Recent research shows that machine learning algorithms are highly susceptible to attacks trying to extract sensitive information about the data used in model training. These attacks called privacy attacks, exploit the model training process. Contemporary defense techniques make alterations to the training algorithm. Such defenses are computationally expensive, cause a noticeable privacy-utility tradeoff, and require control over the training process. This thesis presents a data-centric approach using data augmentations to mitigate privacy attacks. We present privacy-focused data augmentations to change the sensitive data submitted to the model trainer. Compared to traditional defenses, our method provides more control to the individual data owner to protect one's private data. The defense is model-agnostic and does not require the data owner to have any sort of control over the model training. Privacypreserving augmentations are implemented for two attacks namely membership inference and model inversion using two distinct techniques. While the proposed augmentations offer a better privacy-utility tradeoff on CIFAR-10 for membership inference, they reduce the reconstruction rate to ≤ 1% while reducing the classification accuracy by only 2% against model inversion attacks. This is the first attempt to defend model inversion and membership inference attacks using decentralized privacy protection. / Master of Science / Privacy attacks are threats posed to extract sensitive information about the data used to train machine learning models. As machine learning is used extensively for many applications, they have access to private information like financial records, medical history, etc depending on the application. It has been observed that machine learning models can leak the information they contain. As models tend to 'memorize' training data to some extent, even removing the data from the training set cannot prevent privacy leakage. As a result, the research community has focused its attention on developing defense techniques to prevent this information leakage. However, the existing defenses rely heavily on making alterations to the way a machine learning model is trained. This approach is termed as a model-centric approach wherein the model owner is responsible to make changes to the model algorithm to preserve data privacy. By doing this, the model performance is degraded while upholding data privacy. Our work introduces the first data-centric defense which provides the tools to protect the data to the data owner. We demonstrate the effectiveness of the proposed defense in providing protection while ensuring that the model performance is maintained to a great extent.

The cause and distribution of spatial variations in the mechanical properties of the continental lithosphere are fundamental questions for modern geology. In this study variations in long-term lithospheric rigidity have been investigated. These investigations used profile- and grid-based flexural models of the lithosphere’s response to geologically imposed topographic, or buried, loads. These models were constrained by topographic and gravity data allowing recovery of best fitting rigidity values. In Oman a Cretaceous ophiolite acts as a significant load on the continental crust. Flexural models along profiles orthogonal to the ophiolite strike show that the observed gravity data can be best modelled by an elastic beam with standard thickness (T_e) of 30 km. Along strike there is shown to be significant variation in the foreland shape and the observed gravity signal. This, it is proposed, relates to the complex tectonic processes which occurred as the ophiolite was obducted. The Himalayan foreland has been the focus of controversy over the recovered long-term rigidity of the continents, with recovered T_e values ranging from 40 to over 90 km. Both profile- and grid-based techniques show that T_e is high (>70 km) in the foreland region. Across the India-Eurasia collisional system as a whole T_e values are variable. Beneath the Tibetan plateau recovered values are generally low (<10 km), while the plateau margins are marked by regions of higher rigidity. Recovered T_e values across the Arabia-Eurasia collisional system range from over 60 km in the foreland region to close to zero beneath the high Zagros mountains. In the eastern part of the foreland, flexural models match the gravity data; however, they disagree with sediment thickness data for the material infilling the foreland. This discrepancy is interpreted in terms of de-coupling of the flexural lithosphere from the shallower crustal levels, caused by the presence of significant salt deposits in this region. Application of grid-based techniques to South America, North America and Europe recover a broad range of Te values from ∼0 to over 90 km. The low T_e values are explained in active orogenic belts in terms of current processes acting to weaken the lithosphere, and in the continental interiors as the relics of past orogenic events. High T_e values in the continental interiors correlate with ancient cratonic cores which have undergone little deformation since their formation in the Archean. This study shows that T_e variations have a critical influence on the development of large compressional orogenic belts. In the Himalayan and Andean orogens there is a correlation between the over-thrusting of the orogenic belt and high T_e foreland regions. Where lower T_e regions are seen, less over thrusting is apparent, and in the case of the India-Eurasia collisional system out-flow of lower crustal material may be occurring.

No description available.

