Publications

2023

• Tubular structures segmentation of pediatric abdominal-visceral ceCT images with renal tumors: assessment, comparison and improvement
  
  • La Barbera Giammarco
  • Rouet Laurence
  • Boussaid Haithem
  • Lubet Alexis
  • Kassir Rania
  • Sarnacki Sabine
  • Gori Pietro
  • Bloch Isabelle
  Medical Image Analysis, Elsevier, 2023. Renal tubular structures, such as ureters, arteries and veins, are very important for building a complete digital 3D anatomical model of a patient. However, they can be challenging to segment from ceCT images due to their elongated shape, diameter variation and intra-and inter-patient contrast hetereogenity. This task is even more difficult in pediatric and pathological subjects, due to high inter-subject anatomical variations, potential presence of tumors, small volume of these structures compared to the surrounding, and small available labeled datasets. Given the limited literature on methods dedicated to children, and in order to find inspirational approaches, a complete assessment of state-of-the-art methods for the segmentation of renal tubular structures on ceCT images on adults is presented. Then, these methods are tested and compared on a private pediatric and pathological dataset of 79 abdominal-visceral ceCT images with arteriovenous phase acquisitions. To the best of our knowledge, both assessment and comparison in this specific case are novel. Eventually, we also propose a new loss function which leverages for the first time the use of vesselness functions on the predicted segmentation. We show that the combination of this loss function with state-ofthe-art methods improves the topological coherence of the segmentated tubular structures 1 .
• Early Validation of Functional Requirements
  
  • Assioua Yasmine
  • Ameur-Boulifa Rabea
  • Pacalet Renaud
  • Guitton-Ouhamou Patricia
  Africa Insight, Africa Institute of South Africa, 2023, 26 (4), pp.30-32. Technical specifications and intended functionalities are often gathered in documents that include requirements written in constrained natural language, that is, natural‐like language with restricted syntax. In the automotive industry one challenge is the ability to produce safe vehicles, emphasizing the importance of safety by design. In the framework of case studies based on functions of autonomous vehicles, we introduce a systematic process for building formal models from automotive requirements written in constrained natural language, and for verifying them. By allowing formal verification at the earliest stages of the development cycle our aim is to avoid the costly discovery of errors at later stages. (10.1002/inst.12467)
  DOI : 10.1002/inst.12467
• Introducing the 3MT_French Dataset to Investigate the Timing of Public Speaking Judgements
  
  • Biancardi Beatrice
  • Chollet Mathieu
  • Clavel Chloé
  , 2022. Abstract In most public speaking datasets, judgements are given after watching the entire performance, or on thin slices randomly selected from the presentations, without focusing on the temporal location of these slices. This does not allow to investigate how people's judgements develop over time during presentations. This contrasts with primacy and recency theories, which suggest that some moments of the speech could be more salient than others and contribute disproportionately to the perception of the speaker's performance.To provide novel insights on this phenomenon, we present the 3MT_French dataset. It contains a set of public speaking annotations collected on a crowd-sourcing platform through a novel annotation scheme and protocol. Global evaluation, persuasiveness, perceived self-confidence of the speaker and audience engagement were annotated on different time windows (i.e., the beginning, middle or end of the presentation, or the full video). This new resource will be useful to researchers working on public speaking assessment and training. It will allow to fine-tune the analysis of presentations under a novel perspective relying on socio-cognitive theories rarely studied before in this context, such as first impressions and primacy and recency theories. An exploratory correlation analysis on the annotations provided in the dataset suggests that the early moments of a presentation have a stronger impact on the judgements. (10.21203/rs.3.rs-2122814/v1)
  DOI : 10.21203/rs.3.rs-2122814/v1
• Blockchain Adoption in Healthcare : Toward a Patient Centric Ecosystem
  
  • Azzi Rita
  , 2023. The healthcare sector evolves constantly, driven by technological advancement and innovative solutions. From remote patient monitoring to the Internet of Things (IoT), Artificial Intelligence (AI), personalized medicine, mobile health, and electronic records systems, technology has improved patient outcomes and enhanced care delivery. These technologies have shifted the healthcare ecosystem to be more patient-centered, focusing on meeting the patient's needs rather than the needs of the individual organizations within it. However, this transformative shift experienced by the healthcare industry is associated with multiple challenges due to the inherent complexity and fragmentation of the healthcare ecosystem. This dissertation addresses three healthcare ecosystem challenges that significantly impact patients. The first challenge addressed is the problem of counterfeit or falsified drugs that represent a threat to public health, resulting from the vulnerabilities in the pharmaceutical supply chain, notably centralized data management and the lack of transparency. The second challenge addressed is the problem of healthcare data fragmentation that thwarts care coordination and impacts clinical efficiency. This problem results from the dynamic and complex patients' journey in the healthcare system, shaped by their unique health needs and preferences. Patient data are scattered across multiple healthcare organizations within centralized databases and are ruled by policies that hinder data sharing and patients' empowerment over their data. The third challenge addressed is the confidentiality and privacy of healthcare data that, if compromised, shatter the trust relationship between patients and healthcare stakeholders. This challenge results from the healthcare organizations' poor data governance that increases the risk of data breaches and unauthorized access to patient information.The blockchain has emerged as a promising solution to address these critical challenges. It was introduced into the healthcare ecosystem with the promise of enforcing transparency, authentication, security, and trustworthiness. Through comprehensive analysis and case studies, this dissertation assesses the opportunities and addresses the challenges of adopting the blockchain in the healthcare industry. We start with a thorough review of the state of the art covering the blockchain's role in improving supply chain management and enhancing the healthcare delivery chain. Second, we combine theoretical and real-world application studies to develop a guideline that outlines the requirements for building a blockchain-based supply chain. Third, we propose a patient-centric framework that combines blockchain technology with Semantic technologies to help patients manage their health data. Our fourth contribution presents a novel approach to data governance by developing a blockchain-based framework that improves data security and empowers patients to participate actively in their healthcare decisions. In this final contribution, we widen the scope of the proposed framework to include a roadmap for its adoption across diverse domains (banking, education, transportation, and logistics, etc.).
• Collateral-Free Learning of Deep Representations : From Natural Images to Biomedical Applications
  
