The high demand for addressing the required processing power of today's big-data and compute-intensive applications and the cost of powerful processing units have led end-devices to move most of their data processing to clouds, edges, and datacenters. In clouds, edges, and datacenters, hardware accelerators are exploited to drastically augment the processing power of multi-core processors. Among hardware accelerators, Field Programmable Gate Arrays (FPGAs) have attracted significant attention due to their prominent performance and power efficiency. There are two main challenges in exploiting FPGAs as hardware platforms for acceleration: how to design efficient accelerators to gain high performance?, and how to efficiently enable software applications to access FPGA accelerators? In this dissertation, my focus is on the second challenge, where I show the importance of FPGA-based acceleration management on the total gained performance.Unlike Graphics Processing Units (GPUs), FPGAs offer a heterogeneous environment for different types of accelerators. This unique feature allows FPGAs to serve various applications concurrently and makes them suitable to address the demands from diverse applications in clouds, edges, and datacenters. However, due to the limited number of resources and time-consuming process of reconfiguration, the importance of sharing FPGA accelerators among different applications arises. Current state-of-the-art accelerator management schemes do not yet provide accelerator sharing among multiple applications concurrently. To achieve this goal, we need system software support as well as a hardware controller to enable seamless accelerator sharing. In this dissertation, I propose a scalable framework, called UltraShare Express, for the ultimate concurrent sharing of different accelerators among various applications. UltraShare Express provides software like function calls for the seamless FPGA acceleration virtualization. Unlike previous works that exploit a static accelerator allocation at the design-time, UltraShare Express offers a run-time dynamic accelerator allocation that enables maximum utilization of the available resources and accelerators on FPGAs. In Chapter 1, I discuss the essential reasons behind huge attention toward FPGA acceleration in clouds, edges, and datacenters. I also address the challenges of using FPGAs as hardware acceleration platforms for general applications. In Chapter 2, I present the background and history of using FPGAs as acceleration platforms. I also discuss the advantages and disadvantages of previous works in deploying FPGAs for the acceleration. In Chapter 3, I focus on the system software to address the conflicts that happen among multiple applications requesting FPGA accelerators simultaneously. In this chapter, I present our proposed single-command-based framework, called MQMAI, that uses a multi-queue architecture in the software-stack to minimize conflicts among different applications to access different FPGA accelerators. In Chapter 4, I focus on the hardware accelerator controller and specifically address the interleaved/concurrent sharing of multiple accelerators among multiple user applications. In this chapter, I propose our hardware controller, called UltraShare, that enables the ultimate sharing of FPGA accelerators among multiple applications. The proposed hardware controller introduces a dynamic accelerator sharing scheme through an accelerator grouping mechanism. In Chapter 5, I propose UltraShare Express, our full-fledged framework for ultimate sharing of several FPGA accelerators among multiple applications. UltraShare Express inherits the advantages of both MQMAI and UltraShare by combining them. In Chapter 5, I further investigate the opportunities for addressing and eliminating FPGA accelerators' stall times in different scenarios. Focusing on the software stack, I propose a novel multi-queue architecture in the software-stack to avoid possible command blocking scenarios that can significantly degrade the performance of FPGA acceleration when accelerators are shared. I also propose a mechanism to fairly distribute data-link bandwidth among different accelerators concerning the accelerator grouping mechanism proposed in Chapter 4. This mechanism prevents accelerators with higher throughput be sacrificed in accessing data-link bandwidth. Our Experimental results on various accelerator IP cores show significant improvements when the ultimate sharing of FPGA accelerators is enabled. I believe that UltraShare Express provides an important step toward the efficient and easy deploying of FPGAs in heterogeneous architectures. UltraShare Express is an open-source framework available on GitHub and can be used by other research groups.

Unsteady aerodynamics emerges in several applications in mechanical and aerospace engineering such as the flutter problem, dynamic stall, bio-inspired flying robots, helicopters, and wind turbines. If a leap were to happen in the design of these machines, it would have to happen in the preliminary design phase where millions of design alternatives are investigated, which cannot and should not be performed with high-fidelity simulations. Therefore, the development of reduced-order models for unsteady aerodynamic loads will be of paramount importance for advancing these applications, which is the main focus of this Dissertation.

The common (perhaps only) reduced-order models of unsteady aerodynamics in literature are based on potential flow theory, which is neither complete (invokes a closure condition) nor descriptive of viscous effects. While the Kutta condition is quite successful in small-angle of attack and large Reynolds number situations, it was originally devised for steady flows. Hence, its illegitimate use in unsteady environments may only be successful at low-frequencies. Therefore, there are numerous research reports that invoked another closure condition for potential flow in the unsteady case.

