is the most common cardiac arrhythmia worldwide and a leading cause of stroke and heart failure. Early and accurate detection from single-lead electrocardiogram (ECG) signals remains a critical clinical challenge, particularly due to the paroxysmal nature of AFib and the scarcity of large, well-labeled ECG datasets. To address the problem of limited labeled data, we construct a combined dataset by aggregating AFib and normal sinus rhythm (NSR) recordings from six public benchmarks, thereby substantially increasing data diversity and volume for model training. We then propose a hybrid deeplearning model for binary classification of raw single-lead ECG signals. The architecture combines a convolutional neural network (CNN) for local morphological feature extraction, followed by a Transformer encoder network to capture long-range temporal dependencies, and a fully connected classification head for final prediction. Experimental results on the combined dataset using 5-fold cross-validation demonstrate that the proposed model achieves a mean accuracy of 99.09±0.06%, with a precision, recall (sensitivity), specificity, F1-score, and AUC of 99.11±0.15%, 99.07±0.12%, 99.11±0.15%, 0.9909±0.0006, and 0.9988±0.0002, respectively. The total inference time of3 .08× 10 -3 seconds per 10-second recording, corresponding to a throughput of approximately 325 samples per second. These results confirm that the proposed CNN-Transformer model, trained on a strategically aggregated multi-source dataset, offers a robust and generalizable solution for point-of-care AFib detection. To further assess clinical feasibility, the proposed model is deployed on a lowcost AD8232 ECG module interfaced with an Arduino microcontroller, where a 10-second raw ECG recording is acquired and passed directly to the model for real-time binary diagnosis.

S10.2: CBAM-MSRNet: Multi-Scale Residual Attention Network for Automated ECG Arrhythmia Detection.

Saoueb Kerdoudi (university of batna 2)*, larbi GUEZOULI (HNS-RE2SD), Abdessalam HATTAB (university of batna 2); Tahar Dilekh (university of batna 2)

Abstract. Accurate identification of arrhythmias from electrocardiogram (ECG) signals is crucial for early diagnosis and timely treatment of cardiovascular diseases. In this work, we propose CBAM-MSRNet, a hybrid deep learning framework designed to improve feature extraction and classification reliability from ECG recordings. The model integrates a Multi-Scale Residual Network (MSRNet) to learn diverse morphological characteristics across different temporal scales, enabling robust representation of variations in heartbeat shapes. To further capture sequential dependencies and rhythm-related patterns, the extracted features are fed into a Bidirectional Long Short-Term Memory (BiLSTM) module, which models long- and short-range temporal dynamics in both forward and backward directions. In addition, a Convolutional Block Attention Module (CBAM) is employed to adaptively emphasize informative channel and spatial features, suppressing irrelevant noise and enhancing discriminative regions in the learned representations. Extensive experiments conducted on the MIT-BIH Arrhythmia database demonstrate strong performance, achieving 99% accuracy across five arrhythmia categories and surpassing several recent state-of-the-art approaches. Overall, the proposed CBAM-MSRNet shows high robustness and generalization capability, offering a promising solution for dependable automated arrhythmia classification and improved clinical decision support.

S10.3 Efficient Multimodal Deep Learning for Melanoma Detection: Integrating Dermoscopic Images and Clinical Metadata

yassmina Saadna (University Batna 2)*, Saliha Mezzoudj  (University of Algiers 1), Meriem Khelifa (University of ouargla)

Abstract. Melanoma is the most lethal form of skin cancer, with early detection crucial for patient survival. While deep learning has shown promise in automated melanoma detection using dermoscopic images, most approaches neglect complementary clinical metadata that dermatologists routinely consider. This study presents an efficient multimodal deep learning frame-work that integrates EfficientNet-B0 image features with patient metadata (age, sex, anatomical site) through concatenation-based fusion. Trained using patient-stratified 3-fold cross-validation on the ISIC 2020 dataset (33,126 images), our approach achieved an out-of-fold AUC of 0.9297 ± 0.0038, representing a statistically significant +0.032 improvement over an image-only baseline (DeLong test, p < 0.001). At the F1-optimal threshold, the model attained precision 0.600, recall 0.831, and specificity 0.990. Comprehensive subgroup analyses demonstrated robust perfor- mance across age groups (AUC 0.915 0.943) and anatomical sites (AUC 0.900–0.937). With only 4.2M parameters and < hours training time on a single GPU, this work provides a computationally efficient, reproducible baseline for multimodal melanoma detection.