  • Barbano Carlo Alberto Maria
  , 2023. Deep Learning (DL) has become one of the predominant tools for solving a variety of tasks, often with superior performance compared to previous state-of-the-art methods. DL models are often able to learn meaningful and abstract representations of the underlying data. However, it has been shown that they might also learn additional features, which are not necessarily relevant or required for the desired task. This could pose a number of issues, as this additional information can contain bias, noise, or sensitive information, that should not be taken into account (e.g. gender, race, age, etc.) by the model. We refer to this information as collateral. The presence of collateral information translates into practical issues when deploying DL-based pipelines, especially if they involve private users' data. Learning robust representations that are free of collateral information can be highly relevant for a variety of fields and applications, like medical applications and decision support systems.In this thesis, we introduce the concept of Collateral Learning, which refers to all those instances in which a model learns more information than intended. The aim of Collateral Learning is to bridge the gap between different fields in DL, such as robustness, debiasing, generalization in medical imaging, and privacy preservation. We propose different methods for achieving robust representations free of collateral information. Some of our contributions are based on regularization techniques, while others are represented by novel loss functions.In the first part of the thesis, we lay the foundations of our work, by developing techniques for robust representation learning on natural images. We focus on one of the most important instances of Collateral Learning, namely biased data. Specifically, we focus on Contrastive Learning (CL), and we propose a unified metric learning framework that allows us to both easily analyze existing loss functions, and derive novel ones. Here, we propose a novel supervised contrastive loss function, ε-SupInfoNCE, and two debiasing regularization techniques, EnD and FairKL, that achieve state-of-the-art performance on a number of standard vision classification and debiasing benchmarks.In the second part of the thesis, we focus on Collateral Learning in medical imaging, specifically on neuroimaging and chest X-ray images. For neuroimaging, we present a novel contrastive learning approach for brain age estimation. Our approach achieves state-of-the-art results on the OpenBHB dataset for age regression and shows increased robustness to the site effect. We also leverage this method to detect unhealthy brain aging patterns, showing promising results in the classification of brain conditions such as Mild Cognitive Impairment (MCI) and Alzheimer's Disease (AD). For chest X-ray images (CXR), we will target Covid-19 classification, showing how Collateral Learning can effectively hinder the reliability of such models. To tackle such issue, we propose a transfer learning approach that, combined with our regularization techniques, shows promising results on an original multi-site CXRs dataset.Finally, we provide some hints about Collateral Learning and privacy preservation in DL models. We show that some of our proposed methods can be effective in preventing certain information from being learned by the model, thus avoiding potential data leakage.
• Machine Learning for beam Alignment in mmWave massive MIMO
  
  • Ktari Mohamed Aymen
  , 2023. The escalating demand for spectral efficiency driven by the stringent requirements of 5G networks has spurred the development of mmWave MIMO technology, promising significant architectural improvements through advanced precoding techniques. This technology presents substantial gains in spectral and energy efficiencies compared to traditional MIMO systems. However, the transformative potential of mmWave MIMO is hampered by the complex realities of real-world urban environments and the intricate physical properties inherent to mmWave frequencies.Crucially, in mmWave massive MIMO communication, beamforming and combining play pivotal roles: the high bandwidth and operating frequency of mmWave systems necessitate analog domain beamforming/combining, rendering fully digital approaches technically non feasible. At the heart of mmWave large-dimensional MIMO lies the Beam Alignment problem, requiring the identification of optimal transmit and receiver beam pairs that maximize the Signal-to-Noise ratio, ensuring a robust initial link.Existing standards, such as WiGig, employ exhaustive beam sounding methods, testing each possible beam pair to find the one maximizing SNR. Consequently, it leads to substantial pilot-signaling overhead, the major problem we aim to encounter throughout this PhD. Our research revolutionizes Beam Alignment by integrating cutting-edge machine learning techniques for Partial Beam Alignment, significantly reducing the pilot overhead by soundings a subset of beam pairs using sub-sampled codebooks. Therefore, we leverage the received signal energies from these beam pairs soundings, employing shallow neural networks, matrix factorization, and their variants for accurately resolving non-linear and logistic regression problems, crucial for determining the quality of the remaining beam pairs.A fundamental objective of this thesis is to determine the sample complexity for these machine learning methods. This complexity dictates the minimum number of training samples necessary for effective learning and reliable transmission. We delve into the performance of the proposed ML models without prior channel estimation, introducing the concept of Blind Beam Alignment, thus pioneering a paradigm shift. Furthermore, our research delves deep into the nuances of quantization, a vital practical constraint. We then explore critical compromises: identifying the minimum overhead ratio corresponding to the optimal quantization scheme on the one hand and navigating the classic trade-off between accuracy and complexity on the other hand.Through systematic progression, ranging from basic point-to-point narrowband scenarios to intricate wideband multi-user architectures, this PhD thesis offers valuable insights and solutions. The proposed contributions advance the fields of mmWave communications and Machine Learning applications in wireless systems, outperforming existing benchmarks, and encountering the limitations of conventional approaches.
• Physically constrained generative networks for cloud and texture synthesis
  