Three main contributions are achieved in this Dissertation. First, realizing that the lift development and vorticity production are essential viscous processes, we develop a viscous extension of the classical theory of unsteady aerodynamics by matching the potential flow theory with a special boundary layer theory that pays close attention to the flow details in the vicinity of the trailing edge: the triple deck boundary layer theory. Based on this extension, we develop for the first time a Reynolds-number-dependent lift frequency response. The theory is validated against high-fidelity simulations of the Navier-Stokes equations showing a remarkable matching. It is found that viscosity induces more phase lag at high reduced frequencies and low Reynolds numbers where the viscous effects are more pronounced.

Second, to extend the applicability of this viscous unsteady theory to account for arbitrary airfoil shapes, arbitrary kinematics, and wake deformation, a numerical method is developed. The developed numerical technique represents a viscous extension of the classical unsteady vortex lattice method (UVLM). In such an extension, we exploit Hilbert matrix algebra to show the relationship between the Kutta condition and the location of the control and collocation points on the panel. As such, we relax (correct for) the implicit Kutta condition in the UVLM by updating the locations of these points within the panel at each time based on the viscous correction coming from the triple deck boundary layer theory, which is based on the instantaneous airfoil motion.

Finally, due to the fact that many aerial vehicles operate at moderate Reynolds number, which is prone to laminar-to-turbulent transition and theoretical approaches are not expected to perform efficiently in those conditions, the transition effects on the lift and circulation dynamics on a pitching airfoil are investigated numerically. It is found that transition induces significant nonlinearities in the lift and circulation dynamics even at very small amplitudes down to half a degree. It is shown that this nonlinearity in the lift dynamics is attributed to the violation of the Ku0tta condition in this regime. Based on this connection, we delineate how the potential flow theory can be extended with the aid of high-fidelity simulation data to capture the nonlinear transition effects.

Recently, deep learning models have been extensively adopted in numerous applications, from health care to finance and entertainment industry. This wide-spread deployment of deep models raised concern over the privacy of data used to train deep models. This is a huge concern particularly for data-sensitive applications, such as health records, personal data, bio-metric data, etc. As a result, a new direction of research focusing on possible attacks aiming to identify training data of deep models emerged, called membership inference.Membership inference (MI) attacks identify which samples have been used during training and which samples have not. The first generation of membership inference attacks mainly used deep models' prediction confidence as a feature to identify training samples. The intuition is that deep models are more confident on samples they have seen during training than non-training samples. Despite their sound intuition and apparent successful reports, we, along a few other parallel studies, showed that the first generation of membership inference attacks are ineffective in practice for multiple reasons. First, they could not significantly outperform a naive baseline that labels a sample as a member (training sample) if it is correctly classified by the deep model and as a non-member (non-training sample) otherwise. Second, the confidence distribution of correctly classified samples, which cover the majority of a dataset, are not distinguishable between train and non-train samples. Only a small portion of mis-classified samples exhibit discrepant distribution. Third, all these membership inference attacks report average-case success metric (e.g., accuracy or ROC-AUC). However, privacy is not an average case-metric, and it should be treated similar to other security and privacy related problems. Similar to other security problems, the attack is reliable if it can identify a few training samples while almost on non-training samples are falsely labeled as a training sample. In other words, a reliable membership inference attack should have a decent true-positive rate (TPR) at low false-positive rates (FPR). In this dissertation, we aim to move the membership inference research in a more practical direction, either by showing the limitations of the current attacks or by proposing more reliable attacks. As stated earlier, we first show that the current generation of membership inference attacks are not reliable in practice. Then, we propose several new membership inference attacks that achieve more reliable performance in more realistic scenarios. The first attack focuses on the model's behavior in the entire sub-population, instead of a single sample in vacuum. More specifically, we compare the model's confidence on a target sample and other samples from the same sub-population. If the confidence of a sample is significantly higher than the average confidence on that sub-population, that is an indication of a training sample. We show that this attack can achieve moderate true positive with very low false positive. Additionally, we propose a BiGAN architecture to generate samples from the same sub-population, in case it is not available. The second attack aims to focus on user-level MI attack instead of the record-level MI attack. In this scenario, we identify if a user's data has been used during training instead if identifying which samples from the user have been used. Not only this attack is more realistic in privacy domain, but we show that we can achieve the state-of-the-art accuracy if multiple samples from a user are used to draw the membership inference. In another study, we show that MI attacks are generally more successful when deep ensemble is used. We show that deep ensemble shifts the distribution of train and non-train samples in a different way where they become significantly more distinguishable. Finally, we show that are a few simple aggregation mechanisms instead of ensemble averaging that can improve the accuracy and privacy of deep models in deep ensemble context. Finally, we illustrate a fundamental issue with current MI attacks, including the state-of-the-art attacks, that limits their applications in certain scenarios. We elaborate the issues with a practical scenario where membership inference attacks are used by an auditor (investigator) to prove to a judge/jury that the auditee unlawfully used sensitive data during training. Although the current SOTA attacks can identify some training samples with low false positive ratio in a common experimental setting extensively used for MI attacks, an auditee can generate unlimited number of samples on which MI attacks catastrophically fail. This can be used in court to easily discredit the allegation of the auditor and make the case dismissed. Interestingly, we show that auditee does not need to know anything about the auditor's membership inference attack to generate those challenging samples. We called this problem, discredibility. Currently, there is no attack immune to discredibility. We hope that our research sheds light on this newly-discovered issue and encourage researchers to investigate it.

