According to security.org, as of 2023, 65% of credit card (CC) users in the US have been subjected to fraud at some point in their lives, which equates to about 151 million Americans. The proliferation of advanced machine learning (ML) algorithms has also contributed to detecting credit card fraud (CCF). However, using a single or static ML-based defense model against a constantly evolving adversary takes its structural advantage, which enables the adversary to reverse engineer the defense's strategy over the rounds of an iterated game. This paper proposes an adaptive moving target defense (MTD) approach based on deep reinforcement learning (DRL), termed RESONANT to identify the optimal switching points to another ML classifier for credit card fraud detection. It identifies optimal moments to strategically switch between different ML-based defense models (i.e., classifiers) to invalidate any adversarial progress and always stay a step ahead of the adversary. We take this approach in an iterated game theoretic manner where the adversary and defender take turns to take their action in the CCF detection contexts. Via extensive simulation experiments, we investigate the performance of our proposed RESONANT against that of the existing state-of-the-art counterparts in terms of the mean and variance of detection accuracy and attack success ratio to measure the defensive performance. Our results demonstrate the superiority of RESONANT over other counterparts, including static and naïve ML and MTD selecting a defense model at random (i.e., Random-MTD). Via extensive simulation experiments, our results show that our proposed RESONANT can outperform the existing counterparts up to two times better performance in detection accuracy using AUC (i.e., Area Under the Curve of the Receiver Operating Characteristic (ROC) curve) and system security against attacks using attack success ratio (ASR). / Master of Science / According to security.org, as of 2023, 65% of credit card (CC) users in the US have been subjected to fraud at some point in their lives, which equates to about 151 million Americans. The proliferation of advanced machine learning (ML) algorithms has also contributed to detecting credit card fraud (CCF). However, using a single or static ML-based defense model against a constantly evolving adversary takes its structural advantage, which enables the adversary to reverse engineer the defense's strategy over the rounds of an iterated game. This paper proposes an adaptive defense approach based on artificial intelligence (AI), termed RESONANT, to identify the optimal switching points to another ML classifiers for credit card fraud detection. It identifies optimal moments to strategically switch between different ML-based defense models (i.e., classifiers) to invalidate any adversarial progress and always stay a step ahead of the adversary. We take this approach in an iterated game theoretic manner where the adversary and defender take turns to take their action in the CCF detection contexts. Via extensive simulation experiments, we investigate the performance of our proposed RESONANT against that of the existing state-of-the-art counterparts in terms of the mean and variance of detection accuracy and attack success ratio to measure the defensive performance. Our results demonstrate the superiority of RESONANT over other counterparts, showing that our proposed RESONANT can outperform the existing counterparts by up to two times better performance in detection accuracy and system security against attacks.

Cette thèse porte sur la synthèse de séquences de motion capture avec des modèles statistiques. La synthèse de ce type de séquences est une tâche pertinente pour des domaines d'application divers tels que le divertissement, l'interaction homme-machine, la robotique, etc. Du point de vue de l'apprentissage machine, la conception de modèles de synthèse consiste à apprendre des modèles génératifs, ici pour des données séquentielles. Notre point de départ réside dans deux problèmes principaux rencontrés lors de la synthèse de données de motion capture, assurer le réalisme des positions et des mouvements, et la gestion de la grande variabilité dans ces données. La variabilité vient d'abord des caractéristiques individuelles, nous ne bougeons pas tous de la même manière mais d'une façon qui dépend de notre personnalité, de notre sexe, de notre âge de notre morphologie, et de facteurs de variation plus court terme tels que notre état émotionnel, que nous soyons fatigués, etc.Une première partie présente des travaux préliminaires que nous avons réalisés en étendant des approches de l'état de l'art basées sur des modèles de Markov cachés et des processus gaussiens pour aborder les deux problèmes principaux liés au réalisme et à la variabilité. Nous décrivons d'abord une variante de modèles de Markov cachés contextuels pour gérer la variabilité dans les données en conditionnant les paramètres des modèles à une information contextuelle supplémentaire telle que l'émotion avec laquelle un mouvement a été effectué. Nous proposons ensuite une variante d'une méthode de l'état de l'art utilisée pour réaliser une tâche de synthèse de mouvement spécifique appelée Inverse Kinematics, où nous exploitons les