  • Chatillon Pierrick
  , 2023. Evaluating the performance of optical sensors requires large-scale databases of cloud backgrounds, for example to predict optical link availability between ground stations and satellites and detecting small objects like drones against cloudy skies.Deep learning algorithms can be implemented for these purposes, requiring large training databases.Accessing such databases is challenging since passive systems used for atmospheric observations provide only partial views, and databases constructed through physical modeling are costly.We have therefore developed deep learning methods to overcome this need for large quantities of cloud images, with the aim of maintaining the spectral and radiometric properties of the images.Physical simulations are limited in terms of spatial resolution if the area to be covered is large. Hence, we explored two super-resolution approaches to enhance image definition. Both methods belong to internal methods, exploiting information redundancy within a single image at different locations and scales. They leverage the fractal properties of cloud backgrounds and use a generative network as a common model for various resolutions. These methods enable the generation of images exhibiting a power decay of the spectral density, an essential descriptor of cloud textures.A different direction of our research involves exploring texture synthesis methods. We introduce a generative model for cloud image generation based on physical parameters. This model can control the spectral behavior and histogram characteristics of the generated images, given set of physical descriptors. It utilizes an appropriate multi-scale noise weighting to govern the spectral slope. Finally, we delved into texture synthesis from a general perspective, proposing an auto-encoder structure adapted to textures and enriched to handle textures with periodic patterns.Overall, our work contributes to generating realistic cloud images from limited data, preserving spectral and radiometric properties, thanks to multi-scale approaches that leverage the fractal characteristics of clouds.
• Automatic analysis of trust over the course of a human-robot interaction using multimodal features and recurrent neural architectures
  
  • Hulcelle Marc
  , 2023. Trust is an important psychological construct in HRI as it mitagates the relationship qualities between partners of an interaction, as well as the performance of the interaction's task. Research on trust were essentially organized around the study of socio-psychological effects of the robot's design and behavior on users. Trust is usually measured through questionnaires filled by users themselves at the beginning and end of the interaction. In this thesis, we tackle the issue of automatic analysis of trust dynamics during the course of interaction. The standard Psychological approaches used in HRI to study, coming from a mentalist perspective, do not currently allow such analysis. We thus leverage Interactionist Sociology theories to create a coding scheme named TURIN (Trust in hUman Robot INteraction) dedicated to this task. From there, we use Machine Learning tools to develop multimodal models of trust. We propose a new methodology that allows to conduct the analysis over the course of the interaction, first through simple models, then by the design of a specific recurrent neural architecture. We finish by an analysis of ours models to determine which behaviors are the most indicative of trust and understand the types of errors thatthey make.
• Contributions to stochastic analysis for non-diffusive structures
  
  • Vuong Christophe
  , 2023. This thesis is concerned with the study of non-diffusive structures. We focus on two classes of such structures.The first subject deals with Malliavin calculus for conditionally independent random variables, which is a special case of discrete Malliavin calculus. It also generalizes the calculus that has been developed for countable products of probability spaces, for independent random variables.In our case, the interest of such a calculus is to complement results in stochastic analysis with proofs of functional inequalities (Poincaré inequality, McDiarmid's inequality) and limit theorems. One of the main applications is the determination of the convergence rate of central limit theorems via the Stein method.By combining Malliavin calculus with the underlying Dirichlet structure of the random variables, we obtain an integration by parts formula which is key to the derivations of so-called Stein bounds of the rates of convergence. We show quantitative limit theorems, including a fourth moment theorem with remainder. In particular, we discuss an application to the asymptotic normality of motif counting in exchangeable random hypergraphs.The second subject studies functionals of a Poisson measure using the notion of invertibility of transformations of that measure on the sample space of random measures. We use the identification of these measures and the associated marked point processes. Invertible transformations are obtained via the Girsanov's theorem, respecting absolute continuity with respect to the reference measure. This results in an entropy criterion for the invertibility of transformations. Finally, we make the connection with stochastic differential equations driven by Poisson measures.
• Frequency-domain quantum information processing with multimode quantum states of light from integrated sources at telecom wavelengths
  
