Publications

2025

• Software Identification for Cybersecurity: Survey and Recommendations for Regulators
  
  • Barais Olivier
  • Cosmo Roberto Di
  • Mé Ludovic
  • Zacchiroli Stefano
  • Zendra Olivier
  , 2025.
• Reconfigurable intelligent surfaces (RIS) using NOMA with thermal energy harvesting
  
  • Boujemaa Hatem
  • Alhussein Musaed
  • Rekaya Ghaya
  Signal, Image and Video Processing, Springer Verlag, 2025, 19 (5), pp.419:1-419:8. The integration of Reconfigurable Intelligent Surfaces (RIS) with Non-Orthogonal Multiple Access (NOMA) and thermal energy harvesting presents a novel approach to enhancing wireless communication networks. RIS technology optimizes signal propagation and improves network efficiency through programmable surface elements, while NOMA increases spectral efficiency by allowing multiple users to share the same frequency resource. When combined with thermal energy harvesting, which captures ambient heat and converts it into electrical power, this integration offers a sustainable solution to power the RIS infrastructure. This paper explores the synergistic benefits of RIS using NOMA with thermal energy harvesting, examining its impact on network performance, energy efficiency, and sustainability. Through a review of recent advancements and research, we discuss how this combined approach can address key challenges in modern wireless communications and contribute to the development of greener, more efficient networks. (10.1007/s11760-025-03996-x)
  DOI : 10.1007/s11760-025-03996-x
• Multispectral Texture Synthesis using RGB Convolutional Neural Networks
  
  • Ollivier Sélim
  • Gousseau Yann
  • Lefebvre Sidonie
  IEEE Transactions on Geoscience and Remote Sensing, Institute of Electrical and Electronics Engineers, 2025, 63, pp.5402914. State-of-the-art RGB texture synthesis algorithms rely on style distances that are computed through statistics of deep features. These deep features are extracted by classification neural networks that have been trained on large datasets of RGB images. Extending such synthesis methods to multispectral images is not straightforward, since the pre-trained networks are designed for and have been trained on RGB images. In this work, we propose two solutions to extend these methods to multispectral imaging. Neither of them require additional training of the neural network from which the second order neural statistics are extracted. The first one consists in optimizing over batches of random triplets of spectral bands throughout training. The second one projects multispectral pixels onto a 3 dimensional space. We further explore the benefit of a color transfer operation upstream of the projection to avoid the potentially abnormal color distributions induced by the projection. Our experiments compare the performances of the various methods through different metrics. We demonstrate that they can be used to perform exemplar-based texture synthesis, achieve good visual quality and comes close to state-of-the art methods on RGB bands. Code is available at \url{https://github.com/selim2483/multispectral-textureCNN} (10.1109/TGRS.2025.3554931)
  DOI : 10.1109/TGRS.2025.3554931
• Using a Probabilistic Database in an Image Retrieval Application
  
  • Yunus Fajrian
  • Karmakar Pratik
  • Senellart Pierre
  • Abdessalem Talel
  • Bressan Stéphane
  , 2025. ProvSQL is a PostgreSQL extension implementing provenance management and probabilistic database features. ProvSQL seamlessly extends relational database functionality to support the storage, tracking through derivations and transformations, and querying of metadata that explain and qualify the data and query results. In this demonstration, ProvSQL is used to implement a content-based image retrieval system. A deep learning object detection model identifies objects of selected classes located within the images of a large-scale image data set. The uncertainty associated with object detection is recorded. ProvSQL's provenance model incorporates this uncertainty into the retrieval process, thus facilitating the generation of accurate and reliable results and allowing for decision-making in scenarios with incomplete or uncertain information. The demonstration illustrates how ProvSQL handles query processing, uncertainty tracking, and probability computation. It highlights the utility of a probabilistic database for applications dealing with uncertain data, compared to traditional threshold-based approaches.
• Transformer-Based Packet Scheduling under Strict Delay and Buffer Constraints
  
