Publications

2025

• Why Older Adults Integrate Digital Health Technology Into Their Everyday Lives: The Role of Basic Psychological Needs
  
  • Yee Andrew
  • Kwok Bernice
  • Ng Janelle
  • Loy Guan Peng
  • Ng Li Yin
  • Sng Jeremy
  • Perrault Simon
  • Lim Kwan Hui
  • Subburaj Karupppasamy
  Human Behavior and Emerging Technologies, Wiley, 2025, 2025 (1), pp.9711661: 1-13. There is growing consensus on the potential of digital health technology to improve individual health and well‐being outcomes. Existing research examining digital health technology adoption among underserved populations has largely relied upon theoretical lenses with an emphasis on perceived benefits and costs at the preadoption or initial use stage. This study demonstrates an alternative approach, relying on basic psychological needs as an explanatory framework to explore how and why older adults integrate digital health technologies into their everyday lives in a sustained manner. A whole‐of‐community approach was taken to examine this question, with 17 focus groups held with 107 older adults, volunteers, and the people who work with them. Overall, we found that the integration of digital health technology into the day‐to‐day life of older adults is contingent on the different ways in which it satisfies or frustrates the basic needs of competence, autonomy, and relatedness. The social environment, user interface design choices, reward, and feedback systems were found to both satisfy and frustrate needs. Researchers and designers intending to study or implement digital health technology for older adults ought to consider how different implementation decisions impact psychological needs. (10.1155/hbe2/9711661)
  DOI : 10.1155/hbe2/9711661
• Are We Unbiased Yet? A Survey on Model Debiasing for Image Classification
  
  • Ciranni Massimiliano
  • Nahon Rémi
  • Marinelli Davide
  • Murino Vittorio
  • Tartaglione Enzo
  • Pastore Vito Paolo
  , 2025. <div><p>The rapid advancement of machine learning and artificial intelligence (AI) has led to transformative changes across various domains that involve Computer Vision, including healthcare, finance, and autonomous systems. Despite their potential, these technologies raise significant concerns related to privacy, fairness, transparency, reliability, and ethics, especially when dealing with biased data. In recent years, the computer vision community has shown an increasing interest in model debiasing, with strategies designed to identify and mitigate the dependency of deep neural networks on shortcuts corresponding to bias. This survey aims to bring order to this field by providing a comprehensive review of model debiasing methods, with a particular focus on their application in Computer Vision. After defining model bias and discussing its implications, we describe the typical evaluation benchmarks, serving as a foundation for understanding the complexities involved in this critical area of research. We further delve into debiasing methods, providing a categorization while discussing their effectiveness and limitations. Besides, contemporary research trends are explored, among which the growing interest in explainable AI techniques and the ethical implications of biased models are predominant. Throughout this survey, we highlight milestones and open challenges in model debiasing, in an attempt to provide a reference for both experts and researchers interested in this context.</p></div>
• Shape-changing pneumatic interfaces for emotion and trust in Human-Computer interaction
  
  • Liu Yang
  , 2025. This dissertation explores the transformative potential of silicone-based pneumatic haptic interfaces to enhance emotional engagement, trust, and non-verbal communication across three interconnected domains: cinematic experiences, automated driving systems, and human-robot interaction (HRI).At the intersection of affective computing and human-computer interaction (HCI), this research investigates how bio-inspired haptic feedback—mediated through dynamic shape-changing and surface textures—can deepen user experiences by aligning tactile stimuli with contextual emotions, situational awareness, and physiological cues.The first part of this work addresses cinematic emotion enhancement, where traditional audiovisual narratives are augmented with pneumatic haptic feedback to create immersive, multi-sensory cinematic storytelling. Drawing inspiration from biological structures such as sea urchin spines and cat paw pads, the designed interface employs spiked and smooth textures to modulate emotional valence, while inflation rhythms regulate arousal. Controlled experiments reveal that congruent haptic stimuli significantly enhance emotional responses, with high-frequency, goosebump-textured feedback intensifying excitement and low-frequency, smooth textures enhancing calmness. Conversely, non-congruent haptic feedback has negative effects on emotional alignment, underscoring the importance of context-aware design. These findings validate the role of pneumatic haptics as a medium for emotional communication, extending beyond functional alerts to evoke embodied, affective experiences.The second part transitions to automated driving systems, focusing on the design of bio-inspired pneumatic interfaces for Takeover Requests in Level 3 autonomous vehicles. The Pneumatic Haptic Silicon Sac steering wheel, inspired by animal inflation mechanisms like frog vocal sacs, employs gentle rhythmic feedback to reduce cognitive load and foster trust during control transitions. Comparative studies with vibrotactile systems demonstrate that pneumatic cues elicit calmer, smoother driver responses, while vibrotactile alerts achieve faster reaction times at the cost of increased stress.However, directional ambiguity in pneumatic feedback—particularly for left lane changes—highlights the need for hybrid modalities that balance urgency and comfort. These insightsadvance the development of adaptive haptic systems based on situational demands.The third part explores non-verbal human-robot interaction, introducing a pneumatic haptic interface that simulates physiologically-inspired rhythms such as breathing and heartbeat. Integrated into a humanoid robot's chest, this silent, silicone-based system combines voluntary gestures with subconscious physiological cues to enhance emotion recognition. Experimental results show a 16% improvement in accuracy when physiologically-inspired haptics are paired with gestures, particularly for down-tune emotional states like sadness. Participants described the feedback as "intimate" and "lifelike," emphasizing its potential to foster trust and affective rapport in social robotics.Limitations, such as cultural variability in tactile perception and directional biases in driving interfaces, underscore opportunities for future work. Adaptive systems leveraging real-time user states, cross-cultural studies of emotional haptics and sustainable material innovations are critical next steps. By unifying principles of biomimicry, emotional design, and human factors, this research lays the groundwork for haptic technologies that transcend functionality to foster deeper, more intuitive human-computer relationships. (10.70675/d84aad4bzddc1z477cz97f1z3e60e799f8ce)
  DOI : 10.70675/d84aad4bzddc1z477cz97f1z3e60e799f8ce
• The NeWMe Corpus: A gold standard corpus for the study of Word Meaning Negotiation
  