In this thesis, adaptive neural network controllers are designed for a high subsonic cruise missile. Two autopilot designs are included in the study using adaptive neural networks, namely an altitude hold autopilot designed for the longitudinal channel and a directional autopilot designed for heading control. Aerodynamic coefficients are obtained using missile geometry / a 5-Degree of Freedom (5-DOF) simulation model is obtained, and linearized at a single trim condition. An inverted model is used in the controller. Adaptive Neural Network (ANN) controllers namely, model inversion controllers with Sigma-Pi Neural Network, Single Hidden Layer Neural Network and Background Learning implemented Single Hidden Layer Neural Network, are deployed to cancel the modeling error and are applied for the longitudinal and directional channels of the missile. This approach simplifies the autopilot designing process by combining a controller with model inversion designed for a single flight condition with an on-line learning neural network to account for errors that are caused due to the approximate inversion. Simulations are performed both in the longitudinal and directional channels in order to demonstrate the effectiveness of the implemented control algorithms. The advantages and drawbacks of the implemented neural network based controllers are indicated.

Le secteur du bâtiment, responsable de plus de 40% de consommation d’énergie globale et un tiers des émissions de gaz à effet de serre mondial, est un des centres de préoccupations autour des sujets liés au changement climatique et l’indépendance énergétique. Le travail de recherche a exigé l’apport de connaissances supplémentaires et la création d’outils spécifiques orientés sur l’optimisation globale du management énergétique des bâtiments de type tertiaire. Une problématique industrielle est associée à ces enjeux de transitions énergétique et écologique, à savoir le frein observé à la mise en place de plans d’actions de rénovation. En effet, pour des opérations d’optimisation ou de rénovation de petites-moyennes envergures, les coûts initiaux d’études et de métrologie représentent plus de 50% de leur coût global. Cette mise de fonds induit un retour sur investissement très long. Face à ce paramètre financier prohibitif, beaucoup d’entreprises sont réticentes à mettre en place ce type d’action. L’objectif opérationnel a donc été de proposer une solution permettant de réduire drastiquement ces coûts préliminaires.Les aspects abordés dans la thèse sont : l’état de l’art du fonctionnement du bâtiment et des enjeux associés, la création d’un outil de collecte et de remontée des données de fonctionnement et de performance du bâtiment grâce à un réseau de mesure in-situ dédié, concomitant à l’élaboration d’un modèle thermique simplifié adjoint facilitant la compréhension de son comportement, puis l’identification de ses paramètres "observables" de conception et de fonctionnement par méthode inverse, et enfin le calcul de sa consommation énergétique optimale grâce à une méthode d’optimisation. Plus spécifiquement, l’approche sera orientée vers le développement d’outils pour promouvoir un accès facilité à la réduction des consommations unitaires auprès des entreprises au niveau national et l’intégration d’une intelligence pour l’optimisation énergétique des éléments climatiques du bâtiment ou son usage, ou encore une interface ergonomique homme-machine permettant un management efficace de son fonctionnement. Dans les faits, le problème observé est holistique et ne peut pas être pris en compte de manière sectorielle. Il est impératif d’y intégrer tous les processus impliqués dans le bâtiment et son usage (aspect comportemental des usagers). L’approche utilisée a été orientée afin de prendre en compte ultérieurement des paramètres autres que strictement énergétique, tel que les coûts ou le confort / The building field is responsible of about 40% of global energy consumption and a third of world greenhouse gas emissions. It is a main concern subject in climate change issues and fossil fuel independency. The aim of the PhD work is to bring more knowledge about thermal modeling and to create specific tools which are capable of globally optimize the office building energy management. The industrial purpose is associated with its area of expertise, which is advice in energy and ecologic transition. It concerns the difficulty to implement a retrofit action planning. Indeed, for small or middle retrofit actions, the initial study and metrology costs represent over 50% of the overall cost. This down payment induces a long return of investment. Faced with this prohibitive financial parameter, a lot of companies are reluctant to implement this type of actions. The proposed purpose is a solution that drastically reduces preliminary costs. The aspects addressed in this thesis are: the building operation state of art and its associated issues, the creation of reporting and collecting data tool of building operation and performance thanks to a dedicated in-situ measurement network, concomitant with the development of a simplified adjoin thermal model. It facilitates the understanding of its behavior. Then the final aspect are the two steps of optimization. The first is the observable building design and operation parameters with an inverse method, the second is the calculation of optimal energy consumptions. The approach is specifically oriented through the development of tools allowing a facilitated access to energy reduction action for national companies. This should assist the integration of an intelligence for energy optimization for building climatics and thermal equipments or usage. The result could be a new ergonomic man-machine interface for stock building effective management. In the facts, the problem is holistic and cannot be handle sectorally. It is imperative to integrate all the process involved in the building and its use (user behavior). The approach have been oriented to take later into account other parameters than strictly energy, as costs of comfort