  • Henry Antoine
  , 2023. In quantum information, encoding in time and frequency degrees of freedom gives access to a high-dimensional Hilbert space for photonic states, enabling parallel processing of a large number of qubits or even qudits. This is the scope of our work on the generation and manipulation of photonic quantum states at telecom wavelengths with three main achievements. The first one is the efficient generation of photon pairs by second and third-order nonlinear processes in innovative integrated sources: a thin-film, periodically-poled lithium niobate-on-insulator waveguide, and a silicon-on-insulator micro-resonator with a free spectral range of 21 GHz. The second one is the development of concepts, models, and numerical optimizations for the manipulation of photonic qubits and qudits in time-frequency spaces with linear devices. We use programmable filters (PF) and electro-optical phase modulators (EOM). We compare the theoretical performance of 1-qubit gates for two configurations [EOM-PF-EOM] and [PF-EOM-PF] in both time and frequency encoding. The third one is the experimental demonstration of such manipulation of frequency qubits from the silicon microresonator. We use the [EOM-PF-EOM] configuration to implement a reconfigurable and tunable quantum gate. A single tunable parameter is used to go from an identity gate to a Hadamard gate, as well as to a continuum of intermediate gates. We then use these gates to perform quantum tomography of entangled states and to implement a quantum key distribution protocol based on two-photon frequency entanglement. Finally, we demonstrate a frequency-encoded multi-user network without trusted nodes. This experiment constitutes a proof of principle for quantum key distribution in the frequency domain at a rate of 2 bits per second simultaneously for each pair of users in a 5-user network.
• Assessing the Threat Level of Software Supply Chains with the Log Model
  
  • Soeiro Luı́s
  • Robert Thomas
  • Zacchiroli Stefano
  , 2023. The use of free and open source software (FOSS) components in all software systems is estimated to be above 90%. With such high usage and because of the heterogeneity of FOSS tools, repositories, developers and ecosystem, the level of complexity of managing software development has also increased. This has amplified both the attack surface for malicious actors and the difficulty of making sure that the software products are free from threats. The rise of security incidents involving high profile attacks is evidence that there is still much to be done to safeguard software products and the FOSS supply chain. Software Composition Analysis (SCA) tools and the study of attack trees help with improving security. However, they still lack the ability to comprehensively address how interactions within the software supply chain may impact security. This work presents a novel approach of assessing threat levels in FOSS supply chains with the log model. This model provides information capture and threat propagation analysis that not only account for security risks that may be caused by attacks and the usage of vulnerable software, but also how they interact with the other elements to affect the threat level for any element in the model.
• Choosing the Right Time to Learn Evolving Data Streams
  
  • Bernardo Alessio
  • Valle Emanuele Della
  • Bifet Albert
  , 2023, pp.5156--5165. Continuous data generation over time presents new challenges for Machine Learning systems, which must develop real-time models due to memory and latency limitations. Streaming Machine Learning algorithms analyze data streams one sample at a time, progressively updating their models. However, is it necessary to utilize all the data for model updates? This paper introduces the Online Ensemble SPaced Learning (OE-SPL) strategy, an ensemble meta-strategy that combines online ensemble learning and the Spaced Learning heuristic to rapidly learn underlying concepts without using all samples. We evaluated OE-SPL on synthetic and real data streams containing various concept drifts, providing statistical evidence that OE-SPL achieves comparable performance to state-of-the-art ensemble models while recovering from multiple concept drift occurrences more efficiently, using less time and RAM-Hours. (10.1109/BIGDATA59044.2023.10386551)
  DOI : 10.1109/BIGDATA59044.2023.10386551
• Deep learning for remote sensing images and their interpretation
  
  • Meraoumia Ines
  , 2023. Synthetic Aperture Radar (SAR) images are not affected by the presence of clouds or variations of sunlight. They provide very useful information for Earth observation (chapter 1).They are impacted by strong fluctuations called "speckle" which make their interpretation difficult. The speckle is a phenomenon intrinsic to the coherent illumination of the scene by the radar, meaning that speckle-free images can not be captured and used as reference to train models.The properties of speckle are different from that of the traditional additive white Gaussian noise used to model corruptions in optical images, and proper despeckling algorithms are needed. Most of them rely on statistics derived from the Goodman's model (chapter 2). Recently, deep learning based methods have been very successful at despeckling a single SAR image. This work focuses on improving the despeckling performance by jointly processing several input images to exploit the common information while still preventing the propagation of differences from one image to another (chapter 3).The despeckling of Sentinel-1 GRDM Extra Wide images of sea ice is studied in Chapter 4 for sea ice studies. The ice is shifting quickly on the sea and multi-temporal stacks of a specific area can not be used for despeckling purposes due to structural changes. In the images, thermal noise can not be neglected because the reflectivity values of water and ice are very low and close to the thermal noise floor. We propose a dual-polarimetric despeckling framework where HH and HV polarimetric channels are used as input and are jointly despeckled in a single pass. The network is trained in a self-supervised way inspired by the existing SAR2SAR framework and takes corrected images where the thermal noise floor level has been removed as input. Our approach shows a clear improvement over existing image restoration techniques on Sentinel-1 images of the Artic.Despeckling can be improved by combining measurements pertaining to common information within the temporal stack while ignoring data impacted by temporal changes.First, multi-temporal despeckling methods using temporal averaging and the computation of a super-image (i.e. despeckled temporal mean image) are introduced at the beginning of Chapter 5. A generative model is then proposed to explicit the statistics of SAR multi-temporal stacks and account for the spatial and temporal correlations of speckle. A multi-temporal extension of the existing MERLIN framework is derived from this model. The network is fed with additional images of the same area acquired at different dates. It is trained in an unsupervised way inspired by the Noise2Noise framework: the real part (or the imaginary part) of the image and additional dates are fed to the network and the imaginary part (or the real part) is used as a target. Adding more images continuously improves the despeckling performance, but with diminishing gains.A temporal whitening is proposed to prevent the drop of performance of the network when the input channels are temporally correlated.Despeckling methods are hard to evaluate because of the lack of ground truth images. Chapter 6 focuses on uncertainty quantification for despeckling using deep learning.First, works are presented to combine despeckling and estimation of the uncertainty map during the training. Starting from a framework where only one value is predicted for each pixel, we aim at predicting a distribution for each pixel. Parameters of uniform and inverse gamma distributions are estimated. The sharper the distribution, the more certain the network is of its prediction. We discuss the difficulty of estimating uncertainties in a self-supervised learning framework where the noise level is high and the limits faced by our formulations.Working with the MERLIN framework, an estimation of an uncertainty map is proposed based on the expected difference map between predictions from the real and imaginary parts.
• Automatic analysis of image quality criteria in natural scenes using deep neural networks
  