  • Soler Aina Garí
  • Myrendal Jenny
  • Clavel Chloé
  • Larsson Staffan
  , 2025. Abstract Word Meaning Negotiation (WMN) sequences occur when participants focus on clarifying or negotiating the meaning of a word or phrase, often prompted by questions or challenges. These interactions temporarily shift the conversation to explore nuances of meaning - sometimes resulting in quick clarification when due to insufficient understanding of word meaning, and other times leading to extended debates, such as disagreements on what a word can or should mean. This paper presents the largest and freely available manually annotated corpus of WMNs to date, encompassing spoken dyadic and multiparty conversations as well as online discussions. Our methodology combines searching for WMNs using regular expressions with a detailed annotation scheme that categorizes WMNs into types triggered by non-understanding (NONs: Non-understanding WMN) or disagreement (DINs: Disagreement WMN), and distinguishes between negotiations of situated and potential meanings. We also annotate incomplete negotiations and related phenomena, and analyze inter-annotator agreement to evaluate the reliability of the annotation schema. Preliminary investigations of WMNs in the corpus reveal distinct patterns in WMNs across contexts, with NONs prevalent in spoken interactions and DINs dominating online debates. This resource lays a foundation for studying semantic alignment, developing automated WMN detection, and creating adaptive dialogue systems. Our findings highlight the complexity of WMNs and provide practical insights for their identification and analysis. (10.21203/rs.3.rs-5975927/v1)
  DOI : 10.21203/rs.3.rs-5975927/v1
• On the complexity of sabotage games for network security
  
  • Raju Dhananjay
  • Bakirtzis Georgios
  • Topcu Ufuk
  IEEE Transactions on Networking, ieee, 2025, 34, pp.2897-2910. Securing dynamic networks against adversarial actions is challenging because of the need to anticipate and counter strategic disruptions by adversarial entities within complex network structures. Traditional game-theoretic models, while insightful, often fail to model the unpredictability and constraints of real-world threat assessment scenarios. We refine sabotage games to reflect the realistic limitations of the saboteur and the network operator. By transforming sabotage games into reachability problems, our approach allows applying existing computational solutions to model realistic restrictions on attackers and defenders within the game. Modifying sabotage games into dynamic network security problems successfully captures the nuanced interplay of strategy and uncertainty in dynamic network security. Theoretically, we extend sabotage games to model network security contexts and thoroughly explore if the additional restrictions raise their computational complexity, often the bottleneck of game theory in practical contexts. Practically, this research sets the stage for actionable insights for developing robust defense mechanisms by understanding what risks to mitigate in dynamic networks under threat. (10.1109/TON.2025.3628015)
  DOI : 10.1109/TON.2025.3628015
• Multi-view 3D reconstruction in SAR imaging using inverse rendering
  
  • Barbier--Renard Emile
  , 2025. Synthetic Aperture Radar (SAR) imaging is a powerful tool for Earth observation, offering the ability to capture data regardless of daylight or weather conditions. However, traditional methods for 3D reconstruction from SAR images often rely on interferometric techniques, which require strict constraints to guarantee the coherence of the acquisitions.In this thesis, we explore the potential of adapting the Differentiable Inverse Rendering (DIR) approach to SAR. These physics-informed methods have seen recent advancements in the optical domain following the publication of Neural Radiance Fields (NeRF) but have not yet been applied to active imaging systems. By developing a DIR technique for SAR intensity images, we aim to achieve a flexible solution for 3D reconstruction from a few views.First, we introduce a novel differentiable SAR rendering model, a central requirement for developing a DIR method. Our model uses a rasterisation strategy to efficiently synthesise SAR images from a Digital Surface Model (DSM) and a map of backscattering coefficients. It leverages the distinct geometry of SAR to ensure the differentiability of the rendering process, in particular for the computation of shadows. We validate our model using EMPRISE®, a state-of-the-art RADAR simulator developed by ONERA.We subsequently propose two different reconstruction strategies. The first one is directly inspired by NeRF and optimises the parameters of a neural network modelling the scene. The second strategy simplifies the tuning of hyperparameters and improves explainability by modelling the scene with stacks of grids of increasing resolutions. Both strategies use the unique properties of our rendering model to enable coarse-to-fine reconstructions, ensuring faster optimisation and more accurate reconstruction of structures of different scales.We demonstrate our method on images simulated using EMPRISE® under different conditions, before showing applications on Sentinel-1 ascending/descending image pairs. (10.70675/c429b493zd350z474az8f79zaf623a4796bf)
  DOI : 10.70675/c429b493zd350z474az8f79zaf623a4796bf
• Deep Learning for Anatomically-Consistent Airway Tree Modeling : Visual Features, Loss Functions, Learning Strategies, and a Novel Detection Task
  