We present a position estimation scheme for cars based on the integration of Global Positioning System (GPS) with vehicle sensors. The aim is to achieve enough accuracy to enable in vehicle cooperative collision warning, i.e., systems that provides warnings to drivers based on information about the motions of neighboring vehicles obtained by wireless communications from those vehicles, without use of ranging sensors. The vehicle sensors consist of wheel speed sensors, steering angle encoder, and a fiber optic gyro. We fuse these in an extended Kalman filter. The process model is a dynamic bicycle model. We present data from about 60 km of driving in urban environments including stops, intersection turns, U-turns, and lane changes, at both low and high speeds. The data show the filter estimates position, speed, and heading with the accuracies required by cooperative collision warning in all except two kinds of settings. The data also shows GPS and vehicle sensor integration through a bicycle model compares favorably with position estimation by fusing GPS and inertial navigation system (INS) through a kinematic model.

In chapter 1, I study a dynamic innovation race, in which firms invest to be the first toattain a breakthrough invention. I examine how delays in monitoring firms' investments affect their ability to achieve a coordinated under-investment equilibrium enforced by the threat of elevated investment by rival firms. When monitoring delays are small, equilibrium investment can be considerably delayed, matching the first-best solution under cooperation. Even with significant delays in monitoring, equilibrium investment is below, and firm values are well above, the no-monitoring competitive outcome. As a result, the regulatory goal of transparency can conflict with the goal of encouraging investment in innovation. In addition I study how changing the number of firms in the race can affect the outcome. As the number of firms increases, the incentive to preempt escalates, and coordinated effort boundaries decline, as does the maximum delay time in which the first best can be achieved with coordination.

In chapter 2, I study the contracting problem between investors (limited partners or LPs) and venture capitalists ( general partners or GPs) . In real world, GPs are sometimes paid on a deal-by-deal basis and other times on a whole-portfolio basis. When is one method of payment better than the other? I develop a model to see how the method of compensation's payment can affect the behavior of general partners (GP) in a limited partnership agreement (LPA). I show that when assets (projects or firms) are highly correlated or when GPs have low reputation, whole-portfolio contracting is superior to deal-by-deal contracting. In this case, by bundling payouts together, whole- portfolio contracting enhances incentives for GPs to exert effort. Therefore, it is better suited to alleviate the moral hazard problem which is stronger than the adverse selection problem in the case of high correlation of assets or low reputation of GPs. In contrast, for low correlation of assets or high reputation of GPs, information asymmetry concerns dominate and deal-by-deal contracts become optimal, as they can efficiently weed out bad projects one by one. These results shed light on recent empirical findings on the relationship between investors and venture capitalists.

To accommodate the growing air traffic demand, flights will need to be planned and navigated with a much higher level of precision than today's aircraft flight path. The Next Generation Air Transportation System (NextGen) stands to benefit significantly in safety and efficiency from such movement of aircraft along precisely defined paths. Air Traffic Operations (ATO) relying on such precision--the Precision Air Traffic Operations or PATO--are the foundation of high throughput capacity envisioned for the future airports. In PATO, the preferred method is to manage the air traffic by assigning a speed profile to each aircraft in a given fleet in a given airspace (in practice known as (speed control). In this research, an algorithm has been developed, set in the context of a Hybrid Control System (HCS) model, that determines whether a speed control solution exists for a given fleet of aircraft in a given airspace and if so, computes this solution as a collective speed profile that assures separation if executed without deviation. Uncertainties such as weather are not considered but the algorithm can be modified to include uncertainties. The algorithm first computes all feasible sequences (i.e., all sequences that allow the given fleet of aircraft to reach destinations without violating the FAA's separation requirement) by looking at all pairs of aircraft. Then, the most likely sequence is determined and the speed control solution is constructed by a backward trajectory generation, starting with the aircraft last out and proceeds to the first out. This computation can be done for different sequences in parallel which helps to reduce the computation time. If such a solution does not exist, then the algorithm calculates a minimal path modification (known as path control) that will allow separation-compliance speed control. We will also prove that the algorithm will modify the path without creating a new separation violation. The new path will be generated by adding new waypoints in the airspace. As a byproduct, instead of minimal path modification, one can use the aircraft arrival time schedule to generate the sequence in which the aircraft reach their destinations.