processus gaussiens pour encourager le réalisme de chacune des postures d'un mouvement généré. Nos résultats montrent un certain potentiel de ces modèles statistiques pour la conception de systèmes de synthèse de mouvement humain. Pourtant, aucune de ces technologies n'offre la flexibilité apportée par les réseaux de neurones et la récente révolution de l'apprentissage profond et de l'apprentissage Adversarial que nous abordons dans la deuxième partie.La deuxième partie de la thèse décrit les travaux que nous avons réalisés avec des réseaux de neurones et des architectures profondes. Nos travaux s'appuient sur la capacité des réseaux neuronaux récurrents à traiter des séquences complexes et sur l'apprentissage Adversarial qui a été introduit très récemment dans la communauté du Deep Learning pour la conception de modèles génératifs performants pour des données complexes, notamment images. Nous proposons une première architecture simple qui combine l'apprentissage Adversarial et des autoencodeurs de séquences, qui permet de mettre au point des systèmes performants de génération aléatoire de séquences réalistes de motion capture. A partir de cette architecture de base, nous proposons plusieurs variantes d'architectures neurales conditionnelles qui permettent de concevoir des systèmes de synthèse que l'on peut contrôler dans une certaine mesure en fournissant une information de haut niveau à laquelle la séquence générée doit correspondre, par exemple l'émotion avec laquelle une activité est réalisée. Pour terminer nous décrivons une dernière variante qui permet de réaliser de l'édition de séquences de motion capture, où le système construit permet de générer une séquence dans le style d'une autre séquence, réelle. / This thesis focuses on the synthesis of motion capture data with statistical models. Motion synthesis is a task of interest for important application fields such as entertainment, human-computer interaction, robotics, etc. It may be used to drive a virtual character that can be involved in the applications of the virtual reality, animation films or computer games. This thesis focuses on the use of statistical models for motion synthesis with a strong focus on neural networks. From the machine learning point of view designing synthesis models consists in learning generative models. Our starting point lies in two main problems one encounters when dealing with motion capture data synthesis, ensuring realism of postures and motion, and handling the large variability in the synthesized motion. The variability in the data comes first from core individual features, we do not all move the same way but accordingly to our personality, our gender, age, and morphology etc. Moreover there are other short term factors of variation like our emotion, the fact that we are interacting with somebody else, that we are tired etc. Data driven models have been studied for generating human motion for many years. Models are learned from labelled datasets where motion capture data are recorded while actors are performed various activities like walking, dancing, running, etc. Traditional statistical models such as Hidden Markov Models, Gaussian Processes have been investigated for motion synthesis, demonstrating strengths but also weaknesses. Our work focuses in this line of research and concerns the design of generative models for sequences able to take into account some contextual information, which will represent the factors of variation. A first part of the thesis present preliminary works that we realised by extending previous approaches relying on Hidden Markov Models and Gaussian Processes to tackle the two main problems related to realism and variability. We first describe an attempt to extend contextual Hidden Markov Models for handling variability in the data by conditioning the parameters of the models to an additional contextual information such as the emotion which which a motion was performed. We then propose a variant of a traditional method for performing a specific motion synthesis task called Inverse Kinematics, where we exploit Gaussian Processes to enforce realism of each of the postures of a generated motion. These preliminary results show some potential of statistical models for designing human motion synthesis systems. Yet none of these technologies offers the flexibility brought by neural networks and the recent deep learning revolution.The second part of the thesis describes the works we realized with neural networks and deep architectures. It builds on recurrent neural networks for dealing with sequences and on adversarial learning which was introduced very recently in the deep learning community for designing accurate generative models for complex data. We propose a simple system as a basis synthesis architecture, which combines adversarial learning with sequence autoencoders, and that allows randomly generating realistic motion capture sequences. Starting from this architecture we design few conditional neural models that allow to design synthesis systems that one can control up to some extent by either providing a high level information that the generated sequence should match (e.g. the emotion) or by providing a sequence in the style of which a sequence should be generated.