  • Keshavarzi Ali
  , 2025. The human airway tree is a complex, hierarchically organized branching network that conducts air from the trachea to the alveoli through successive generations of bronchi and bronchioles. Its geometry and connectivity are tightly linked to pulmonary function, and structural alterations in this tree serve as key indicators of diseases such as chronic obstructive pulmonary disease (COPD), asthma, cystic fibrosis, and interstitial lung disease. Accurate and anatomically consistent modeling of the airway tree from chest Computed Tomography (CT) is therefore essential for quantitative pulmonary analysis, disease monitoring, bronchoscopic surgery planning, and the development of digital airway twins.However, this task remains challenging due to the fine and tortuous nature of distal bronchi, high anatomical variability, and substantial differences in acquisition protocols across scanners and populations. Conventional deep learning approaches often struggle to reconstruct a continuous airway tree and to generalize across domains, producing anatomically incomplete or fragmented structures that limit their clinical reliability.This thesis introduces a series of complementary methods to enhance the robustness, anatomical fidelity, and generalizability of deep airway-tree modeling through both segmentation and detection perspectives, reformulating the task from pure segmentation to structured detection. First, convolutional sparse priors are proposed as data-driven structural biases to encode recurring airway patterns. The learned sparse dictionaries generate compact representations that capture generalizable geometric primitives, thereby improving the model's understanding of bronchial morphology—particularly in few-shot settings where data scarcity hinders structural learning—and stabilizing training across heterogeneous datasets. Second, a Boundary-Emphasized Loss (BEL) is designed to reinforce morphological boundaries and prevent distal branch breakages through dynamic weighting, enabling topology-enhancing segmentation without explicit centerline supervision. Third, a complexity-based curriculum domain adaptation framework is developed to improve performance on the source domain and progressively transfer knowledge between healthy and pathological cohorts by leveraging scan-level complexity scores to guide training from simple to challenging cases. Finally, this thesis presents BifDet, the first large-scale dataset and benchmark for 3D airway bifurcation detection, comprising over 7,500 annotated bifurcations and establishing standardized evaluation protocols for bifurcation analysis.Comprehensive experiments on healthy and diseased cohorts demonstrate consistent improvements in distal airway reconstruction, topological continuity, and cross-domain generalization. Collectively, these contributions establish a robust and anatomically grounded framework for airway-tree modeling, bridging the gap between geometric accuracy and topological completeness. Beyond airway analysis, the proposed methodologies provide a foundation for topology-aware and structure-preserving modeling in other tubular and multi-organ systems, paving the way toward interpretable and clinically reliable digital anatomical twins. (10.70675/976008a6zb55bz4f9dza453z091f9cd95be6)
  DOI : 10.70675/976008a6zb55bz4f9dza453z091f9cd95be6
• Side channel resistance of Europe's choices for lattice-based post-quantum cryptography : FALCON and FrodoKEM
  
  • Berthet Pierre-Augustin
  , 2025. There is a high chance that a quantum computer will be able to break current asymmetric cryptography primitives in the coming decade. To address this threat, new cryptography primitives, i.e., Post-Quantum Cryptography (PQC), are studied. In this thesis, we look at lattice-based PQC selected by European agencies, specifically the signature FALCON and the Key Encapsulation Mechanism (KEM) FrodoKEM. We study their resistance to Side Channel Analysis (SCA).The first contribution of this thesis is the study of SCA countermeasures for the FALCON signature. This signature uses floating-point arithmetic, and a masking countermeasure for the floor function is proposed. Similarly, a masking design based on the Newton-Raphson method is used to protect the inversion. Both operands are used in FALCON's Gaussian sampler. Formal security proofs using the MIMO-SNI model are provided to demonstrate the resistance of the design.In another chapter, ciphertext malleability in lattice-based KEMs is used to recover the secret message by targeting the decoding function in FrodoKEM's decapsulation. By combining chosen ciphertexts and side channel analysis, new strategies are proposed to improve the efficiency of the attack: more optimal choices of ciphertext alterations, adaptative attack, partial recovery with brute-force of remaining bits.A final contribution of this manuscript proposes to re-use the same ciphertext malleability but as a countermeasure to protect the secret message. By performing the attack in a random and controlled manner as the defender, it is possible to disrupt the attacker. The countermeasure requires several adjustments, and it is recommended to deploy it alongside a shuffling countermeasure to protect the secret key as well. (10.70675/10827b58zabb1z4b34za600zf6fd733bc828)
  DOI : 10.70675/10827b58zabb1z4b34za600zf6fd733bc828
• Acoustics-aware hybrid deep neural dereverberation
  
  • Bahrman Louis
  , 2025. The aim of this thesis is to leverage room acoustics models in deep-learning-based approaches for dereverberation. Audio signals are often altered by reverberation effects induced by objects and walls of the room in which they propagate, leading to a loss in intelligibility. However, most deep learning methods developed to tackle this problem can be considered as black-box systems, as they are purely data-driven and not interpretable from a physical perspective. After studying whether neural dereverberators are consistent with physical reverberation models, we propose two hybrid approaches to train a dereverberation model in a physically realistic manner. The first one regularizes the training loss to encourage a deep neural network to produce realistic solutions, and the second is motivated by a maximum-likelihood formulation of the problem and consists in an unsupervised learning strategy that integrates a reverberation model into a deep learning framework. (10.70675/5f2d2e3fzbe12z428fzb7afzfcd5584e781f)
  DOI : 10.70675/5f2d2e3fzbe12z428fzb7afzfcd5584e781f
• RibbonSculpt: Voronoi Ball based 3D Sculpting from Sparse VR Ribbons
  