  • Nérondat Sylvain
  • Leturc Xavier
  • Le Martret Christophe
  • Ciblat Philippe
  , 2025, pp.1-6. <div><p>This paper presents a packet scheduler for managing multiple links with varying channel capacities, where each link carries multiple data flows with finite buffers and strict delay constraints. Packet loss can result from buffer overflow or delay violations. We propose a deep reinforcement learning scheduler based on an encoder-only transformer (EOT) architecture, capable of handling a variable number of links without dedicated training. Using deep Q-learning, the scheduler minimizes the packet loss rate (PLR). Simulations show that our approach outperforms a state-of-the-art fully connected (FC) scheduler, delivering better performance under diverse configurations of links, packet arrival rates, and channel capacities.</p></div> (10.1109/WCNC61545.2025.10978237)
  DOI : 10.1109/WCNC61545.2025.10978237
• Role of deep learning into three-dimensional reconstruction for breast tomosynthesis
  
  • Quillent Arnaud
  , 2025. Digital breast tomosynthesis (DBT) is an X-ray imaging technique introduced in the 2010s that provides a three-dimensional reconstruction of tissues. Commercial systems rely on acquiring images from a limited number of views sampled every 1 to 3 degrees, evenly distributed over a scanning cone of 15 to 40°. The acquired projections are then processed by a reconstruction algorithm that generates the volume reviewed by the radiologist. Numerous analytical and neural network reconstruction algorithms exist. However, the narrow opening of the acquisition cone limits the quality of the reconstructed volume, particularly its resolution in the thickness of the breast. Orthogonal slices thus become unreadable with current techniques :the gradual extinction of artefacts from an object outside its focal planes disrupts the reading of objects located in neighbouring planes. Thus, there is currently no commercial solution effectively addressing limited-angle artefacts in breast tomosynthesis.The use of deep learning faces several challenges. Indeed, paired data in real conditions that would allow supervised learning are absent,and volumes from other breast imaging modalities cannot be used directly . Questions also arise regarding the properties of the obtained methods. On the one hand, the reconstructed volumes are not always consistent to the measurements, a property well mastered by conventional methods. Moreover, the uncertainty of the image reconstructed by deep learning is almost never explicit. In this thesis, we first propose a post-processing approach for conventional reconstruction. To develop our method, we create a synthetic database consisting of digital phantoms. We simulate their X-ray projections and reconstruct 3D volumes using conventional methods,thus creating a dataset that we use to train a convolutional neural network in a supervised manner. We demonstrate that the proposed strategy significantly improves the quality of orthogonal planes and is therefore promising to address the problem of reconstruction in breast tomosynthesis. Secondly, we seek to obtain an estimate of the reliability of the volumes predicted by the neural network. We adopt a Bayesian perspective, differentiating between aleatory and epistemic uncertainties. We model the first term using a Laplace distribution and the second by approximating the posterior predictive distribution, then compare the results obtained with a Monte Carlo Dropout method and a deep ensemble. We show that the computed uncertainty is a good approximation of the actual error and reuse to minimise a data consistency term. Finally, we improve the realism of the synthetic volumes by adapting images from another imaging modality that does not present the same artefacts as tomosynthesis. After appropriate denoising, we segment these images and simulate their compression to replicate the conditions of the medical examination. We redesign the neural network to allow 3D learning that considers information from the three anatomical directions, and solve the convergence problems related to the calculation of the uncertainty terms. We then impose the fidelity constraint to the projections by limiting the contribution of the neural model to the kernel space of the projection operator. Finally, we evaluate our methods on volumes with specific geometric properties as well as on clinical images, and highlight the benefits and limitations of our deep learning reconstruction approaches for breast tomosynthesis.
• Adaptive Passive Beamforming in RIS-Aided Communications with Q-Learning
  
