Defense Notices

The advent of advanced natural language processing (NLP) techniques has revolutionized the way we handle textual data. This project presents the implementation of exploring contextual sentence understanding on the Quora Insincere Questions dataset using the pretrained BERT architecture. In this study, we explore the application of BERT, a bidirectional transformer model, for text classification tasks. The goal is to classify if a question contains hateful, disrespectful or toxic content. BERT represents the state-of-the-art in language representation models and has shown strong performance on various natural language processing tasks. In this project, the pretrained BERT base model is fine-tuned on a sample of the Quora dataset for next sentence prediction. Results show that with just 1% of the data (around 13,000 examples), the fine-tuned model achieves over 90% validation accuracy in identifying insincere questions after 4 epochs of training. This demonstrates the effectiveness of leveraging BERT for text classification tasks with minimal labeled data requirements. Being able to automatically detect toxic, hateful or disrespectful content is important to maintain healthy online discussions. However, the nuances of human language make this a challenging natural language processing problem. Insincere questions may contain offensive language, hate speech, or misinformation, making their identification crucial for maintaining a positive and safe online environment. In this project, we explore using the pretrained Bidirectional Encoder Representations from Transformers (BERT) model for next sentence prediction on the task of identifying insincere questions.

The research investigates the utilization of Synthetic Minority Oversampling Techniques (SMOTE) in the context of machine learning models applied to biomedical datasets, particularly focusing on mitigating demographic data disparities. The study is most relevant to underrepresented demographic data. The primary objective is to enhance the SMOTE methodology, traditionally designed for addressing class imbalances, to specifically tackle ethnic imbalances within feature representation. In contrast to conventional approaches that merely exclude race as a fundamental or additive factor without rectifying misrepresentation, this work advocates an innovative modification of the original SMOTE framework, emphasizing dataset augmentation based on participants' demographic backgrounds. The predominant aim of the project is to enhance and reshape the distribution to optimize model performance for unspecified demographic subgroups during training. However, the outcomes indicate that despite the application of feature balancing in this adapted SMOTE method, no statistically significant enhancement in accuracy was discerned. This observation implies that while rectifying imbalances is crucial, it may not independently suffice to overcome challenges associated with heterogeneity in species representation within machine learning models applied to biomedical databases. Consequently, further research endeavors are necessary to identify novel methodologies aimed at enhancing sampling accuracy and fairness within diverse populations.

Quantum computing presents an opportunity to accelerate machine learning (ML) tasks on quantum processors in a similar vein to existing classical accelerators, such as graphical processing units (GPUs). In the classical domain, convolutional neural networks (CNNs) effectively exploit data locality using the convolution operation to reduce the number of fully-connected operations in multi-layer perceptrons (MLPs). Preserving data locality enables the pruning of training parameters, which results in reduced memory requirements and shorter training time without compromising classification accuracy. However, contemporary quantum machine learning (QML) algorithms do not leverage the data locality of input features in classification workloads, particularly for multidimensional data. This work presents a multidimensional quantum convolutional classifier (MQCC) that adapts the CNN structure to a variational quantum algorithm (VQA). The proposed MQCC uses quantum implementations of multidimensional convolution, pooling based on the quantum Haar transform (QHT) and partial measurement, and fully-connected operations. Time-complexity analysis will be presented to demonstrate the speedup of the proposed techniques in comparison to classical convolution and pooling operations on modern CPUs and/or GPUs. Experimental work is conducted on state-of-the-art quantum simulators from IBM Quantum and Xanadu modeling noise-free and noisy quantum devices. High-resolution multidimensional images are used to demonstrate the correctness and scalability of the convolution and pooling operations. Furthermore, the proposed MQCC model is tested on a variety of common datasets against multiple configurations of related ML and QML techniques. Based on standard metrics such as log loss, classification accuracy, number of training parameters, circuit depth, and gate count, it will be shown that MQCC can deliver a faithful implementation of CNNs on quantum machines. Additionally, it will be shown that by exploiting data locality MQCC can achieve improved classification over contemporary QML methods. 