  • Sureshkumar Anandhu
  • Parakkat Amal Dev
  • Bonneau Georges-Pierre
  • Hahmann Stefanie
  • Cani Marie-Paule
  , 2025, pp.1-11. We introduce RibbonSculpt, the first method for interactive freeform shape design in VR through progressive sketching of sparse, oriented ribbons. Instead of reconstructing a surface from a fully drawn VR sketch, our method allows the real-time creation and progressive refinement of a closed surface of any topological genus, thanks to the continuous update of a volumetric proxy. The latter corresponds to a filtered subset of the Voronoi balls defined by the user-sketched ribbons. At each visualization step, a mesh extracted from the proxy is beautified through Laplacian-based energy minimization, yielding a smooth surface that interpolates the ribbons. Guided by this surface, users can easily refine their design by adding or removing ribbons, which sculpts, in return, the set of Voronoi balls forming the proxy. Our results, supported by user studies, show that RibbonSculpt allows VR users to easily and quickly draft the 3D shapes they have in mind. (10.1145/3757377.3763994)
  DOI : 10.1145/3757377.3763994
• Deep learning for SAR image enhancement : multi-modal fusion and interferometric denoising
  
  • Gaya Victor
  , 2025. This thesis focuses on the denoising of synthetic aperture radar (SAR) images, targeting speckle in amplitude imaging and phase noise in SAR interferometry (InSAR). These issues are central to many remote sensing applications, from land use mapping to terrestrial displacement estimation.While traditional methods based on statistical models and non-local filters have made it possible to better preserve structures, they are reaching their limits in terms of current requirements for accuracy and generalization. Deep learning offers new perspectives by learning directly from representations adapted to the specificities of the radar signal.We propose two major methodological contributions. The first concerns SAR denoising through multimodal SAR-optical fusion. By exploiting the complementarity between modalities, we develop a self-supervised approach that guides filtering using optical data, preserving geometric and textural details. This method does not require ground truth by exploiting the independence between the real and imaginary parts of the radar data.The second contribution concerns the joint denoising of phase and coherence in InSAR, a critical step for estimating displacement or topography. Based on projected multi-channel data, our method reduces the multichannel problem to single-channel processing while integrating appropriate spatial regularization, stabilizing the estimation despite strong radiometric or temporal variations.The experimental results demonstrate significant quantitative and qualitative improvements for downstream applications (deformation detection, environmental monitoring, mapping). This thesis bridges the gap between classical radar signal processing and modern machine learning, highlighting the complementarity between physical modeling, multimodal data, and self-supervision. (10.70675/cb154572z701fz4e19z83dezf85196761d51)
  DOI : 10.70675/cb154572z701fz4e19z83dezf85196761d51
• Adaptive Compression : From Visual Data to Efficient and Transferable Models
  
  • Spadaro Gabriele
  , 2025. The exponential growth of visual content has made compression a fundamental challenge in modern communication systems. While traditional codecs achieved remarkable success, their rigid design limits their performance. Learned Image Compression emerged as a data-driven alternative, in which models directly minimize a rate-distortion loss function. Despite their results, these methods suffer from limited flexibility, since models are trained to attain a fixed rate-distortion trade-off, as well as poor generalization across novel visual domains and a lack of perceptual control. This thesis aims to investigate deep learning-based compression methods and to address the key limitations that currently hinder their deployment. Moreover, we go beyond the traditional definition of compression, proposing strategies that enhance efficiency, adaptability, and generalization capabilities by compressing the models and their internal representations.In this context, we show how the integration of learning-based modules can significantly enhance compression performance. This improvement occurs not only by replacing specific components of standardized codecs, but also by defining end-to-end methods in which the entire compression pipeline consists of learnable modules. Interestingly, in this latter scenario, we demonstrate how the use of alternative graph-based paradigms can be effectively applied for compression tasks, while also showing their potential as general-purpose backbones for visual feature extraction.Beyond improving compression, this thesis also proposes a unified adapter-based strategy to overcome the structural limitations of learned codecs. Considering a model-adaptation perspective, we demonstrate how adapters enable continuous control over the rate-distortion and distortion-perception trade-offs. Furthermore, they enhance the generalization capability of a pre-trained model to novel visual domains.These advances make learned codecs more versatile for heterogeneous real-world applications.Finally, we demonstrate how model and representation pruning methods allow not only to reduce the complexity of a model, but also to improve generalization and transferability capabilities of a pre-trained model.Here, the notion of compression is extended to models and their representations. This perspective highlights its role not only as a tool for efficiency but also as a principle for designing adaptive and robust neural models. (10.70675/ae4a7203z4b30z4b6azb64dz25b215b2f878)
  DOI : 10.70675/ae4a7203z4b30z4b6azb64dz25b215b2f878
• Approximating Queries on Probabilistic Graphs
  
  • Amarilli Antoine
  • van Bremen Timothy
  • Gaspard Octave
  • Meel Kuldeep S.
  Logical Methods in Computer Science, Logical Methods in Computer Science Association, 2025, 21 (4). Query evaluation over probabilistic databases is notoriously intractable -- not only in combined complexity, but often in data complexity as well. This motivates the study of approximation algorithms, and particularly of combined FPRASes, with runtime polynomial in both the query and instance size. In this paper, we focus on tuple-independent probabilistic databases over binary signatures, i.e., probabilistic graphs, and study when we can devise combined FPRASes for probabilistic query evaluation. We settle the complexity of this problem for a variety of query and instance classes, by proving both approximability results and (conditional) inapproximability results doubled with (unconditional) DNNF provenance circuit size lower bounds. This allows us to deduce many corollaries of possible independent interest. For example, we show how the results of Arenas et al. on counting fixed-length strings accepted by an NFA imply the existence of an FPRAS for the two-terminal network reliability problem on directed acyclic graphs: this was an open problem until now. We also show that one cannot extend a recent result of van Bremen and Meel that gives a combined FPRAS for self-join-free conjunctive queries of bounded hypertree width on probabilistic databases: neither the bounded-hypertree-width condition nor the self-join-freeness hypothesis can be relaxed. We last show how our methods can give insights on the evaluation and approximability of regular path queries (RPQs) on probabilistic graphs in the data complexity perspective, showing in particular that some of them are (conditionally) inapproximable. (10.46298/lmcs-21(4:30)2025)
  DOI : 10.46298/lmcs-21(4:30)2025
• Contextual knowledge representation for neurosymbolic Artifical Intelligence reasoning
  