  • Chêne Thomas
  • Bounhar Oumaïma
  • Othman Ghaya Rekaya-Ben
  • Damen Oussama
  , 2025, pp.1-6. Reconfigurable Intelligent Surfaces (RIS) appear as a promising solution to combat wireless channel fading and interferences. However, the elements of the RIS need to be properly oriented to boost the data transmission rate. In this work, we propose a new strategy to adaptively configure the RIS without Channel State Information (CSI). Our goal is to minimize the number of RIS configurations to be tested to find the optimal one. We formulate the problem as a stochastic shortest path problem, and use Q-Learning to solve it. (10.1109/WCNC61545.2025.10978715)
  DOI : 10.1109/WCNC61545.2025.10978715
• Stochastic Geometry-Based MCS Adaption Analysis for Uplink Cellular Networks
  
  • Guo Xinyi
  • Liu Qiong
  • Wang Shanshan
  • You Li
  , 2025, pp.1-6. The link adaptation plays a crucial role in the fifth generation (5G) and future wireless networks, where adaptive modulation and coding (AMC) is vital for significantly increasing the data transmission rate and quality of service (QoS) by adjusting the modulation and coding scheme (MCS). In this work, we investigate the stochastic geometry-based MCS adaption for the uplink cellular networks with Poisson distributed base stations (BS) and user equipments (UE). We first define the conditional received rate by quantizing the channel quality, i.e., signal to interference ratio (SIR), using the sets of thresholds. Basically, higher SIR indicates better channel condition and applys higher order of modulation scheme, which leads to higher received rate. We then derive the framework of meta distribution on the conditional received rate, the spatially-average spectral efficiency (SE), and the variance of the SE. In addition, beta approximation and several bounds are presented to simplify the calculation of meta distribution. We validate the proposed framework by numerical simulations under different system parameters. (10.1109/WCNC61545.2025.10978118)
  DOI : 10.1109/WCNC61545.2025.10978118
• Towards semantically enriched embeddings for knowledge graph completion
  
  • Alam Mehwish
  • van Harmelen Frank
  • Acosta Maribel
  Neurosymbolic Artificial Intelligence, 2025, 1, pp.1-16. Embedding based Knowledge Graph (KG) completion has gained much attention over the past few years. Most of the current algorithms consider a KG as a multidirectional labeled graph and lack the ability to capture the semantics underlying the schematic information. This position paper revises the state of the art and discusses several variations of the existing algorithms for KG completion, which are discussed progressively based on the level of expressivity of the semantics utilized. The paper begins with analysing various KG completion algorithms considering only factual information such as transductive and inductive link prediction and entity type prediction algorithms. It then revises the algorithms utilizing Large Language Models as background knowledge. Afterwards, it discusses the algorithms progressively utilizing semantic information such as class hierarchy information within the KGs and semantics represented in different description logic axioms. The paper concludes with a critical reflection on the current state of work in the community, where we argue that the aspects of semantics, rigorous evaluation protocols, and bias against external sources have not been sufficiently addressed in the literature, which hampers a more thorough understanding of advantages and limitations of existing approaches. Lastly, we provide recommendations for future directions. (10.3233/NAI-240731)
  DOI : 10.3233/NAI-240731
• Noise and dynamics of hybrid plasmonic semiconductor lasers for InP-based quantum integrated optical communications
  