Monitoring biophysical variables at a global scale over long time periods is vital to address the climatechange and food security challenges. Leaf Area Index (LAI) is a structure variable giving a measure of the canopysurface for radiation interception and canopy-atmosphere interactions. LAI is an important variable in manyecosystem models and it has been recognized as an Essential Climate Variable. This thesis aims to provide globaland continuous estimates of LAI from satellite observations in near-real time according to user requirements to beused for diagnostic and prognostic evaluations of vegetation state and functioning. There are already someavailable LAI products which show however some important discrepancies in terms of magnitude and somelimitations in terms of continuity and consistency. This thesis addresses these important issues. First, the nature ofthe LAI estimated from these satellite observations was investigated to address the existing differences in thedefinition of products. Then, different temporal smoothing and gap filling methods were analyzed to reduce noiseand discontinuities in the time series mainly due to cloud cover. Finally, different methods for near real timeestimation of LAI were evaluated. Such comparison assessment as a function of the level of noise and gaps werelacking for LAI.Results achieved within the first part of the thesis show that the effective LAI is more accurately retrievedfrom satellite data than the actual LAI due to leaf clumping in the canopies. Further, the study has demonstratedthat multi-view observations provide only marginal improvements on LAI retrieval. The study also found that foroptimal retrievals the size of the uncertainty envelope over a set of possible solutions to be approximately equal tothat in the reflectance measurements. The results achieved in the second part of the thesis found the method withlocally adaptive temporal window, depending on amount of available observations and Climatology as backgroundestimation to be more robust to noise and missing data for smoothing, gap-filling and near real time estimationswith satellite time series.

Le suivi des variables biophysiques à l’échelle globale sur de longues périodes de temps est essentiellepour répondre aux nouveaux enjeux que constituent le changement climatique et la sécurité alimentaire. L’indice foliaire (LAI) est une variable de structure définissant la surface d’interception du rayonnement incident et d’échanges gazeux avec l’atmosphère. Le LAI est donc une variable importante des modèles d’écosystèmes et a d’ailleurs été reconnue comme variable climatique essentielle (ECV). Cette thèse a pour objectif de fournir des estimations globales et continues de LAI à partir d’observations satellitaires en temps quasi-réel en réponse aux besoins des utilisateurs pour fournir des diagnostiques et pronostiques de l’état et du fonctionnement de la végétation. Quelques produits LAI sont déjà disponibles mais montrent des désaccords et des limitations en termes de cohérence et de continuité. Cette thèse a pour objectif de lever ces limitations. Dans un premier temps, on essaiera de mieux définir la nature des estimations de LAI à partir d’observations satellitaires. Puis, différentes méthodes de lissage te bouchage des séries temporelles ont été analysées pour réduire le bruit et les discontinuités principalement liées à la couverture nuageuse. Finalement quelques méthodes d’estimation temps quasi réel ont été évaluées en considérant le niveau de bruit et les données manquantes.Les résultats obtenus dans la première partie de cette thèse montrent que la LAI effectif et bien mieux estimé que la valeur réelle de LAI du fait de l’agrégation des feuilles observée au niveau du couvert. L’utilisation d’observations multidirectionnelles n’améliore que marginalement les performances d’estimation. L’étude montre également que les performances d’estimation optimales sont obtenues quand les solutions sont recherchées à l’intérieur d’une enveloppe définie par l’incertitude associée aux mesures radiométriques. Dans la deuxième partie consacrée à l’amélioration de la continuité et la cohérence des séries temporelles, les méthodes basées sur une fenêtre temporelle locale mais de largeur dépendant du nombre d’observations présentes, et utilisant la climatologie comme information a priori s’avèrent les plus intéressantes autorisant également l’estimation en temps quasi réel. / Monitoring biophysical variables at a global scale over long time periods is vital to address the climatechange and food security challenges. Leaf Area Index (LAI) is a structure variable giving a measure of the canopysurface for radiation interception and canopy-atmosphere interactions. LAI is an important variable in manyecosystem models and it has been recognized as an Essential Climate Variable. This thesis aims to provide globaland continuous estimates of LAI from satellite observations in near-real time according to user requirements to beused for diagnostic and prognostic evaluations of vegetation state and functioning. There are already someavailable LAI products which show however some important discrepancies in terms of magnitude and somelimitations in terms of continuity and consistency. This thesis addresses these important issues. First, the nature ofthe LAI estimated from these satellite observations was investigated to address the existing differences in thedefinition of products. Then, different temporal smoothing and gap filling methods were analyzed to reduce noiseand discontinuities in the time series mainly due to cloud cover. Finally, different methods for near real timeestimation of LAI were evaluated. Such comparison assessment as a function of the level of noise and gaps werelacking for LAI.Results achieved within the first part of the thesis show that the effective LAI is more accurately retrievedfrom satellite data than the actual LAI due to leaf clumping in the canopies. Further, the study has demonstratedthat multi-view observations provide only marginal improvements on LAI retrieval. The study also found that foroptimal retrievals the size of the uncertainty envelope over a set of possible solutions to be approximately equal tothat in the reflectance measurements. The results achieved in the second part of the thesis found the method withlocally adaptive temporal window, depending on amount of available observations and Climatology as backgroundestimation to be more robust to noise and missing data for smoothing, gap-filling and near real time estimationswith satellite time series.