Smart wearables, known as emerging paradigms for vital big data capturing, have been attracting intensive attentions. However, one crucial problem is their power-hungriness, i.e., the continuous data streaming consumes energy dramatically and requires devices to be frequently charged. Targeting this obstacle, we propose to investigate the biodynamic patterns in the data and design a data-driven approach for intelligent data compression. We leverage Deep Learning (DL), more specifically, Convolutional Autoencoder (CAE), to learn a sparse representation of the vital big data. The minimized energy need, even taking into consideration the CAE-induced overhead, is tremendously lower than the original energy need. Further, compared with state-of-the-art wavelet compression-based method, our method can compress the data with a dramatically lower error for a similar energy budget. Our experiments and the validated approach are expected to boost the energy efficiency of wearables, and thus greatly advance ubiquitous big data applications in era of smart health.<br><div>In recent years, there has also been a growing interest in edge intelligence for emerging instantaneous big data inference. However, the inference algorithms, especially deep learning, usually require heavy computation requirements, thereby greatly limiting their deployment on the edge. We take special interest in the smart health wearable big data mining and inference. <br></div><div><br></div><div>Targeting the deep learning’s high computational complexity and large memory and energy requirements, new approaches are urged to make the deep learning algorithms ultra-efficient for wearable big data analysis. We propose to leverage knowledge distillation to achieve an ultra-efficient edge-deployable deep learning model. More specifically, through transferring the knowledge from a teacher model to the on-edge student model, the soft target distribution of the teacher model can be effectively learned by the student model. Besides, we propose to further introduce adversarial robustness to the student model, by stimulating the student model to correctly identify inputs that have adversarial perturbation. Experiments demonstrate that the knowledge distillation student model has comparable performance to the heavy teacher model but owns a substantially smaller model size. With adversarial learning, the student model has effectively preserved its robustness. In such a way, we have demonstrated the framework with knowledge distillation and adversarial learning can, not only advance ultra-efficient edge inference, but also preserve the robustness facing the perturbed input.</div>

Machine learning models are vulnerable to adversarial attacks that add perturbations to the input data. Here we model and simulate power flow in a power grid test case and generate adversarial attacks for these measurements in three different ways. This is to compare the effect of attacks of different sizes constructed using various levels of knowledge of the model to see how this affects how often the attacks are detected. The three methods being one where the attacker has full knowledge of model, one where the attacker only has access to the measurements of the model, and the third method where the attacker has no knowledge of the model. By comparing these methods through how often they are detected by a residual-based detection scheme, one can argue that a data-driven attack only knowing the measurements is enough to add an error without being detected by the detection scheme. Using a linearized version of a state estimation is shown to be insufficient for generating attacks with full knowledge of the system, and further research is needed to compare the performance of the full knowledge attacks and the data-driven attacks. The attacks generated without knowledge of the system perform poorly and are easily detected.

<p>Over the past decade, distributed representation learning has emerged as a popular alternative to conventional centralized machine learning training. The increasing interest in distributed representation learning, specifically federated learning, can be attributed to its fundamental property that promotes data privacy and communication savings. While conventional ML encourages aggregating data at a central location (e.g., data centers), distributed representation learning advocates keeping data at the source and instead transmitting model parameters across the network. However, since the advent of deep learning, model sizes have become increasingly large often comprising million-billions of parameters, which leads to the problem of communication latency in the learning process. In this thesis, we propose to tackle the problem of communication latency in two different ways: (i) learning private representation of data to enable its sharing, and (ii) reducing the communication latency by minimizing the corresponding long-range communication requirements.</p> <p><br></p> <p>To tackle the former goal, we first start by studying the problem of learning representations that are private yet informative, i.e., providing information about intended ''ally'' targets while hiding sensitive ''adversary'' attributes. We propose Exclusion-Inclusion Generative Adversarial Network (EIGAN), a generalized private representation learning (PRL) architecture that accounts for multiple ally and adversary attributes, unlike existing PRL solutions. We then address the practical constraints of the distributed datasets by developing Distributed EIGAN (D-EIGAN), the first distributed PRL method that learns a private representation at each node without transmitting the source data. We theoretically analyze the behavior of adversaries under the optimal EIGAN and D-EIGAN encoders and the impact of dependencies among ally and adversary tasks on the optimization objective. Our experiments on various datasets demonstrate the advantages of EIGAN in terms of performance, robustness, and scalability. In particular, EIGAN outperforms the previous state-of-the-art by a significant accuracy margin (47% improvement), and D-EIGAN's performance is consistently on par with EIGAN under different network settings.