Defense Notices

Deep learning (DL) has dramatically improved the state-of-the-art performances in broad applications of computer vision, such as image recognition, object detection, semantic/instance segmentation, and point cloud analysis. However, the reasons for such huge empirical success of DL still keep elusive theoretically. In this dissertation, to understand DL and improve its efficiency, robustness, and interpretability, we theoretically investigate optimization algorithms for training deep models and empirically explore deep learning for unsupervised learning tasks in point-cloud analysis and image classification. 

1). Optimization for Training Deep Models: Neural network training is one of the most difficult optimization problems involved in DL. Recently, to understand the global optimality in DL has attracted a lot of attention. However, we observe that conventional DL solvers have not been developed intentionally to seek for such global optimality. In this dissertation, we propose a novel approximation algorithm, BPGrad, towards optimizing deep models globally via branch and pruning. Empirically an efficient solver based on BPGrad for DL is proposed as well, and it outperforms conventional DL solvers such as Adagrad, Adadelta, RMSProp, and Adam in the tasks of object recognition, detection, and segmentation. 

2). Deep Learning for Unsupervised Learning Tasks: The architecture of neural networks is of central importance for many visual recognition tasks. In this dissertation, we focus on the emerging field of unsupervised learning for point clouds analysis and image classification. 

2.1) For point cloud analysis, we propose a novel unsupervised approach to jointly learn the 3D object model and estimate the 6D poses of multiple instances of the same object in a single end-to-end deep neural network framework, with applications to depth-based instance segmentation. Extensive experiments evaluate our technique on several object models and a varying number of instances in 3D point clouds. Compared with popular baselines for instance segmentation, our model not only demonstrates competitive performance, but also learns a 3D object model that is represented as a 3D point cloud. 

2.2) For low-quality image classification, we propose a simple while effective unsupervised deep feature transfer network to address the degrading problem of the state-of-the-art recognition algorithms on low-quality images. No fine-tuning is required in our method. Our network can be embedded into the state-of-the-art deep neural networks as a plug-in feature enhancement module. It preserves data structures in feature space for high-resolution images, and transfers the distinguishing features to low-resolution features space. Extensive experiments show that the proposed transfer network achieves significant improvements over the baseline method. 

In today's world of digital data, the field of data mining has come into the limelight. In data mining, patterns in data are found and accordingly can be analyzed further. Processing data as deep as possible is relevant in case of pattern recognition in huge data sets. In this whole process, we try to understand the data well in order to gain some useful results out of it. For the data to be analyzed correctly, it is better if it is complete and consistent.

We compare the effect of incomplete and inconsistent data in this project. The algorithm used for generating rules is the Modified Learning from Example Module version 2 (MLEM2). We used a single local probabilistic approach for all the datasets. We took 141 datasets into consideration for the error rate comparison of incomplete and inconsistent data. We used ten-fold cross validation and computed average error rate for each of the datasets. From our experiments, we observed that the error rate for incomplete data is greater than the error rate for inconsistent data.