  • Cui Di
  , 2025. Integrating plasmonics into photonic integrated circuits (PICs) enables the development of ultra-compact devices while maintaining high performance. Plasmonic structures enhance light-matter interactions, leading to advanced functionalities such as high-sensitivity sensing and nonlinear optics. By offering subwavelength control and faster modulation speeds, they overcome the limitations of silicon photonics in light confinement and modulation efficiency.This thesis presents a comprehensive experimental study of hybrid plasmonic semiconductor lasers, emphasizing their potential applications in plasmonic PICs. Key characteristics such as linewidth enhancement, relaxation oscillations, and relative intensity noise (RIN) are examined, along with nonlinear dynamics induced by external perturbations like optical feedback and optical injection. Experimental results reveal that hybrid plasmonic lasers exhibit greater resistance to optical feedback compared to conventional quantum well (QW) lasers, demonstrating reduced dynamic instabilities and superior feedback tolerance. This suggests they could be integrated into photonic circuits without requiring optical isolators.Optical injection experiments show that hybrid plasmonic lasers deviate from the conventional chaotic behavior observed in semiconductor lasers, instead exhibiting sustained feedback-induced oscillations. Sensitivity tests under optical feedback further confirm their reluctance to transition into chaotic states, even under destabilizing conditions. These findings highlight the significant role of surface plasmon polariton (SPP) interactions in enhancing nonlinear effects. The resonance properties of the metal coating and underdamped relaxation oscillations in the surface plasmon waveguide contribute to the laser's unique nonlinear behavior. With their strong resistance to optical feedback, absence of chaotic oscillations, and distinct dynamic properties, hybrid plasmonic lasers emerge as promising candidates for large-scale CMOS-compatible photonic integration, particularly in eliminating the need for bulky optical isolators in PICs.
• Self-Supervised Learning of Audio Representations for Musical Applications
  
  • Riou Alain
  , 2025. The goal of this PhD is to propose new paradigms for training deep neural networks to produce audio representations that are suited for diverse musical applications, with a special focus on self-supervised learning (SSL) approaches.In our first work, we focus on the task of monophonic pitch estimation. We solve the task by training a Siamese architecture on pairs of pitch-shifted Constant-Q Transform (CQT) frames. The network is trained with a novel class-based equivariance criterion, using the (known) shift between the frames as a supervision signal. This strategy enables our model to directly predict pitch distributions in a fully self-supervised way, without access to any annotated data. Despite being extremely lightweight (30k parameters), our model significantly outperforms previous self-supervised baselines and is on par with supervised ones.Then, we investigate the paradigm of Joint-Embedding Predictive Architectures (JEPA) for music-related applications. From given context/target pairs, an encoder and a predictor are jointly trained to produce latent representations of the pair and predict the target representations from the context ones. Opposite to contrastive learning, JEPAs do not require any negative samples and can learn a richer latent space thanks to the predictor. By using the different sources of music tracks as context/target pairs, we show that these architectures can capture both local and global musical features, making them useful for a variety of tasks such as compatibility estimation, musical stem retrieval, track alignment, genre classification, auto-tagging and beat-tracking.
• Identifying quantum resources in encoded computations
  
  • Davis Jack
  • Fabre Nicolas
  • Chabaud Ulysse
  , 2024. What is the origin of quantum computational advantage? Providing answers to this far-reaching question amounts to identifying the key properties, or quantum resources, that distinguish quantum computers from their classical counterparts, with direct applications to the development of quantum devices. The advent of universal quantum computers, however, relies on error-correcting codes to protect fragile logical quantum information by robustly encoding it into symmetric states of a quantum physical system. Such encodings make the task of resource identification more difficult, as what constitutes a resource from the logical and physical points of view can differ significantly. Here we introduce a general framework which allows us to correctly identify quantum resources in encoded computations, based on phase-space techniques. For a given quantum code, our construction provides a Wigner function that accounts for how the symmetries of the code space are contained within the transformations of the physical space, resulting in an object capable of describing the logical content of any physical state, both within and outside the code space. We illustrate our general construction with the Gottesman--Kitaev--Preskill encoding of qudits with odd dimension. The resulting Wigner function, which we call the Zak-Gross Wigner function, is shown to correctly identify quantum resources through its phase-space negativity. For instance, it is positive for encoded stabilizer states and negative for the bosonic vacuum. We further prove several properties, including that its negativity provides a measure of magic for the logical content of a state, and that its marginals are modular measurement distributions associated to conjugate Zak patches. (10.48550/arXiv.2407.18394)
  DOI : 10.48550/arXiv.2407.18394
• Fundamental limits and practical algorithms for wireless distributed computation and estimation systems
  