  • Tworski Marcelin
  , 2023. As smartphone cameras became more prevalent than traditional camera systems, the demand for precise measurements increased. This Ph.D. dissertation proposes using deep learning systems to evaluate image-quality criteria specific to smartphone camera evaluation, and more specifically texture evaluation. This dissertation addresses several limitations in current image-quality assessment methods for smartphone cameras. Deep learning systems struggle with computational complexity due to high-resolution smartphone images, and downsizing would lead to information loss for evaluating noise and details preservation. Consequently, it is essential to find the relevant image regions to assess a camera attribute to alleviate these problems. Additionally, the lack of suitable datasets hinders the development of learning-based methods aimed at benchmarking smartphone cameras.Furthermore, when comparing cameras, it is essential to capture the same content to facilitate direct comparison. In standard camera benchmarking protocols, multiple shots are collected from the same content. This setting deviates from traditional machine learning approaches, where training and test data are assumed to be independent and identically distributed (iid). However, the non-independent nature of our data is frequently overlooked in the image-quality assessment literature.To overcome these challenges, this research introduces several contributions: (i). A region selection method is introduced to automatically detect relevant regions for evaluating specific quality attributes. Adapting the class activation map method for a regression problem, we outperform traditional chart-based approaches in evaluating texture quality and permitting the usage of deep learning methods on charts shot in laboratory conditions. In this work, we use texture quality as an illustrative example of camera quality attributes. However, our methodology is designed to be applicable to other attributes, such as noise, as well. (ii) A new in-the-wild dataset is created to accurately reflect the complex mixture of defects commonly found in smartphone camera images and reflect the scenario of camera benchmarking, where several different scenes are shot by multiple camera devices. This dataset, annotated through pairwise comparisons, allows us to perform a large evaluation of different methods in different practical scenarios, setting guidelines for the usage of deep learning systems for camera quality evaluation. (iii) We introduce a new image quality assessment setup and method where we go beyond the traditional iid assumption. We consider multiple images with varying quality of the same content available at test time. We use the specificity of this camera quality estimation setting to enhance the quality prediction accuracy by introducing a batch-based pseudo-reference which allows us to use full-reference methods in the no-reference setting.
• High-Order Collision Attack Vulnerabilities in Montgomery Ladder Implementations of RSA
  
  • Varillon Arnaud
  • Sauvage Laurent
  • Danger Jean-Luc
  , 2024, 14412, pp.139-161. This paper describes a straightforward methodology which allows mounting a specific kind of single-trace attacks called collision attacks. We first introduce the methodology (which operates at the algorithmic level) and then provide empirical evidence of its soundness by locating the points of interest involved in all existing collisions and then attacking an unmasked RSA implementation whose modular exponentiation is based on the Montgomery Ladder. The attacks we performed, albeit slightly worse than the theoretical prediction, are very encouraging nonetheless: the whole secret exponent can be retrieved (i.e., a success rate equal to 100%) using only 10 traces. Lastly, we describe how this could allow for the introduction of high-order attacks, which are known to break some protected implementations of symmetric cryptography, in the context of asymmetric cryptography. (10.1007/978-3-031-51583-5_9)
  DOI : 10.1007/978-3-031-51583-5_9
• Methods and frameworks of annotation cost optimization for deep learning algorithms applied to medical imaging
  
  • Ruppli Camille
  , 2023. In recent years, the amount of medical imaging data has kept on growing. In 1980, 30 minutes of acquisition were necessary to obtain 40 medical images.Today, 1000 images can be acquired in 4 seconds. This growth in the amount of data has gone hand in hand with the development of deep learning techniques which need quality labels to be trained. In medical imaging, labels are much more expensive to obtain as they require the expertise of a radiologist whose time is limited. The goal of this thesis is to propose and develop methods to limit the annotation load in medical imaging while maintaining a high performance of deep learning algorithms.In the first part of this thesis, we focus on self-supervised learning methods which introduce pretext tasks of various types: generation based, context based and self-distillation approaches. These tasks are used to pretrain a neural network with no additional annotations to take advantage of the amount of available unannotated data. Most of these tasks use perturbations often quite generic, unrelated to the objective task and sampled at random in a fixed list with fixed parameters. How to best combine and choose these perturbations and their parameters remains unclear. Furthermore, some perturbations can be detrimental to the target supervised task. Some works mitigate this issue by designing pretext tasks for a specific supervised task, especially in medical imaging. But these tasks do not generalize well to other problems.A balance must be found between perturbation or pretext task optimization for a given supervised problem and method generalization ability.Among context-based methods, contrastive learning approaches propose an instance-level discrimination task: the latent space is structured with instance similarity. Defining instance similarity is the main challenge of these approaches and has been widely explored.When defining similarity through perturbed versions of the same image, the same questions of perturbations optimization arise.We introduce a perturbation generator optimized for contrastive pre-training guided by a small amount of supervision.Class labels and metadata have been used to condition instance similarity, but these data can be subject to annotator variability, especially in the medical domain. Some methods have been proposed to use confidence in fully supervised and self-supervised training, but it is mostly based on loss function values. However, confidence on labels and metadata is often linked to a priori domain knowledge such as data acquisition, annotators experience and agreement. This is even more relevant for medical data.In the second part of this thesis, we focus we design an adapted contrastive loss introducing annotation confidence for the specific problem of prostate cancer lesion detection.Finally, we explore some approaches to apply self-supervised and contrastive learning to prostate cancer lesion segmentation.
• Applications of Artificial Intelligence to Control and Analyze the Performance of Fiber-Optic Transmission Systems
  