  • Coumes Simon
  , 2025. The field of Knowledge Representation and Reasoning is concerned with the representation of information about reality in a form that is both human-readable and machine-processable.It has been a part of artificial intelligence since its inception, and has produced many important formalisms and systems.One key aspect of knowledge is the context in which it is expressed. This has been identified early on in the field and matches with our common experience: understanding a statement or judging its validty often require to know in what context it was meant.Historically, there has been some work aiming at producing logics implementing a general notion of context. None of them saw a lot of adoption, in part because they lack either sufficient expressive power or because they were not sufficiently usable.This dissertation presents Qiana, a logic of context powerful enough for almost all types of context representation. It is also compatible with various automated reasoning tools, as is demonstrated by the code provided which allows automated reasoning with Qiana. This makes use of the pre-existing theorem prover Vampire -- though any other compatible prover can freely be used instead.By providing a powerful logic for context representation with the possibility of concrete computations without (much) overhead, Qiana paves the way for larger prevalence of logics of context, making it possible to build other reasoning on top of such logics like has been done --for example-- for epistemic logics or description logics. (10.70675/556210bfzbf29z4847z8f40z0181b318cd1c)
  DOI : 10.70675/556210bfzbf29z4847z8f40z0181b318cd1c
• Digital Signal Processing and Parameter Estimation for Practical Quantum Communications
  
  • Ricard Guillaume
  , 2025. Continuous-Variable Quantum Key Distribution (CV-QKD) is a form of quantum cryptography that encodes information in the amplitude and phase of light, enabling provably secure key sharing over optical channels. Building on previous research, subcarrier multiplexing configurations have been proposed to allow a quantum channel and a classical data channel to share a single optical carrier and transmitter. In these schemes, advanced Digital Signal Processing (DSP) enables the high Signal-to-Noise Ratio (SNR) classical signal to serve as a phase reference for the quantum channel, facilitating phase recovery without dedicated pilot tones. This thesis integrates these concepts into a complete system architecture that combines CV-QKD with One-Time Pad encrypted communication over a single fiber and transmitter, bringing practical implementation closer to a scenario where classical and quantum communication modes can be selected purely via software. A detailed link-budget analysis quantifies trade-offs between classical and quantum data rates and ensures sufficient bandwidth for essential post-processing tasks such as reverse reconciliation and parameter estimation.Beyond system integration and experimental improvements, this work rigorously addresses practical security. Three contributions enhance parameter estimation under realistic conditions. First, the Multi-Level Parameter Estimation scheme allows continuous receiver-noise calibration via an intensity modulator without interrupting key generation, significantly increasing throughput. Second, a DSP-induced mode mismatch is characterized: the receiver's digital filters and hardware response define the measured quantum mode, which may deviate from the sender's calibration and bias critical parameter estimation. Third, the Time-Gated Variance formalism quantifies calibrations in the presence of realistic non-white noise, accounting for the impact of calibration duration on variance estimation. Together, these methods provide a more accurate, quantitative foundation for assessing transmittance and excess noise, improving both security and performance.Although demonstrated on a single-fiber, subcarrier-multiplexed system, these techniques are broadly applicable to other CV-QKD architectures. By combining such system designs with rigorous, quantitative analysis of noise calibrations, this thesis takes a step toward scalable, certifiable quantum-secured communication, bridging the gap between theoretical security and real-world deployment. (10.70675/2c2004bcz2d99z4da8zbb31z78568956852a)
  DOI : 10.70675/2c2004bcz2d99z4da8zbb31z78568956852a
• Bypass Synchronization Primitives: GhostRace Attack and Mitigation on RISC-V
  
  • Khan Mahreen
  • Priscilla Maria
  • Victor Matheus
  • Mushtaq Maria
  • Pacalet Renaud
  • Apvrille Ludovic
  • Naviner Lirida
  , 2025. <div><p>In communication and network security systems, synchronization primitives (such as mutexes, semaphores, and spinlocks) are essential for ensuring safe access to shared data. They prevent data races and enforce memory consistency across threads, which is critical for communication stacks and cryptographic libraries. However, the recent GhostRace vulnerability exploits speculative execution to microarchitecturally bypass synchronization primitives, creating speculative race conditions that leak data across threads. This fundamentally undermines the security guarantees of critical communication infrastructure. While GhostRace is tested on x86 architecture, its implications for RISC-V remain unexplored. RISC-V is increasingly adopted in networking and communication hardware, making its security analysis essential. In this paper, we evaluate GhostRace on RISC-V using both the BeagleV-Fire RISC-V board and the gem5 full-system simulator. We demonstrate successful speculative synchronization bypasses and validate the instruction serialization mitigation. We also analyze microarchitectural behavior, including cache misses and branch mispredictions, using gem5 full system simulations. This paper provides the first comprehensive characterization of GhostRace on RISC-V, highlighting the need for secure hardware-software co-design.</p></div>
• Opcode Analysis of Real Encryption-Based Microarchitectural Attacks Using gem5
  