</p> <p><br></p> <p>We next tackle the latter objective - reducing the communication latency - and propose two timescale hybrid federated learning (TT-HF), a semi-decentralized learning architecture that combines the conventional device-to-server communication paradigm for federated learning with device-to-device (D2D) communications for model training. In TT-HF, during each global aggregation interval, devices (i) perform multiple stochastic gradient descent iterations on their individual datasets, and (ii) aperiodically engage in consensus procedure of their model parameters through cooperative, distributed D2D communications within local clusters. With a new general definition of gradient diversity, we formally study the convergence behavior of TT-HF, resulting in new convergence bounds for distributed ML. We leverage our convergence bounds to develop an adaptive control algorithm that tunes the step size, D2D communication rounds, and global aggregation period of TT-HF over time to target a sublinear convergence rate of O(1/t) while minimizing network resource utilization. Our subsequent experiments demonstrate that TT-HF significantly outperforms the current art in federated learning in terms of model accuracy and/or network energy consumption in different scenarios where local device datasets exhibit statistical heterogeneity. Finally, our numerical evaluations demonstrate robustness against outages caused by fading channels, as well favorable performance with non-convex loss functions.</p>

<p>Artificial neural networks (ANNs) have achieved extraordinary performance in various domains in recent years. However, some studies reveal that ANNs may be vulnerable in three aspects: label scarcity, perturbations, and open-set emerging classes. Noisy labeling and self-supervised learning approaches address the label scarcity issues, but most of the work couldn't handle the perturbations. Adversarial training methods, topological denoising methods, and mechanism designing methods aim to mitigate the negative effects caused by perturbations. However, adversarial training methods can barely train a robust model under the circumstance of extensive label scarcity; topological denoising methods are not efficient on dynamic data structures; and mechanism designing methods often depend on heuristic explorations. Detection-based methods devote to identifying novel or anomaly instances for further downstream tasks. Nonetheless, such instances may belong to open-set new emerging classes. To embrace the aforementioned challenges, we address the robustness issues of ANNs from two aspects. First, we propose a series of Bayesian label transition models to improve the robustness of Graph Neural Networks (GNNs) in the presence of label scarcity and perturbations in the graph domain. Second, we propose a new non-exhaustive learning model, named NE-GM-GAN, to handle both open-set problems and class-imbalance issues in network intrusion datasets. Extensive experiments with several datasets demonstrate that our proposed models can effectively improve the robustness of ANNs.</p>

Ce travail expose plusieurs études dans les domaines de la reconnaissance de la parole et compréhension du langage parlé. La compréhension sémantique du langage parlé est un sous-domaine important de l'intelligence artificielle. Le traitement de la parole intéresse depuis longtemps les chercheurs, puisque la parole est une des charactéristiques qui definit l'être humain. Avec le développement du réseau neuronal artificiel, le domaine a connu une évolution rapide à la fois en terme de précision et de perception humaine. Une autre étape importante a été franchie avec le développement d'approches bout en bout. De telles approches permettent une coadaptation de toutes les parties du modèle, ce qui augmente ainsi les performances, et ce qui simplifie la procédure d'entrainement. Les modèles de bout en bout sont devenus réalisables avec la quantité croissante de données disponibles, de ressources informatiques et, surtout, avec de nombreux développements architecturaux innovateurs. Néanmoins, les approches traditionnelles (qui ne sont pas bout en bout) sont toujours pertinentes pour le traitement de la parole en raison des données difficiles dans les environnements bruyants, de la parole avec un accent et de la grande variété de dialectes. Dans le premier travail, nous explorons la reconnaissance de la parole hybride dans des environnements bruyants. Nous proposons de traiter la reconnaissance de la parole, qui fonctionne dans un nouvel environnement composé de différents bruits inconnus, comme une tâche d'adaptation de domaine. Pour cela, nous utilisons la nouvelle technique à l'époque de l'adaptation du domaine antagoniste. En résumé, ces travaux antérieurs proposaient de former des caractéristiques de manière à ce qu'elles soient distinctives pour la tâche principale, mais non-distinctive pour la tâche secondaire. Cette tâche secondaire est conçue pour être la tâche de reconnaissance de domaine. Ainsi, les fonctionnalités entraînées sont invariantes vis-à-vis du domaine considéré. Dans notre travail, nous adoptons cette technique et la modifions pour la tâche de reconnaissance de la parole dans un environnement bruyant. Dans le second travail, nous développons une méthode générale pour la régularisation des réseaux génératif récurrents. Il est connu que les réseaux récurrents ont souvent des difficultés à rester sur le même chemin, lors de la production de sorties longues. Bien qu'il soit possible d'utiliser des réseaux bidirectionnels pour une meilleure traitement de séquences pour l'apprentissage des charactéristiques, qui n'est pas applicable au cas génératif. Nous avons développé un moyen d'améliorer la cohérence de la production de longues séquences avec des réseaux récurrents. Nous proposons un moyen de construire un modèle similaire à un réseau bidirectionnel. L'idée centrale est d'utiliser une perte L2 entre les réseaux récurrents génératifs vers l'avant et vers l'arrière. Nous fournissons une évaluation expérimentale sur une multitude de tâches et d'ensembles de données, y compris la reconnaissance vocale, le sous-titrage d'images et la modélisation du langage. Dans le troisième article, nous étudions la possibilité de développer un identificateur d'intention de bout en bout pour la compréhension du langage parlé. La compréhension sémantique du langage parlé est une étape importante vers le développement d'une intelligence artificielle de type