  • Bi Yue
  , 2025. Distributed systems are at the core of modern computing applications, enabling collaborative task execution across interconnected components. However, their distributed nature presents major challenges in communication efficiency. This thesis addresses these challenges by analyzing the theoretical limits of information and proposing solutions to enhance the performance of distributed computing (DC) and distributed estimation systems. In DC systems, task parallelization significantly reduces execution time, but the shuffle phase remains a bottleneck, particularly in wireless networks. This thesis introduces new coding schemes to optimize the computation-communication tradeoff in such environments, leveraging interference alignment (IA) and establishing theoretical bounds. Regarding distributed estimation, where multiple nodes collaborate to estimate a common parameter, two scenarios are explored: with and without a fusion center. In the first case, a framework is proposed to optimize multi-bit quantization and minimize the Cramer-Rao bound. In the second, a synchronous al- ´ gorithm with stochastic activation is developed to improve convergence while reducing data collisions. In summary, this thesis deepens the understanding of theoretical limits and proposes practical coding strategies for distributed systems, enhancing their efficiency and robustness across various environments.
• Statistical wave field theory: Curvature term
  
  • Badeau Roland
  Journal of the Acoustical Society of America, Acoustical Society of America, 2025, 157 (3), pp.1650-1664. In a recent research paper, we introduced the statistical wave field theory, which establishes the statistical laws of waves propagating in a bounded volume. These laws hold after many reflections on the boundary surface and at high frequency. The statistical wave field theory is the first statistical theory of reverberation that provides the closed-form expression of the power distribution and the correlations of the wave field jointly over time, frequency and space, in terms of the geometry and the specific admittance of the boundary surface. In this paper, we refine the theory predictions, by investigating the impact of a curved boundary surface on the wave field statistics. In particular, we provide an improved closed-form expression of the reverberation time in room acoustics that holds at lower frequency. (10.1121/10.0036053)
  DOI : 10.1121/10.0036053
• Approches hybrides en cryptographie quantique
  
  • Alléaume Romain
  • Nemoz Tristan
  Photoniques, EDP Sciences, 2025 (130), pp.46-48. La cryptographie quantique s’est largement définie comme visant une sécurité inconditionnelle, en alternative à la cryptographie dite classique reposant sur la difficulté calculatoire conjecturée de certains problèmes mathématiques. Plutôt que d’opposer cryptographie quantique et classique, hybrider approches calculatoires post-quantiques (PQC) et cryptographie quantique (QC) ouvre des perspectives nouvelles, pour une cryptographie pratique, plus sûre et offrant plus de fonctionnalités. (10.1051/photon/202513046)
  DOI : 10.1051/photon/202513046
• Is This You, LLM? Recognizing AI-written Programs with Multilingual Code Stylometry
  
  • Gurioli Andrea
  • Gabbrielli Maurizio
  • Zacchiroli Stefano
  , 2025. With the increasing popularity of LLM-based code completers, like GitHub Copilot, the interest in automatically detecting AI-generated code is also increasing-in particular in contexts where the use of LLMs to program is forbidden by policy due to security, intellectual property, or ethical concerns. We introduce a novel technique for AI code stylometry, i.e., the ability to distinguish code generated by LLMs from code written by humans, based on a transformer-based encoder classifier. Differently from previous work, our classifier is capable of detecting AI-written code across 10 different programming languages with a single machine learning model, maintaining high average accuracy across all languages (84.1% ± 3.8%). Together with the classifier we also release H-AIRosettaMP, a novel open dataset for AI code stylometry tasks, consisting of 121 247 code snippets in 10 popular programming languages, labeled as either human-written or AI-generated. The experimental pipeline (dataset, training code, resulting models) is the first fully reproducible one for the AI code stylometry task. Most notably our experiments rely only on open LLMs, rather than on proprietary/closed ones like ChatGPT. (10.48550/arXiv.2412.14611)
  DOI : 10.48550/arXiv.2412.14611
• Bidding Efficiently in Simultaneous Ascending Auctions With Budget and Eligibility Constraints Using Simultaneous Move Monte Carlo Tree Search
  