  • Awais Muhammad
  • Khan Mahreen
  • Mushtaq Maria
  • Naviner Lirida
  • Yahya Jawad Haj
  • Bruguier Florent
  • Pacalet Renaud
  • Apvrille Ludovic
  , 2025, pp.377-384. Can we trust architectural simulators to predict real-world security vulnerabilities? This work utilizes gem5 cycle-accurate simulation to dissect the side-channel leakage of specific AES-NI hardware instructions. By analyzing opcodes like AESENC and AESKEYGENASSIST, we achieved an 89% key recovery rate in simulation, closely mirroring results on physical Intel hardware (96%). Our findings provide a roadmap for identifying microarchitectural footprints and hardening software against Flush+Reload attacks before the code ever touches physical silicon. (10.1109/ComComAp68359.2025.11353163)
  DOI : 10.1109/ComComAp68359.2025.11353163
• Opcode Analysis of Real Encryption-Based Microarchitectural Attacks Using gem5
  
  • Awais Muhammad
  • Khan Mahreen
  • Mushtaq Maria
  • Naviner Lirida
  • Haj-Yahya Jawad
  • Bruguier Florent
  • Pacalet Renaud
  • Apvrille Ludovic
  , 2025, pp.377-384. Modern processors use performance-boosting techniques like speculative execution and multi-level caching. Unfortunately, these same optimizations can unintentionally create security vulnerabilities known as microarchitectural sidechannels. The core novelty of this work is a precise, simulationdriven method to pinpoint exactly how individual cryptographic instructions leak information. We present a systematic analysis of common AES instructions (such as AESENC, AESDEC, and AESKEYGENASSIST), measuring their vulnerability to Flush+Reload side-channel attacks. By using cycle-accurate gem5 simulations, we were able to isolate and analyze each instruction's unique leakage pattern. The AESENC instruction causes significant timing variations and L1 data cache misses (76%) due to its S-box table lookups. The AESKEYGENASSIST instruction shows the highest latency, stemming from its iterative key expansion process. Nearby branch instructions (like JNE) worsen the leakage by introducing frequent mispredictions. A key finding is that our gem5 simulation environment reliably mirrors real hardware behaviour, achieving 89% key recovery accuracy compared to 96% on a physical Intel CPU. This validation confirms that simulations can effectively capture these subtle vulnerabilities without real-world noise. We also show that smaller caches amplify leakage by concentrating memory accesses, while larger caches only reduce, not eliminate, the risk. Ultimately, our findings provide actionable insights for developers to harden cryptographic code and for designers creating future detection tools. This work demonstrates the critical role of architectural simulation in identifying and understanding security vulnerabilities before chips are even built. (10.1109/ComComAp68359.2025.11353163)
  DOI : 10.1109/ComComAp68359.2025.11353163
• Domain Adaptation in the era of Foundation Models
  
  • Benigmim Mohammed-Yasser
  , 2025. Artificial intelligence models that analyze images need to understand what each pixel represents, such as identifying cars and pedestrians in street scenes. Building such models for semantic segmentation typically requires collecting many manually labeled images, a time-consuming and expensive process. To address this bottleneck, domain adaptation has emerged: models are trained on accessible source data (such as synthetic images) and adapted to realworld target domains. However, this approach faces domain shift, where models trained on one data distribution fail to generalize to another at deployment. This challenge is especially pronounced in urban driving due to variations in weather, lighting, and geography. This thesis explores how Foundation Models, large AI systems pre-trained on millions of images, can overcome domain shift. These models learn powerful general-purpose representations that can be adapted to diverse applications. We investigate four practical scenarios under severe resource constraints. First, in one-shot domain adaptation, we generate synthetic datasets from a single target image by personalizing text-to-image diffusion models to capture target style while introducing diverse scenes. Second, we develop a collaborative framework where multiple foundation models work synergistically to generalize to unseen environments. Third, we distill knowledge from proprietary black-box AI services accessed via APIs without access to model internals. Finally, we improve open-vocabulary semantic segmentation by identifying which text descriptions work best for each object type. This work enables effective computer vision with unlabeled data and limited resources, democratizing access to advanced AI capabilities. (10.70675/cac8b35fzd07dz45b0z97c1z12f35ecc450d)
  DOI : 10.70675/cac8b35fzd07dz45b0z97c1z12f35ecc450d
• Kramers-Kronig detection in the quantum regime
  
  • Pousset Thomas
  • Federico Maxime
  • Alléaume Romain
  • Fabre Nicolas
  Physical Review Research, American Physical Society, 2025, 7 (4), pp.043287. We investigate the quantization of the Kramers-Kronig (KK) detection technique, initially developed for classical coherent communications. This detection method involves combining the state of interest with a local oscillator on an unbalanced beamsplitter, followed by direct detection and digital signal processing. Intensity measurements in the case of spectrally engineered fields allow for the "digital measurement" of the phase of classical and quantum fields. We show that, to the first order of the local oscillator's amplitude, KK detection is a Gaussian measurement that allows for the estimation of both quadratures of the electric field, similar to double homodyne detection. We study in detail how KK detection operates in the case of bosonic coherent states, pure single-mode and mixed states, as well as the nature of the phase information it measures. Finally, we propose an alternative spectral tomography technique for single-photon states inspired by KK detection. (10.48550/arXiv.2407.20827)
  DOI : 10.48550/arXiv.2407.20827
• Deep learning and SAR tomography for monitoring forest structures
  