  • Ye Xiaoyan
  , 2023. The surging demands for internet traffic have necessitated continuous expansion in opticalfiber communication systems capacity, cornerstone of global communication networks. This thesisdelves into innovative solutions addressing the challenges posed by ultra-wideband (UWB) amplificationand precise noise estimation in optical transmission systems. Optical fiber communication systems haveundergone significant evolution to meet escalating capacity requirements. Progressing from optical amplifiersand coherent detection to advanced modulationformat and digital signal processing (DSP) algorithms. To meet the need for higher traffic demands in opticalnetworks, integrating UWB schemes and implementing low-margin network designs have becomeprimordial. This work explores fundamental aspects of UWB amplification. Accurate prediction of Ramangain profiles and optimal pump configurations design is paramount, yet conventional methods prove computationallyintensive. Here, Machine Learning (ML) emerges as a powerful tool, simplifying complexityand enhancing accuracy in these scenarios. Additionally, the thesis addresses the challenge of designinglow-margin systems by developing a reliable Quality of Transmission (QoT) tool. Optical fiber transmissionsystems contend with diverse impairments such as fiber attenuation, ASE noise, laser phase noise, nonlinearinterference (NLI), etc. While linear impairments can be effectively mitigated and characterized, traditionalmethods may falter in estimating some major nonlinear impairments, posing challenges in accuracyand complexity. Consequently, this work delves into data-driven approaches, including ML frameworks,to provide effective estimation of Kerr nonlinear impairments and electronically enhanced phase noise(EEPN) In summary, this thesis leverages ML and data-driven methods to enhance the performance ofoptical transmission systems. These advancements are poised to shape the future of optical communicationsystems, facilitating higher capacities and more reliable transmissions in our rapidly evolving digitalenvironment.
• Optimization of High Data Rate Ground to Satellite Links Pre-compensated by Adaptive Optics
  
  • Lognoné Perrine
  , 2023. In a context of growing digital needs, optical satellite communications serve as a complementary tool to existing terrestrial communication infrastructures. Establishing optical links with GEO-stationary satellites would enable data exchange at rates on the order of Terabits per second between Earth and space. One of the primary limitations this optical link is the disturbance of the optical wave during its propagation through the atmosphere. The effect of atmospheric turbulence results in spatiotemporal fluctuations in the phase and amplitude of this wave. This translates into a highly disturbed beam with speckles evolving over time in the satellite's plane. Consequently, the coupled flux onboard the satellite fluctuates significantly, leading to long and deep signal fades that degrade the information signal.Solutions exist to mitigate these information losses. Physical means, such as adaptive optics, can minimize coupling losses, while digital techniques can enhance information reliability through coding and interleaving. These techniques have been applied to the downlink in previous works (Lucien Canuet). Concerning the uplink, the envisioned optical technique is the pre-compensation by adaptive optics. However, due to the geometry of the link, where the optical up and downlink path are separated by a point-ahead angle, this pre-compensation identical to the downlink AO correction is currently suboptimal. As a result, deep and long signal fades persist.In this thesis, we have designed new methods to optimize the pre-compensation phase at the point-ahead angle, thereby improving the channel statistics. These methods design and evaluation rely on a reciprocal formalism that allows for an analytical description of the pre-compensation phase error and associated coupled flux. To optimize the pre-compensation phase at the point-ahead angle, we have developed four methods that exploit information obtained from available measurements at the optical ground station. All the proposed methods show to greatly reduce the pre-compensation phase error and therefore improve the statistics of the coupled flux aboard the satellite. Additionally, we evaluate the telecommunication performance of the links using the developed pre-compensation methods. Finally, we develop the statistical channel model of the AO pre-compensated link.
• Approches statistiques pour les communications centrées sur l'utilisateur
  