Software vulnerabilities, stemming from coding weaknesses and poor development practices, have become increasingly prevalent. These vulnerabilities could be exploited by attackers to pose risks to the confidentiality, integrity, and availability of software. To protect themselves, end-users of software may have an interest in knowing if the software they buy and use is secure from such attacks. Our work is motivated by this need to automatically assess and rate the security properties of binary software.

To increase user trust in third-party software, researchers have devised several techniques and tools to identify and mitigate coding weaknesses in binary software. Therefore, our first task in this work is to assess the current landscape and comprehend the capabilities and challenges faced by binary-level techniques aimed at detecting critical coding weaknesses in software binaries. We categorize the most important coding weaknesses in compiled programming languages, and conduct a comprehensive survey, exploration, and comparison of static techniques designed to locate these weaknesses in software binaries. Furthermore, we perform an independent assessments of the efficacy of open-source tools using standard benchmarks.

Next, we develop techniques to assess if secure coding principles were adopted during the generation of the software binary. Towards this goal, we first develop techniques to determine the high-level source language used to produce the binary. Then, we check the feasibility of detecting the use of secure coding best practices during code development. Finally, we check the feasibility of detecting the vulnerable regions of code in any binary executable. Our ultimate future goal is to employ all of our developed techniques to rate the security-quality of the given binary software.

With threat actors ever evolving, the need for secure communications continues to grow. By using non-traditional means as a way of a communication network, it is possible to securely communicate within a region using the bluetooth mesh protocol. The goal is to automatically place these mesh devices in a defined region in order to ensure the integrity and reliability of the network, while also ensuring the least number of devices are placed. By placing a provisioner node, the rest of the specified region populates with mesh nodes that act as relays, creating a network allowing users to communicate within. By utilizing Dijkstra’s algorithm, it is possible to calculate the Time to Live (TTL) between two given nodes in the network, which is an important metric as it directly affects how far apart two users can be within the region. When placing the nodes, a range for the nodes being used is specified and accounted for, which impacts the number of nodes needed within the region. Results show that when nodes are placed at coordinate points given by the generated map, users are able to communicate effectively across the specified region. In this project, a web interface is created in order to allow a user to specify the TTL, range, and the number of nodes to use, and proceeds to place each device within the region drawn by the user.

Large Language Models (LLMs) have become integral to natural language processing, involving initial broad pretraining on generic data followed by fine-tuning for specific tasks or domains. While advancements in Parameter Efficient Fine-Tuning (PEFT) techniques have made strides in reducing resource demands for LLM fine-tuning, they possess individual constraints. This project addresses the challenges posed by PEFT in the context of transformers architecture for sequence-to-sequence tasks, by integrating two pivotal techniques: Low-Rank Adaptation (LoRA) for computational efficiency and adaptive layers for task-specific customization. To overcome the limitations of LoRA, we introduce a simple yet effective hyper alignment adapter, that leverages a hypernetwork to generate decoder inputs based on encoder outputs, thereby serving as a crucial bridge to improve alignment between the encoder and the decoder. This fusion strikes a balance between the fine-tuning complexity and task performance, mitigating the individual drawbacks while improving the encoder-decoder alignment. As a result, we achieve more precise and contextually relevant sequence generation. The proposed solution improves the overall efficiency and effectiveness of LLMs in sequence-to-sequence tasks, leading to better alignment and more accurate output generation.