  • Berenger Zoé
  , 2025. Synthetic Aperture Radar (SAR) images are widely used for Earth observation because they are not affected by clouds or variations in sunlight. In forests, L- and P-band wavelengths can penetrate the canopy, providing information on both trees and ground. However, radar measurements are degraded by speckle and by the mixing of signals from different heights, making interpretation challenging. Tomographic SAR (TomoSAR) addresses this by combining several images from different baselines to reconstruct vertical reflectivity profiles. These profiles reveal the distribution of scatterers in height, which is key for forest monitoring and biomass estimation (Chapter 1). Traditional TomoSAR methods face clear limitations, especially in forests, where reflectivity is volumetric and not sparse. Non-parametric estimators such as beamforming and Capon are simple and fast but yield coarse results. Regularized inversion methods, including wavelet-based compressive sensing, offer sharper reconstructions but are computationally expensive and sensitive to parameter tuning. The ESA BIOMASS mission, which will provide global P-band acquisitions, makes it urgent to design inversion techniques that are both accurate and scalable (Chapter 2).This thesis investigates whether deep learning can provide such a solution. Neural networks have shown promise in urban TomoSAR, but their use in forests has remained limited due to the lack of ground truth and the complexity of volumetric scattering (Chapter 3). Our first contribution is a supervised deep learning framework for tomographic reconstruction, presented in Chapter 4. Because no ground-truth vertical profiles exist for forests, a generative model is designed to simulate realistic training data. An autoencoder is first tested to learn latent representations of reflectivity profiles, and then an encoder-decoder network is trained to map beamforming reconstructions to high-resolution vertical profiles. The method is thoroughly evaluated on simulations and on three airborne campaigns: BioSAR-2 (boreal forest), DLR data on a temperate forest in Traunstein, and TropiSAR (tropical forest). Across these diverse conditions, the neural network significantly improves the separation of crown tree and ground compared to classical approaches. It shows good agreement with LiDAR data, while requiring much less computation than iterative optimization methods. Additional analyses explore robustness to data correlation, sensitivity to the number of acquisitions, and generalization to varying baselines and polarizations.To overcome the reliance on simulated training data, a second method is developed in Chapter 5: a self-supervised learning approach based on Equivariant Imaging. A novel loss function is introduced, combining data fidelity with regularization, so that the network learns directly from radar measurements without external labels. This allows training on both simulated and airborne data, bypassing the lack of ground truth. The framework is again applied to BioSAR-2, Traunstein, and TropiSAR datasets, where it reconstructs detailed vertical structures, often matching the performance of the supervised method and exceeding that of conventional techniques. Although training is computationally heavier, inference remains efficient, showing strong potential for adapting the method to the BIOMASS mission. These results establish deep learning as both accurate and efficient for large-scale TomoSAR applications over forests.Chapter 6 outlines perspectives at both methodological and application levels. Future work includes extending networks to better-resolved inputs including spatial information, polarimetric TomoSAR and improving generalization across acquisition geometries. Application perspectives focus on adapting these approaches to spaceborne data, particularly BIOMASS data, and directly retrieving biophysical parameters such as tree height and biomass proxies, for operational forest monitoring worldwide. (10.70675/2bd334cdz37aez4826z84a3z9fd469ca7da7)
  DOI : 10.70675/2bd334cdz37aez4826z84a3z9fd469ca7da7
• Deep reinforcement learning for joint optimization of scheduling and resource allocation in mobile ad hoc networks
  
  • Nérondat Sylvain
  , 2025. The central theme of the thesis concerns the use of deep reinforcement learning techniques to solve the joint scheduling and resource allocation problem in mobile ad hoc communication systems. The context of this work is clustered ad hoc networks where nodes are grouped in clusters whose composition can vary dynamically over time. Within each cluster, the scheduling and resource allocation task is performed by a dedicated node that receives the communication requests. The problem we propose to solve is to find the optimal joint scheduling and resource allocation under certain quality of service (QoS) constraints. The global optimization problem has a high dimensional state and action space and cannot be solved by analytical methods, hence the use of new reinforcement learning techniques using deep neural networks. The general objective of the thesis is to study new solutions and their performance for the implementation of joint scheduling and resource allocation algorithms at the radio access level in the cross-optimization paradigm applied to clustered mobile ad hoc networks. (10.70675/1154e2f3za405z4212za8a3z740ceea05854)
  DOI : 10.70675/1154e2f3za405z4212za8a3z740ceea05854
• Mission aware cyber‐physical security
  
  • Bakirtzis Georgios
  • Carter Bryan
  • Fleming Cody
  • Elks Carl
  Systems Engineering, Wiley, 2025, 29 (2), pp.355-367. Perimeter cybersecurity, while essential, has proven insufficient against sophisticated, coordinated, and cyber‐physical attacks. In contrast, mission‐centric cybersecurity emphasizes finding evidence of attack impact on mission success, allowing for targeted resource allocation to mitigate vulnerabilities and protect critical assets. Mission Aware is a systems‐theoretic cybersecurity analysis that identifies components that, if compromised, destabilize the overall mission. It generates evidence by finding potential attack vectors relevant to mission‐linked elements and traces this evidence to mission requirements, prioritizing high‐impact vulnerabilities relative to mission objectives. Mission Aware is an informational tool for system resilience by unifying cybersecurity analysis with core systems engineering goals. (10.1002/sys.70018)
  DOI : 10.1002/sys.70018
• Automatically Interpreting LLM Judgments Using Linguistic Insights : Case of Public Speaking
  