S11.1: OPT(n): Memory-Efficient Partitioned Trees.

Fahd Mustapha Meguellati (Ecole nationale Supérieure d’Informatique ESI)*, Djamel Eddine Zegour (Ecole nationale Supérieure d'Informatique ESI) 

Abstract. Balanced binary search trees are fundamental components of many information systems, where efficient updates and predictable per-formance are critical. However, most deterministic balanced trees rely on explicit structural metadata, increasing memory consumption and limiting their applicability in constrained environments. This paper presents OPT(n), a memory-aware optimization of the Partitioned Tree family, denoted PT(n), that eliminates explicit height storage while preserving deterministic balancing guarantees. OPT(n) encodes structural information using a single bit per node and relies on bounded local recomputation of heights within partitioned subtrees. An experimental evaluation comparing OPT(n) with PT(n), AVL trees, and Red-Black trees shows that OPT(n) reduces balancing metadata from 8 bits to 1 bit per node, yielding memory savings of up to 875 KB for trees containing one million nodes. For moderate parameter values (n= 4), OPT(n) maintains update performance within a limited overhead range relative to PT(n), while larger values ofn lead to a marked slowdown due to deeper local height recomputation. These results position OPT(n) as an effective space–time compromise for memory-constrained systems requiring deterministic behavior.

S11.2: Hysteresis Threshold M/M/c Queue with Setup and Power-Saving Modes in Cloud Computing.

Sabrina MOUSLI (University of Bejaia)*, Karima ADEL-AISSANOU (University of Bejaia)

Abstract. Cloud computing has become a key paradigm for providing ubiquitous, on-demand access to shared computing resources without requiring heavy local hardware, but its rapid growth has driven a massive expansion of large data centers whose electricity use already forms a significant share of global consumption. In typical deployments, server utilization remains relatively low, while idle machines still draw a substantial fraction of their peak power, making the management of energy consumption—and in particular the reduction of idle energy—crucial for sustainable cloud operation. Existing approaches combine software-based consolidation, dynamic task scheduling, and queueing-theoretic mechanisms based on threshold and vacation policies to switch servers between active, sleep, or off states, thereby improving the energy–performance trade-off. Within this context, the M/M/c/Set–Saving model captures a multi-server queue with setup times and power-saving modes, in which idle servers process jobs at a reduced rate before being set up to the normal mode when needed. In this paper, we extend this model and augment it with activation thresholds that depend on both the number of active servers and the total number of jobs in the system, with the aim of limiting unnecessary setup operations and further enhancing the balance between energy consumption and quality of service in data-center environments.

S11.3 Semantic compression for Knowledge Graph Attention Networks via Formal Concept Analysis for efficient recommendation.

Naouel Manaa (Badji Mokhtar Annaba University)*, hassina seridi (Badji Mokhtar Annaba-University);  Mohamed Said Mehdi MENDJEL (Badji Mokhtar Annaba-University), Nour Elislem  Karabadji (National Higher School of Technology and Engineering)

Abstract. Knowledge Graph Attention Networks have emerged as a powerful framework for recommendation by modeling high order relational dependencies via attention-based message passing. However, their scalability is hindered by the computational cost of attending over large, often redundant entity relation graphs. This paper introduces semantic compression via Formal Concept Analysis as a principled method to reduce KGAT’s graph size while preserving semantic structure. By extracting formal concepts from knowledge graph triples and merging entities with identical relational incidence patterns, the proposed method reduces graph density and eliminates structural redundancy. Experiments show substantial reductions in entity and triple counts, accompanied by significant training time savings. These results demonstrate that concept lattice-based abstraction provides an effective and theoretically grounded strategy for improving the scalability of graph neural recommender systems.