  • Pacaud Alexandre
  • Bechler Aurelien
  • Coupechoux Marceau
  IEEE Transactions on Games, Institute of Electrical and Electronics Engineers, 2025, 17 (1), pp.210-223. For decades, simultaneous ascending auction (SAA) has been the most popular mechanism used for spectrum auctions. It has recently been employed by many countries for the allocation of 5G licences. Although SAA presents relatively simple rules, it induces a complex strategic game for which the optimal bidding strategy is unknown. Considering the fact that sometimes billions of euros are at stake in an SAA, establishing an efficient bidding strategy is crucial. In this work, we model the auction as a n-player simultaneous move game with complete information and propose the first efficient bidding algorithm that tackles simultaneously its four major strategic issues: the exposure problem, the own price effect, budget constraints, and the eligibility management problem. Our solution, called SMSα, is based on simultaneous move Monte Carlo Tree Search and relies on a new method for the prediction of closing prices. By introducing a new reward function in SMSα, we give the possibility to bidders to define their own level of risk-aversion. Through extensive numerical experiments on instances of realistic size, we show that SMSα largely outperforms state-of-the-art algorithms, notably by achieving higher expected utility while taking less risks. (10.1109/TG.2024.3424246)
  DOI : 10.1109/TG.2024.3424246
• Unlearning Personal Data from a Single Image
  
  • de Min Thomas
  • Mancini Massimiliano
  • Lathuilière Stéphane
  • Roy Subhankar
  • Ricci Elisa
  Transactions on Machine Learning Research Journal, [Amherst Massachusetts]: OpenReview.net, 2022, 2025. Machine unlearning aims to erase data from a model as if the latter never saw them during training. While existing approaches unlearn information from complete or partial access to the training data, this access can be limited over time due to privacy regulations. Currently, no setting or benchmark exists to probe the effectiveness of unlearning methods in such scenarios. To fill this gap, we propose a novel task we call One-Shot Unlearning of Personal Identities (1-SHUI) that evaluates unlearning models when the training data is not available. We focus on unlearning identity data, which is specifically relevant due to current regulations requiring personal data deletion after training. To cope with data absence, we expect users to provide a portraiting picture to aid unlearning. We design requests on CelebA, CelebA-HQ, and MUFAC with different unlearning set sizes to evaluate applicable methods in 1-SHUI. Moreover, we propose MetaUnlearn, an effective method that meta-learns to forget identities from a single image. Our findings indicate that existing approaches struggle when data availability is limited, especially when there is a dissimilarity between the provided samples and the training data.
• ELMGS: Enhancing Memory and Computation Scalability Through coMpression for 3D Gaussian Splatting
  
  • Ali Muhammad Salman
  • Bae Sung-Ho
  • Tartaglione Enzo
  , 2025, pp.2591-2600. 3D models have recently been popularized by the potentiality of end-to-end training offered first by Neural Radiance Fields and most recently by 3D Gaussian Splatting models. The latter has the big advantage of naturally providing fast training convergence and high editability. However, as the research around these is still in its infancy, there is still a gap in the literature regarding the model's scalability. In this work, we propose an approach enabling both memory and computation scalability of such models. More specifically, we propose an iterative pruning strategy that removes redundant information encoded in the model. We also enhance compressibility for the model by including a differentiable quantization and entropy coding estimator in the optimization strategy. Our results on popular benchmarks showcase the effectiveness of the proposed approach and open the road to the broad deployability of such a solution even on resource-constrained devices. (10.1109/WACV61041.2025.00257)
  DOI : 10.1109/WACV61041.2025.00257
• CATALOG: A Camera Trap Language-guided Contrastive Learning Model
  