The dynamic wave model (the complete form of the saint-Venant equations), as applied to flow routing in irrigation canals or flood routing in natural channels, is associated with parameter and model uncertainties. The parameter uncertainty arises due to imprecision in the estimation of Manning’s n used for calculating the friction slope (sf) in the momentum equation of the dynamic wave model. Accurate estimation of n is difficult due to its dependence on several channel and flow characteristics. The model uncertainty of the dynamic wave model arises due to difficulty in applying the momentum equation to curved channels, as it is a vector equation. The one-dimensional form of the momentum equation is derived assuming that the longitudinal axis of the channel is a straight line, so that the net force vector is equal to the algebraic sum of the forces involved. Curved channel reaches have to be discretized into small straight sub-reaches while applying the momentum equation. Otherwise, two- or three-dimensional forms of the momentum equation need to be adopted. A main objective of the study presented in the thesis is to develop a fuzzy dynamic wave model (FDWM), which is capable of overcoming the parameter and model uncertainties of the dynamic wave model mentioned above, specifically for problems of flow routing in irrigation canals and flood routing in natural channels. It has been demonstrated earlier in literature that the problem of parameter uncertainty in infiltration models can be addressed by replacing the momentum equation by a fuzzy rule based model while retaining the continuity equation in its complete form. The FDWM is developed by adopting the same methodology: i.e. By replacing the momentum equation of the dynamic wave model by a fuzzy rule based model while retaining the continuity equation in its complete form. The fuzzy rule based model is developed based on fuzzification of a new equation for wave velocity, to account for the model uncertainty and backwater effects. A fuzzy dynamic wave routing model (FDWRM) is developed based on application of the FDWM to flow routing in irrigation canals. The fuzzy rule based model is developed based on the observation that inertia dominated gravity wave predominates in irrigation canal flows. Development of the FDWRM and the method of computation are explained. The FDWRM is tested by applying it to cases of hypothetical flow routing in a wide rectangular channel and also to a real case of flow routing in a field canal. For the cases of hypothetical flow routing in the wide rectangular channel, the FDWRM results match well with those of an implicit numerical model (INM), which solves the dynamic wave model; but the accuracy of the results reduces with increase in backwater effects. For the case of flow routing in the field canal, the FDWRM outputs match well with measured data and also are much better than those of the INM. A fuzzy dynamic flood routing model (FDFRM) is developed based on application of the FDWM to flood routing in natural channels. The fuzzy rule based model is developed based on the observation that monoclinal waves prevail during floods in natural channels. The natural channel reach is discredited into a number of approximately uniform sub-reaches and the fuzzy rule based model for each sub-reach is obtained using the discharge (q)–area (a) relationship at its mean section, based on the kleitz-seddon principle. Development of the FDFRM and the method of computation are explained. The FDFRM is tested by applying it to cases of flood routing in fictitious channels and to flood routing in a natural channel, which is described in the HEC-RAS (hydrologic engineering center – river analysis system) application guide. For the cases of flood routing in the fictitious channels, the FDFRM outputs match well with the INM results. For the case of flood routing in the natural channel, optimized fuzzy rule based models are derived using a neuro-fuzzy algorithm, to take the heterogeneity of the channel sub-reaches into account. The resulting FDFRM outputs are found to be comparable to the HEC-RAS outputs. Also, in literature, the dynamic wave model has been applied in the inverse direction for the development of centralized control algorithms for irrigation canals. In the present study, a centralized control algorithm based on inversion of the fuzzy dynamic wave model (FDWM) is developed to overcome the drawbacks of the existing centralized control algorithms. A fuzzy logic based dynamic wave model inversion algorithm (FDWMIA) is developed for this purpose, based on the inversion of the FDWM. The FDWMIA is tested by applying it to two canal control problems reported in literature: the first problem deals with water level control in a fictitious canal with a single pool and the second, with water level control in a real canal with a series of pools (ASCE Test Canal 2). In both cases, the FDWMIA results are comparable to those of the existing centralized control algorithms.