  • Youssef Badre
  , 2023. Les réseaux corporels sans fil (WBAN) sont étudiés depuis plus d'une vingtaine d'année. Les domaines d'applications sont multiples et ce type de communication repose sur des liaisons radio entre plusieurs terminaux à proximité du corps humain. On distingue principalement 3 types de liaisons selon le positionnement des antennes (in-on, on-on et on-off). Leurs spécificités sont d'une part, l'impact du corps humain sur le canal de propagation, c'est un perturbateur électromagnétique puissant qui est l'origine d'atténuations importantes et de mécanismes de propagation particuliers et d'autre part, le nombre de sources de variabilités qui est très important : le sujet, la nature du lien radio, les antennes, la fréquence et l'environnement proche. Tout ceci explique la complexité de ce type de canal et les difficultés à le modéliser de manière générique.A notre connaissance, les travaux menés jusqu'à présent ont permis d'extraire des modèles fondés la plupart du temps sur une approche plutôt classique en distance ayant l'avantage d'avoir une explication physique ou encore une approche fondée sur des scénarios qui permet d'avoir une orientation « applications ». Certaines sources comme l'environnement et la morphologie sont insuffisamment étudiées selon nous, tant qualitativement que quantitativement alors que leur influence est souvent significative et parfois importante selon le lien radio.L'objectif principal de la thèse est donc de proposer des modèles paramétriquo-statistiques pour différents observables, pour ces sources de variabilités, en prenant en compte un échantillon statistique plus important que ce qui a été fait aujourd'hui dans un contexte de liens radio. C'est ainsi que nous avons mis en place une méthodologie associant, la définition d'un plan d'expériences, d'une base de données suffisante issue de simulations et de mesures expérimentales permettant de disposer de données de références. L'approche expérimentale devient rapidement irréaliste lorsque la diversité des situations possibles doit être considérée car la combinatoire est explosive. On souligne que dans les publications fondées sur des mesures, les modèles obtenus sont propres à l'environnement considéré, ce qui n'est pas le but de notre approche.Pour l'environnement, nous avons développé un code de Ray Tracing simplifié permettant l'étude de n'importe quel environnement parallélépipédique vide et type de lien radio. Sachant que la taille des pièces a une influence, notre objectif est de déterminer si quantitativement cet effet est significatif ou de 2nd ordre, et éventuellement sous quelles conditions. Les environnements sont « catégorisés » et leurs paramètres explicatifs (dimensions et caractéristiques des murs) sont renseignés dans des plans d'expériences construits à partir d'informations de la littérature spécialisée. La méthode de l'hypercube latin, jugée simple et plus efficace que la méthode de Monte Carlo, a été utilisée pour l'échantillonnage de l'espace stochastique. Les éventuelles dépendances entre variables d'entrée sont traitées au moyen de copule.Concernant la variabilité morphologique, nous avons également adopté une approche par simulation en utilisant CST Studio Suite®. Nous avons fait l'hypothèse simplificatrice de considérer des fantômes homogènes. Cette approche « raisonnable » permet surtout d'obtenir « facilement » une variabilité anthropométrique importante à partir de logiciels de synthèses d'images. La variabilité des sujets a été considérée au moyen de deux critères anthropométriques : l'Indice de Masse Corporel et la Circonférence Abdominale au moyen de données représentatives et d'études statistiques.Pour ces variabilités et les observables étudiées, l'ajustement des modèles paramétriquo-statistiques obtenus par régression multilinéaire est satisfaisant à très satisfaisant selon le lien radio considéré. Ils sont de complexité modérée et définis par un nombre limité de paramètres.
• PerfectDart: Automatic Dart Design for Garment Fitting
  
  • de Malefette Charles
  • Qi Anran
  • Parakkat Amal Dev
  • Cani Marie-Paule
  • Igarashi Takeo
  , 2023, pp.1-4. Dart, a triangle-shaped folded and stitched tuck in a garment, is a common sewing technique used to provide custom-fit garments. Unfortunately, designing and optimally placing these darts requires knowledge and practice, making it challenging for novice users. We propose a novel computational dart design framework that takes rough user cues (the region where the dart will be inserted) and computes the optimal dart configurations to improve fitness. To be more specific, our framework utilizes the body-garment relationship to quantify the fitting using a novel energy composed of three geometric terms: 1) closeness term encoding the proximity between the garment and the target body, 2) stretchability term favouring area-preserving cloth deformation, and 3) smoothness term promoting an unwrinkled and unfolded garment. We evaluate these three geometric terms via off-the-shelf cloth simulation and use it to optimize the dart configuration by minimizing the energy. As demonstrated by our results, our method is able to automatically generate darts to improve fitness for various garment designs and a wide range of body shapes, including animals. CCS CONCEPTS • Computing methodologies → Computer graphics. (10.1145/3610543.3626154)
  DOI : 10.1145/3610543.3626154
• Quantum Security of the UMTS-AKA Protocol and its Primitives, Milenage and TUAK
  
  • Frixons Paul
  • Canard Sébastien
  • Ferreira Loïc
  , 2023. The existence of a quantum computer is one of the most significant threats cryptography has ever faced. However, it seems that real world protocols received little attention so far with respect to their future security. Indeed merely relying upon post-quantum primitives may not suffice in order for a security protocol to be resistant in a full quantum world. In this paper, we consider the fundamental UMTS key agreement used in 3G but also in 4G (LTE), and in the (recently deployed) 5G technology. We analyze the protocol in a quantum setting, with quantum communications (allowing superposition queries by the involved parties), and where quantum computation is granted to the adversary. We prove that, assuming the underlying symmetric-key primitive is quantum-secure, the UMTS key agreement is also quantum-secure. We also give a quantum security analysis of the underlying primitives, namely Milenage and TUAK. To the best of our knowledge this paper provides the first rigorous proof of the UMTS key agreement in a strong quantum setting. Our result shows that in the quantum world to come, the UMTS technology remains a valid scheme in order to secure the communications of billions of users.
• AUTOMATIC GENERATION of 3D ANATOMICAL MODELS
  
  • Delmonte Alessandro
  • Sarnacki Sabine
  • Bloch Isabelle
  • Gori Pietro
  • Muller Cecile
  • Virzi Alessio
  • La Barbera Giammarco
  , 2023.
• Active Bipartite Ranking
  