  • Santamaria Julian
  • Isaza Claudia
  • Giraldo Jhony
  , 2025, pp.1197-1206. <div><p>Foundation Models (FMs) have been successful in various computer vision tasks like image classification, object detection and image segmentation. However, these tasks remain challenging when these models are tested on datasets with different distributions from the training dataset, a problem known as domain shift. This is especially problematic for recognizing animal species in camera-trap images where we have variability in factors like lighting, camouflage and occlusions. In this paper, we propose the Camera Trap Language-guided Contrastive Learning (CATALOG) model to address these issues. Our approach combines multiple FMs to extract visual and textual features from camera-trap data and uses a contrastive loss function to train the model. We evaluate CATALOG on two benchmark datasets and show that it outperforms previous state-of-theart methods in camera-trap image recognition, especially when the training and testing data have different animal species or come from different geographical areas. Our approach demonstrates the potential of using FMs in combination with multi-modal fusion and contrastive learning for addressing domain shifts in camera-trap image recognition. The code of CATALOG is publicly available at https://github.com/Julian075/CATALOG.</p></div> (10.1109/WACV61041.2025.00124)
  DOI : 10.1109/WACV61041.2025.00124
• Enabling Incremental SysML Model Verification: Managing Variability and Complexity Through Tagging and Model Reduction
  
  • Sultan Bastien
  • Apvrille Ludovic
  • Hotescu Oana
  • de Saqui-Sannes Pierre
  , 2025, pp.224-233. <div><p>Designing complex software systems with model-based approaches encounters the recognized state space explosion problem. Typically, only a subset of models can be formally verified, forcing reliance on simulation or testing to verify the entire system. Furthermore, most formal verification tools require a complete reevaluation of properties after even minor modifications to a model. Although incremental formal verification, particularly the incremental model-checking approach of TTool, has been proposed, it still requires modelers to manually select sub-models not facing state space explosion. Unfortunately, this manual model selection is susceptible to errors. This paper presents a twofold contribution to SysML models of software product lines. First, we introduce a SysML model tagging feature that enables designers to explicitly differentiate between various subsystems, such as core and optional features. Second, we develop and implement a model reduction algorithm using dependency graphs (DGs). This algorithm automatically deactivate model elements linked to specific tags, removing both the specified elements and all their logical dependencies provided the DG is acyclic. These two contributions are evaluated for their effectiveness in generating model variants. Together, they facilitate the creation of a core model and an associated set of models, each extended by additional model elements, and make it possible to rely on incremental model-checking. We have implemented the contributions in TTool and applied it to an integrated modular avionics system. This application enables to compare-both manual and automated-model reduction strategies and assess their benefits for TTool users. a</p></div> (10.5220/0013182300003896)
  DOI : 10.5220/0013182300003896
• WiGNet: Windowed Vision Graph Neural Network
  
  • Spadaro Gabriele
  • Grangetto Marco
  • Fiandrotti Attilio
  • Tartaglione Enzo
  • Giraldo Jhony
  , 2024, pp.859-868. In recent years, Graph Neural Networks (GNNs) have demonstrated strong adaptability to various real-world challenges, with architectures such as Vision GNN (ViG) achieving state-of-the-art performance in several computer vision tasks. However, their practical applicability is hindered by the computational complexity of constructing the graph, which scales quadratically with the image size. In this paper, we introduce a novel Windowed vision Graph neural Network (WiGNet) model for efficient image processing. WiGNet explores a different strategy from previous works by partitioning the image into windows and constructing a graph within each window. Therefore, our model uses graph convolutions instead of the typical 2D convolution or self-attention mechanism. WiGNet effectively manages computational and memory complexity for large image sizes. We evaluate our method in the ImageNet-1k benchmark dataset and test the adaptability of WiGNet using the CelebA-HQ dataset as a downstream task with higher-resolution images. In both of these scenarios, our method achieves competitive results compared to previous vision GNNs while keeping memory and computational complexity at bay. WiGNet offers a promising solution toward the deployment of vision GNNs in real-world applications. We publicly released the code and pre-trained models at https://github.com/EIDOSLAB/WiGNet. (10.1109/WACV61041.2025.00093)
  DOI : 10.1109/WACV61041.2025.00093
• Efficient Progressive Image Compression with Variance-Aware Masking
  