humain. Nous avons vu que les approches de bout en bout montrent des performances élevées sur les tâches, y compris la traduction automatique et la reconnaissance de la parole. Nous nous inspirons des travaux antérieurs pour développer un système de bout en bout pour la reconnaissance de l'intention. / This work presents several studies in the areas of speech recognition and understanding. The semantic speech understanding is an important sub-domain of the broader field of artificial intelligence. Speech processing has had interest from the researchers for long time because language is one of the defining characteristics of a human being. With the development of neural networks, the domain has seen rapid progress both in terms of accuracy and human perception. Another important milestone was achieved with the development of end-to-end approaches. Such approaches allow co-adaptation of all the parts of the model thus increasing the performance, as well as simplifying the training procedure. End-to-end models became feasible with the increasing amount of available data, computational resources, and most importantly with many novel architectural developments. Nevertheless, traditional, non end-to-end, approaches are still relevant for speech processing due to challenging data in noisy environments, accented speech, and high variety of dialects. In the first work, we explore the hybrid speech recognition in noisy environments. We propose to treat the recognition in the unseen noise condition as the domain adaptation task. For this, we use the novel at the time technique of the adversarial domain adaptation. In the nutshell, this prior work proposed to train features in such a way that they are discriminative for the primary task, but non-discriminative for the secondary task. This secondary task is constructed to be the domain recognition task. Thus, the features trained are invariant towards the domain at hand. In our work, we adopt this technique and modify it for the task of noisy speech recognition. In the second work, we develop a general method for regularizing the generative recurrent networks. It is known that the recurrent networks frequently have difficulties staying on same track when generating long outputs. While it is possible to use bi-directional networks for better sequence aggregation for feature learning, it is not applicable for the generative case. We developed a way improve the consistency of generating long sequences with recurrent networks. We propose a way to construct a model similar to bi-directional network. The key insight is to use a soft L2 loss between the forward and the backward generative recurrent networks. We provide experimental evaluation on a multitude of tasks and datasets, including speech recognition, image captioning, and language modeling. In the third paper, we investigate the possibility of developing an end-to-end intent recognizer for spoken language understanding. The semantic spoken language understanding is an important step towards developing a human-like artificial intelligence. We have seen that the end-to-end approaches show high performance on the tasks including machine translation and speech recognition. We draw the inspiration from the prior works to develop an end-to-end system for intent recognition.

<p>  </p> <p>Hyperspectral Imaging is one of the most popular technologies in plant phenotyping due to its ability to predict the plant physiological features such as yield biomass, leaf moisture, and nitrogen content accurately, non-destructively, and efficiently. Various kinds of hyperspectral imaging systems have been developed in the past years for both greenhouse and field phenotyping activities. Developing the plant physiological prediction model such as relative water content (RWC) using hyperspectral imaging data requires the adoption of machine learning-based calibration techniques. Convolutional neural networks (CNNs) have been known to automatically extract the features from the raw data which can lead to highly accurate physiological prediction models. Once a reliable prediction model has been developed, sharing that model across multiple hyperspectral imaging systems is very desirable since collecting the large number of ground truth labels for predictive model development is expensive and tedious. However, there are always significant differences in imaging sensors, imaging, and environmental conditions between different hyperspectral imaging facilities, which makes it difficult to share plant features prediction models. Calibration transfer between the imaging systems is critically important. In this thesis, two approaches were taken to address the calibration transfer from the greenhouse to the field. First, direct standardization (DS), piecewise direct standardization (PDS), double window piecewise direct standardization (DPDS) and spectral space transfer (SST) were used for standardizing the spectral reflectance to minimize the artifacts and spectral differences between different greenhouse imaging systems. A linear transformation matrix estimated using SST based on a small set of plant samples imaged in two facilities reduced the root mean square error (RMSE) for maize physiological feature prediction significantly, i.e., from 10.64% to 2.42% for RWC and from 1.84% to 0.11% for nitrogen content. Second, common latent space features between two greenhouses or a greenhouse and field imaging system were extracted in an unsupervised fashion. Two different models based on deep adversarial domain adaptation are trained, evaluated, and tested. In contrast to linear standardization approaches developed using the same plant samples imaged in two greenhouse facilities, the domain adaptation extracted non-linear features common between spectra of different imaging systems. Results showed that transferred RWC models reduced the RMSE by up to 45.9% for the greenhouse calibration transfer and 12.4% for a greenhouse to field transfer. The plot scale evaluation of the transferred RWC model showed no significant difference between the measurements and predictions. The methods developed and reported in this study can be used to recover the performance plummeted due to the spectral differences caused by the new phenotyping system and to share the knowledge among plant phenotyping researchers and scientists.</p>