  • Barkar Alisa
  , 2025. We live in a world where generative AI is employed in over 30% of daily tasks. Indeed, large language models (LLMs) offer new opportunities for automatic evaluation that previously relied on hand-crafted features. In public speaking assessment, for instance, LLMs can apply criteria formulated by experts directly in natural language, generating both judgments and explanations. This shift underpins the growing use of LLMs in commercial applications such as Poised, PolymnIA, Yoodli, and others. Yet concerns about explainability and accuracy remain among the most pressing issues. These concerns are especially acute in subjective domains, where defining objective accuracy is inherently challenging. In such contexts, human-like feedback from large language models (LLMs) can be misleading, particularly for users lacking the expertise to detect subtle inaccuracies.This thesis investigates the strengths and weaknesses of LLMs in the case of automatic public speaking assessment. We develop a criterion-based framework grounded in public speaking theory and expert feedback, and construct a new French dataset with dense expert annotations. Using this resource, we introduce a three-layer evaluation methodology combining bias analysis, model self-consistency, human-model agreement, and linguistic interpretation. Beyond statistical comparison, we probe how LLMs explain their evaluations and manipulate persuasiveness, revealing the rhetorical strategies encoded in their behaviour.The contributions of this thesis are threefold. First, we provide a novel text-based framework and dataset for criterion-based public speaking assessment, a resource that can support future work on expert versus lay judgments and alternative evaluation methods. Second, we deliver the first systematic evaluation of flagship LLMs in this domain, showing both their capacity to reproduce certain expert preferences and their limitations, such as overreliance on emotional manipulation and simplification. These findings inform not only future LLM development but also ethical debates around their use in communication training. Third, we propose a linguistically interpretable methodology and an extended set of linguistic features for public speaking evaluation, enabling analyses that go beyond accuracy to uncover the cues shaping human versus model judgments. This approach can be applied to other models to test whether similar evaluative patterns emerge. Our results show that LLMs, while self-consistent, are strongly biased toward positive evaluations and exhibit low agreement with the expert. Whereas the expert prioritised high-level organisation and content framing, LLMs leaned on lexical accessibility, transition markers, and emotional vocabulary. These tendencies persist even when models are guided by expert examples, raising concerns about their reliability as evaluators.Taken together, this thesis demonstrates that LLMs do not mirror expert evaluators but instead rely on a distinct persuasion model. This insight has implications for research, providing a robust foundation for future studies, and for practice, where it cautions against uncritical reliance on LLMs in socially consequential evaluation tasks. (10.70675/4a56b558z1a62z4cbdz9824zf44dd525e2f9)
  DOI : 10.70675/4a56b558z1a62z4cbdz9824zf44dd525e2f9
• On the Role and Robustness of Self-Supervised and Multiple Instance Learning Approaches for Digital Pathology - Application to Sjögren's Syndrome
  
  • Mammadov Ali
  , 2025. Digital pathology has revolutionized medical diagnosis by enabling the digitization of tissue samples into Whole Slide Images (WSIs). However, WSIs are extremely large and high-resolution, which makes applying deep learning models directly very challenging. A common solution is to divide each slide into smaller patches, treating each slide as a bag of patches with a single global label rather than requiring costly pixel-level annotations. This approach naturally leads to Multiple Instance Learning (MIL), which has become the preferred method for WSI classification. MIL has two main paradigms: instance-based and embedding-based. In instance-based MIL, each patch is classified independently, and the patch-level predictions are aggregated to determine the slide label. In embedding-based MIL, patch embeddings are first combined into a single representation before classifying the slide. While instance-based methods are more interpretable, embedding-based methods have traditionally been favored due to their robustness with weak feature extractors. However, recent advances in Self-Supervised Learning (SSL) have dramatically improved feature quality. Despite this, many studies still prioritize embedding-based MIL. To investigate this, we conducted 710 experiments across four datasets, comparing 12 MIL strategies, six SSL methods with four backbones, four foundation models, and several pathology-specific techniques. We also introduce four novel instance-based MIL methods not previously used in pathology. Our results show that, with high-quality SSL features, simple instance-based MIL methods with few parameters match or outperform state-of-the-art embedding-based MIL methods. These findings suggest that efforts should focus on designing effective SSL feature extractors rather than developing increasingly complex embedding-based MIL models, especially since instance-based methods are naturally more interpretable for clinicians. A second challenge in MIL is high run-to-run performance variability, which can reach 10-15 AUC points due to random weight initialization, batch ordering, and learning rate differences. To address this, we propose a Multi-Fidelity Model Fusion strategy: multiple models are trained for a few epochs, and the most promising ones are selected and ensembled based on validation scores. This approach reduces variability, improves reproducibility, simplifies hyperparameter tuning, and maintains computational efficiency. We validated it extensively with over 2,000 experiments on multiple datasets, initialization strategies, and MIL models, confirming its effectiveness. Finally, we apply these techniques to Sjögren's syndrome (SjS), a chronic autoimmune disease affecting salivary and lacrimal glands. The clinical gold standard involves quantifying lymphocytic foci in labial salivary gland biopsies, but manual scoring is time-consuming and prone to inter-observer variability. While SSL-enhanced MIL can achieve high performance, it frames the problem as a classification task and is sometimes not sufficiently interpretable for clinical use in cases like SjS diagnosis, which requires step-by-step reasoning. To address this, we propose a fully automated, clinically interpretable pipeline that replicates the pathologist's workflow: candidate foci are detected using unsupervised clustering, filtered with a supervised tissue classifier, and aggregated to compute a focus score. Using a curated dataset from Saint-Joseph Hospital, our method achieves strong diagnostic performance while maintaining transparency and clinical relevance. (10.70675/500eff04z8e2bz47a8zab97zb9ea5105a432)
  DOI : 10.70675/500eff04z8e2bz47a8zab97zb9ea5105a432