  • Presta Alberto
  • Tartaglione Enzo
  • Fiandrotti Attilio
  • Grangetto Marco
  • Cosman Pamela
  , 2025, pp.7692-7700. Learned progressive image compression is gaining momentum as it allows improved image reconstruction as more bits are decoded at the receiver. We propose a progressive image compression method in which an image is first represented as a pair of base-quality and top-quality latent representations. Next, a residual latent representation is encoded as the element-wise difference between the top and base representations. Our scheme enables progressive image compression with element-wise granularity by introducing a masking system that ranks each element of the residual latent representation from most to least important, dividing it into complementary components, which can be transmitted separately to the decoder in order to obtain different reconstruction quality. The masking system does not add further parameters or complexity. At the receiver, any elements of the top latent representation excluded from the transmitted components can be independently replaced with the mean predicted by the hyperprior architecture, ensuring reliable reconstructions at any intermediate quality level. We also in-troduced Rate Enhancement Modules (REMs), which refine the estimation of entropy parameters using already decoded components. We obtain results competitive with state-of-the-art competitors, while significantly reducing computational complexity, decoding time, and number of parameters. (10.1109/WACV61041.2025.00747)
  DOI : 10.1109/WACV61041.2025.00747
• Till the Layers Collapse: Compressing a Deep Neural Network Through the Lenses of Batch Normalization Layers.
  
  • Liao Zhu
  • Hezbri Nour
  • Quétu Victor
  • Nguyen Van-Tam
  • Tartaglione Enzo
  , 2025, 39 (18), pp.18702-18710. Today, deep neural networks are widely used since they can handle a variety of complex tasks. Their generality makes them very powerful tools in modern technology. However, deep neural networks are often overparameterized. The usage of these large models consumes a lot of computation resources. In this paper, we introduce a method called Till the Layers Collapse (TLC), which compresses deep neural networks through the lenses of batch normalization layers. By reducing the depth of these networks, our method decreases deep neural networks' computational requirements and overall latency. We validate our method on popular models such as Swin-T, MobileNet-V2, and RoBERTa, across both image classification and natural language processing (NLP) tasks. (10.1609/aaai.v39i18.34058)
  DOI : 10.1609/aaai.v39i18.34058
• HYGENE: A Diffusion-Based Hypergraph Generation Method
  
  • Gailhard Dorian
  • Tartaglione Enzo
  • Naviner Lirida
  • Giraldo Jhony
  , 2025, 39 (16), pp.16682-16690. Hypergraphs are powerful mathematical structures that can model complex, high-order relationships in various domains, including social networks, bioinformatics, and recommender systems. However, generating realistic and diverse hypergraphs remains challenging due to their inherent complexity and lack of effective generative models. In this paper, we introduce a diffusion-based Hypergraph Generation (HYGENE) method that addresses these challenges through a progressive local expansion approach. HYGENE works on the bipartite representation of hypergraphs, starting with a single pair of connected nodes and iteratively expanding it to form the target hypergraph. At each step, nodes and hyperedges are added in a localized manner using a denoising diffusion process, which allows for the construction of the global structure before refining local details. Our experiments demonstrated the effectiveness of HYGENE, proving its ability to closely mimic a variety of properties in hypergraphs. To the best of our knowledge, this is the first attempt to employ diffusion models for hypergraph generation. (10.1609/aaai.v39i16.33833)
  DOI : 10.1609/aaai.v39i16.33833