  • Cheshire James
  • Clémençon Stéphan
  • Laurent Vincent
  , 2023. In this paper, we develop an active learning framework for the bipartite ranking problem. Motivated by numerous applications, ranging from supervised anomaly detection to credit-scoring through the design of medical diagnosis support systems, and usually formulated as the problem of optimizing (a scalar summary of) the ROC curve, bipartite ranking has been the subject of much attention in the passive context. Various dedicated algorithms have been recently proposed and studied by the machine-learning community. In contrast, active bipartite ranking rule is poorly documented in the literature. Due to its global nature, a strategy for labeling sequentially data points that are difficult to rank w.r.t. to the others is required. This learning task is much more complex than binary classification, for which many active algorithms have been designed. It is the goal of this article to provide a rigorous formulation of such a selective sampling approach. We propose a dedicated algorithm, referred to as active-rank, which aims to minimise the distance between the ROC curve of the ranking function built and the optimal one, w.r.t. the sup norm. We show that, for a fixed confidence level ε and probability δ, active-rank is PAC(ε, δ). In addition, we provide a problem dependent upper bound on the expected sampling time of active-rank and also demonstrate a problem dependent lower bound on the expected sampling time of any PAC(ε, δ) algorithm. Beyond the theoretical analysis carried out, numerical results are presented, providing strong empirical evidence of the performance of the algorithm proposed, which compares favorably with more naive approaches.
• LegalBench: A Collaboratively Built Benchmark for Measuring Legal Reasoning in Large Language Models
  
  • Guha Neel
  • Nyarko Julian
  • Ho Daniel E.
  • Ré Christopher
  • Chilton Adam
  • Narayana Aditya
  • Chohlas-Wood Alex
  • Peters Austin
  • Waldon Brandon
  • Rockmore Daniel N.
  • Zambrano Diego
  • Talisman Dmitry
  • Hoque Enam
  • Surani Faiz
  • Fagan Frank
  • Sarfaty Galit
  • Dickinson Gregory M.
  • Porat Haggai
  • Hegland Jason
  • Wu Jessica
  • Nudell Joe
  • Niklaus Joel
  • Nay John
  • Choi Jonathan H.
  • Tobia Kevin
  • Hagan Margaret
  • Ma Megan
  • Livermore Michael
  • Rasumov-Rahe Nikon
  • Holzenberger Nils
  • Kolt Noam
  • Henderson Peter
  • Rehaag Sean
  • Goel Sharad
  • Gao Shang
  • Williams Spencer
  • Gandhi Sunny
  • Zur Tom
  • Iyer Varun
  • Li Zehua
  , 2023, 4583531, pp.1-143. The advent of large language models (LLMs) and their adoption by the legal community has given rise to the question: what types of legal reasoning can LLMs perform? To enable greater study of this question, we present LegalBench: a collaboratively constructed legal reasoning benchmark consisting of 162 tasks covering six different types of legal reasoning. LegalBench was built through an interdisciplinary process, in which we collected tasks designed and hand-crafted by legal professionals. Because these subject matter experts took a leading role in construction, tasks either measure legal reasoning capabilities that are practically useful, or measure reasoning skills that lawyers find interesting. To enable cross-disciplinary conversations about LLMs in the law, we additionally show how popular legal frameworks for describing legal reasoning -- which distinguish between its many forms -- correspond to LegalBench tasks, thus giving lawyers and LLM developers a common vocabulary. This paper describes LegalBench, presents an empirical evaluation of 20 open-source and commercial LLMs, and illustrates the types of research explorations LegalBench enables. (10.2139/ssrn.4583531)
  DOI : 10.2139/ssrn.4583531
• Fault-tolerant four-dimensional constellation for coherent optical transmission systems
  
  • Liu Jingtian
  • Awwad Élie
  • Jaouën Yves
  Optics Express, Optical Society of America - OSA Publishing, 2023, 31 (26), pp.43449-43461. We propose a 4-dimensional 2-ary amplitude ring-switched modulation format with 64 symbols, which is denoted as 4D-2A-RS64 encoded over two polarization tributaries to improve the transmission performance over long-haul optical fibers in the presence of the non-linear Kerr effect. At a spectral efficiency of 6 bits per 4D, simulation results show that this format outperforms the polarization division multiplexed (PDM) 8QAM-star modulation as well as the 4D-2A-8PSK over links without inline dispersion management. We evaluate the performance for a WDM transmission of 11 × 90~Gbaud channels over a multi-span SSMF link. For an achievable information rate of 4.8bit/s/Hz, the maximum transmission distance is improved by 10.6% (400 km) and 4% (160 km) compared to PDM-8QAM-star and 4D-2A-8PSK respectively. The achieved gains are composed of a linear part and a non-linear part, respectively from the improved Euclidean-distance distribution and the constant power property of the 4D modulation. The geometric shaping of the proposed scheme is easy to implement and is robust to Mach-Zehnder modulator (MZM) imbalances and quantization errors stemming from the finite digital-to-analog converter (DAC) resolution. This robustness is compared to the one of other geometric-shaped non-linearity tolerant 4D schemes such as the 4D-2A-8PSK and the 4D-64PRS that can be both outperformed by our scheme in severe conditions. (10.1364/OE.504506)
  DOI : 10.1364/OE.504506