Trends toward large-scale integration and the high-power application of green energy resources necessitate the advent of efficient power converter topologies, multilevel converters. Multilevel inverters are effective solutions for high power and medium voltage DC-to-AC conversion due to their higher efficiency, provision of system redundancy, and generation of near-sinusoidal output voltage waveform. Recently, modular multilevel converter (MMC) has become increasingly attractive. To improve the harmonic profile of the output voltage, there is the need to increase the number of output voltage levels. However, this would require increasing the number of submodules (SMs) and power semi-conductor devices and their associated gate driver and protection circuitry, resulting in the overall multilevel converter to be complex and expensive. Specifically, the need for large number of bulky capacitors in SMs of conventional MMC is seen as a major obstacle. This work proposes an MMC-based multilevel converter that provides the same output voltage as conventional MMC but has reduced number of bulky capacitors. This is achieved by introduction of an extra middle arm to the conventional MMC. Due to similar dynamic equations of the proposed converter with conventional MMC, several previously developed control methods for voltage balancing in the literature for conventional MMCs are applicable to the proposed MMC with minimal effort. Comparative loss analysis of the conventional MMC and the proposed multilevel converter under different power factors and modulation indexes illustrates the lower switching loss of proposed MMC. In addition, a new voltage balancing technique based on carrier-disposition pulse width modulation for modular multilevel converter is proposed.
The second part of this work focuses on an improved control of MMC-based high-power DC/DC converters. Medium-voltage DC (MVDC) and high-voltage DC (HVDC) grids have been the focus of numerous research studies in recent years due to their increasing applications in rapidly growing grid-connected renewable energy systems, such as wind and solar farms. MMC-based DC/DC converters are employed for collecting power from renewable energy sources. Among various developed DC/DC converter topologies, MMC-based DC/DC converter with medium-frequency (MF) transformer is a valuable topology due to its advantages. Specifically, they offer a significant reduction in the size of the MMC arm capacitors along with the ac-link transformer and arm inductors due to the ac-link transformer operating at medium frequencies. As such, this work focuses on improving the control of isolated MMC-based DC/DC (IMMDC) converters. The single phase shift (SPS) control is a popular method in IMMDC converter to control the power transfer. This work proposes conjoined phase shift-amplitude ratio index (PSAR) control that considers amplitude ratio indexes of MMC legs of MF transformer’s secondary side as additional control variables. Compared with SPS control, PSAR control not only provides wider transmission power range and enhances operation flexibility of converter, but also reduces current stress of medium-frequency transformer and power switches of MMCs. An algorithm is developed for simple implementation of the PSAR control to work at the least current stress operating point. Hardware-in-the-loop results confirm the theoretical outcomes of the proposed control method.

To design highly precise 3D object detection approaches for autonomous vehicle has been a crucial topic recently. Shallow machine learning methods such as clustering, support vector machines fail to accomplish multi-modal tasks for self-driving vehicle, while deep-learning based methods gain great success in regressing accurate 3D bound boxes and pose estimation of objects in complicated road scene. Though deep neural networks designed for LiDAR points and monocular-view inputs achieve highest performance in 3D object detection, binocular-views based networks suffer from intrinsic ambiguities therefore yielding less precise regressions. To remedy the ambiguities, we propose an efficient module to bridge the gap between 2D objection detection on stereopsis and real LiDAR points. Experiments on challenging KITTI dataset show that our method outperforms state-of-the-arts binocular-views based methods.

For the past several years, with almost every system being upgraded and digitized, data are getting generated and collected in huge amounts. But there is no use of collecting huge amounts of data unless we can make sense out of it. Generating rules from datasets helps to predict the possible outcomes from given datasets. The predictions are never error free and so all we can do is create rules from datasets that are as accurate as possible. Rule induction from large datasets can be based on different principles and rules induced from applying different methods lead to rulesets with different levels of accuracy. Some rules are more accurate than others. In this project, the goal is to compare two approaches of rule induction, from characteristic sets and from maximal consistent blocks. The aim is to study the error rates of rules generated by taking either of the two approaches. To validate the error rates of the rulesets, 10-fold cross validation method is applied. 

In George Orwell's 1984, there is fear regarding what “Big Brother”, knows due to the fact that even thoughts could be “heard”. Though we are not quite to this point, it should concern us all in what data we are transferring, both intentionally and unintentionally, and whether or not that data is being “leaked”. In this work, we consider the evolving landscape of IoT devices and the threat posed by the pervasive botnets that have been forming over the last several years. We look at two specific cases in this work. One being the practical application of a botnet system actively executing a Man in the Middle Attack against SSH, and the other leveraging the same paradigm as a case of eavesdropping on Internet Protocol (IP) cameras. For the latter case, we construct a web portal for interrogating IP cameras directly for information that they may be exposing. ​

With the rise of software defined radios (SDR) and the trend towards integrating more RF components into MMICs the cost and complexity of multichannel radar development has gone down. High-speed RF data converters have seen continuous increases in both sampling rate and resolution, further rendering a growing subset of components in an RF chain unnecessary. A recent development in this trend is the Xilinx RFSoC, which integrates multiple high speed data converters into the same package as an FPGA. The Center for Remote Sensing of Ice Sheets (CReSIS) is regularly upgrading its suite of sensor platforms spanning from HF depth sounders to Ka band altimeters. A radar platform was developed around the RFSoC to demonstrate the capabilities of the chip when acting as a digital backend and evaluate its role in future radar designs at CReSIS. A new ultra-wideband (UWB) FMCW RF frontend was designed that consists of multiple transmit and receive modules operating at microwave frequencies with multi-GHz bandwidth. An antenna array was constructed out of printed-circuit elements to validate radar system performance. Firmware developed for the RFSoC enables radar features that will prove useful in future sensor platforms used for the remote sensing of snow, soil moisture, or crop canopies.