Having control over nano- and micro-sized objects' position inside a suspension is crucial in many applications such as: trapping and manipulating microscopic objects, sorting particles and living microorganisms, and building microscopic size 3D crystal structures and lattices. This control can be achieved by judiciously engineering optical forces and light-matter interactions inside colloidal suspensions that result in optical trapping. However, in the current techniques, to confine and transport particles in 3D, the use of high NA (Numerical Aperture) optics is a must. This in turn leads to several disadvantages such as alignment complications, narrow field of view, low stability values, and undesirable thermal effects. Hence, here we study a novel optical trapping method that we named asymmetric counter-propagating beams where optical forces are engineered to overcome the aforementioned limitations of existing methods. This novel system is significantly easier to align due to its utilization of much lower NA optics in combination with engineered beams which create a very flexible manipulating system. This new approach allows the trapping and manipulation of different shape objects, sizing from tens of nanometers to hundreds of micrometers by exploiting asymmetrical optical fields with high stability. In addition, this technique also allows for significantly larger particle trapping volumes. As a result, we can apply this method to trapping much larger particles and microorganisms that have never been trapped optically before as well as building 3D lattices and crystal structures of microscopic-sized particles. Finally, this novel approach allows for the integration of a variety of spectroscopy and microscopy techniques, such as light-sheet fluorescence microscopy, to extract time-sensitive information and acquire images with detailed features from trapped entities.

We describe an autonomous driving system that uses reinforcement learning to train a car to drive without the need for collecting training input from human drivers.  We achieve this by using the Advantage Actor Critic reinforcement system that trains the car based on continuously adapting the model to minimize the penalty received by the car.  This penalty is determined if the car intersected the borders of the track on which it is driving.  We show the resilience of the proposed autonomously trained system to noisy sensor inputs and variations in the shape of the track.

The Center for Oldest Ice Exploration (COLDEX) is an NSF-funded multi-institution collaboration to explore Antarctica for the oldest possible continuous ice record. It comprises of exploration and modelling teams that are using instruments like radars, lidars, gravimeters, and magnetometers to select candidate locations to collect a continuous 1.5-million-year ice core. To assist in this search for old ice, the Center for Remote Sensing and Integrated Systems (CReSIS) at the University of Kansas developed a new airborne higher-power version of the 600-900 MHz Accumulation Radar with a much larger multichannel cross-track antenna array. The fuselage portion of the antenna array is a 64-element 0.9 m by 3.8 m array with 4 elements in along-track and 16 elements in cross-track. Each element is a dual-polarized microstrip antenna and each column of 4 elements is power combined into a single channel resulting in 16 cross-track channels. Power is transmitted across 4 cross-track channels on either side of the fuselage array alternatingly to produce a total peak power of 6.4 kW (before losses). Three additional antennas are integrated on each wing to lengthen the antenna aperture. A novel receiver concept is developed using limiters to compress the dynamic range to simultaneously capture the strong ice surface and weak ice bottom returns. This system was flown on a Basler aircraft at the South Pole during the 2022-2023 Austral Summer season and will be flown again during the upcoming 2023-2024 season for repeat interferometry. This work describes the current radar system design and proposes to develop improvements to the compact, high-power divider and large multichannel polarimetric array used by the radar. It then proposes to develop and implement a system engineering perspective on the calibration of this multi-pass imaging radar.

Even for a fixed time-bandwidth product there are infinite possible spectrally-shaped random FM (RFM) waveforms one could generate due to their being phase-continuous. Moreover, certain RFM classes rely on an imposed basis-like structure scaled by underlying parameters that can be optimized (e.g. gradient descent and greedy search have been demonstrated). Because these structures must include oversampling with respect to 3-dB bandwidth to account for sufficient spectral roll-off (necessary to be physically realizable in hardware), they are not true bases (i.e. not square). Therefore, any individual structure cannot represent all possible waveforms, with the waveforms generated by a given structure tending to possess similar attributes. Unless of course we consider over-coded polyphaser-coded FM (PCFM), which increases the number of elements in the parameter vector, while maintaining the relationship between waveform samples and the time-bandwidth product. Which presents the potential for a true bases, if there is a constraint either explicit or implicit that will constrain the spectrum. Here we examine waveforms possessing different attributes, as well as the potential for a true basis which may inform their selection for given radar applications.