S12.1: Traffic Signal Control in a Connected Road Network: A constrained multi-objective and Reinforcement Learning Approach.

Nouara DJERROUD (LaMOS Research Unit)*, Mohammed Said RADJEF (LaMOS Research Unit), Nassim TOUCHE ( LaMOS Research Unit)
 

Abstract. Urban traffic congestion remains a major challenge for intelligent transportation systems. This paper proposes a constrained multiobjective optimization model for decentralized traffic signal control. Using SUMO, the model is evaluated on an isolated four-phase intersection with 12 movements and a network of four interconnected intersections where each junction independently optimizes its signal timings. At each cycle, green times are allocated across the four phases by solving a linear program derived from a weighted Tchebycheff scalarization, using detector-based inflow, outflow, and queue measurements. Performance is assessed through key traffic indicators, including inserted and arrived
vehicles, vehicles in the network, average travel time, fuel consumption, and CO emissions. Results show that the proposed four-phase multiobjective approach significantly outperforms a conventional two-phase non-cooperative model based on Cournot equilibrium in both configurations, reducing congestion and environmental costs. In addition, we investigate reinforcement learning with Proximal Policy Optimization (PPO) as a complementary adaptive perspective for traffic signal control under dynamic traffic conditions.

S12.2: Investigation and analysis of Existing Approaches for Road Traffic Congestion Prediction, Detection and Avoidance.

Asma BOUALI (Univesity of Batna 2)*, Soundes BELKACEM (University of Batna2), Sofiane Aouag (University of Batna2)

Abstract. Vehicular Ad-Hoc Networks (VANETs) are essential components of Intelligent Transportation Systems (ITS). VANET’s enable communication between vehicles and infrastructure, enhance road safety, increase traffic efficiency, and provide real-time information to drivers. However, one of the significant challenges faced by VANETs is traffic congestion. In fact, it can result in delays, increased travel times, and higher emissions. Hence addressing congestion in VANETs is crucial for reaching their objectives, as effective communication and coordination among vehicles can help to optimize traffic flow and reduce traffic jams. Considering the strong interconnection between VANETs, Internet of Things (IoT), and smart cities, the integration of these domains is crucial for building efficient and sustainable transportation systems. This paper investigates and analyzes existing strategies for traffic congestion detection, prediction, and avoidance in VANETs, IoT and smart cities. It highlights research gaps, identifies limitations, and suggests future directions for more efficient congestion management.

S12.3 Swin Transformers for In-the-Wild Plant Disease Classification: A Multi-Architecture Statistical Evaluation.

Ghrieb Nawel (echahid cheikh larbi tebessi university)*

Abstract. Plant disease classification in real-world conditions remains challenging due to noisy backgrounds, variable illumination, and occlusions. This study presents a statistically rigorous evaluation of vision architectures on the in-thewild PlantDoc dataset. We compare four state-of-the-art architectures —EfficientNet-B4 (CNN baseline), ConvNeXt (modern CNN), ViT (Vision Transformer), and Swin (hierarchical Transformer), using rigorous 5-fold stratified cross-validation with independent test set evaluation via ensemble aggregation and comprehensive statistical testing. Beyond architectural comparison, we investigate the impact of training strategies through an ablation study and assess generalization on external data. Our results demonstrate that the Swin Transformer ensemble achieves promising absolute performance (F1-macro = 0.7504). Ablation results reveal that our optimized training configuration (Focal Loss + Label Smoothing) is a primary driver of success, yielding a +4.22 percentage points (pp) F1 improvement over standard cross-entropy on Swin-Base. While limited test set size yields non-significant differences among modern architectures after strict Bonferroni correction, insufficient statistical power prevents claims of equivalence; empirical trends and consistent training stability favor hierarchical transformer architectures for real-world agricultural deployment.