Though languages like C and C++ are known to be memory unsafe, they are still used widely in industry because of their memory management features, low level nature and performance benefits. Also, as most of the systems software has been written using these languages, replacing them with memory safe languages altogether is currently impossible. Memory safety violations are commonplace, despite the fact that that there have been numerous attempts made to conquer them using source code, compiler and post compilation based approaches. SoftBound is a compiler-based technique that enforces spatial memory safety for C/C++ programs. However, SoftBound needs and depends on program information available in the high-level source code. The goal of our work is to develop a mechanism to efficiently and effectively implement a technique, like SoftBound, to provide spatial memory safety for binary executables. Our approach employs a combination of static-time analysis (using Ghidra) and dynamic-time instrumentation checks (using PIN). Softbound is a pointer based approach, which stores base and bound information per pointer. Our implementation determines the array and pointer access patterns statically using reverse engineering techniques in Ghidra. This static information is used by the Pin dynamic binary instrumentation tool to check the correctness of each load and store instruction at run-time. Our technique works without any source code support and no hardware or compiler alterations are needed. We evaluate the effectiveness, limitations, and performance of our implementation. Our tool detects spatial memory errors in about 57% of the test cases and induces about 6% average overhead over that caused by a minimal pintool.

Observational data produced in scientific experimentation and in day to day life is a valuable source of information for research. It can be challenging to extract meaningful inferences from large amounts of data. Data mining offers many algorithms to draw useful inferences from large pools of information based on observable patterns.

In this project I have implemented one such data mining algorithm for determining a single local probabilistic approximation, which also computes the corresponding ruleset; and compared it with two versions of the MLEM2 algorithm which induce a certain rule set and a possible rule set respectively. For experimentation, eight data sets with 35% missing values were used to induce corresponding rulesets and classify unseen cases. Two different interpretations of missing values were used, namely, lost values and do not care conditions. k-fold cross validation technique was employed with k=10 to identify error rates in classification. 

The goal of this project was to compare how accurately unseen cases are classified by the rulesets induced by each of the aforementioned algorithms. Error rate calculated from the k-fold cross validation technique was also used to observe how each type of interpretation of missing values affects the ruleset.

Time and again, fiber optic medium has proved to be the best means for transporting global data traffic which is following an exponential growth trajectory. Rapid development of high bandwidth applications since the past decade based on virtual reality, 5G and big data to name a few have resulted in a sudden surge of research activities across the globe to maximize effective utilization of available fiber bandwidth which until then was supporting low speed services like voice and low bandwidth data traffic. To this end, higher order modulation formats together with multi-subcarrier superchannel based fiber optic transmission systems have proved to enhance spectral efficiency and achieve multi terabit per second data rates. However, spectrally efficient systems are extremely sensitive to transmission impairments stemming from both optical devices and fiber itself. Therefore, such systems mandate the use of robust digital signal processing (DSP) to compensate and/or mitigate the undesired artifacts, thereby extending the transmission reach. The central theme of this dissertation is to propose and validate few efficient DSP techniques to compensate specific impairments as delineated in the next three paragraphs.
For short reach applications, we experimentally demonstrate a digital compensation technique to undo semiconductor optical amplifier (SOA) and photodiode nonlinearity effects by digitally backpropagating the received signal through a virtual SOA with inverse gain characteristics followed by an iterative algorithm to cancel signal-signal beat interference arising from photodiode. We characterize the phase dynamics of comb lines from a quantum dot passive mode locked laser based on a novel multiheterodyne coherent detection technique. In the context of multi-subcarrier, Nyquist pulse shaped, superchannel transmission system with coherent detection, we demonstrate through measurements and numerical simulations an efficient phase noise compensation technique called “Digital Mixing” that operates using a shared pilot tone exploiting the mutual phase coherence among the comb lines.
Finally, we propose and experimentally validate a practical pilot aided relative phase noise compensation technique for forward pumped distributed Raman amplified, digital subcarrier multiplexed coherent transmission systems.