Defense Notices

With the exponential growth of user-generated content, online platforms face unprecedented challenges in moderating toxic and harmful comments. Due to this, Automated content moderation has emerged as a critical application of machine learning, enabling platforms to ensure user safety and maintain community standards. Despite its importance, challenges such as severe class imbalance, contextual ambiguity, and the diverse nature of toxic language often compromise moderation accuracy, leading to biased classification performance.

This project presents a comparative analysis of machine learning approaches for a Multi-Label Toxic Comment Classification System using the Toxic Comment Classification dataset from Kaggle.  The study examines the performance of traditional algorithms, such as Logistic Regression, Random Forest, and XGBoost, alongside deep architectures, including Bi-LSTM, CNN-Bi-LSTM, and DistilBERT. The proposed approach utilizes word-level embeddings across all models and examines the effects of architectural enhancements, hyperparameter optimization, and advanced training strategies on model robustness and predictive accuracy.

The study emphasizes the significance of loss function optimization and threshold adjustment strategies in improving the detection of minority classes. The comparative results reveal distinct performance trade-offs across model architectures, with transformer models achieving superior contextual understanding at the cost of computational complexity. At the same time, deep learning approaches(LSTM models) offer efficiency advantages. These findings establish evidence-based guidelines for model selection in real-world content moderation systems, striking a balance between accuracy requirements and operational constraints.

Quantum computing has the potential to revolutionize computational problem-solving by leveraging the quantum mechanical phenomena of superposition and entanglement, which allows for processing a large amount of information simultaneously. This capability is significant in the numerical solution of complex and/or multidimensional partial differential equations (PDEs), which are fundamental to modeling various physical phenomena. There are currently many quantum techniques available for solving partial differential equations (PDEs), which are mainly based on variational quantum circuits. However, the existing quantum PDE solvers, particularly those based on variational quantum eigensolver (VQE) techniques, suffer from several limitations. These include low accuracy, high execution times, and low scalability on quantum simulators as well as on noisy intermediate-scale quantum (NISQ) devices, especially for multidimensional PDEs.

 In this work, we propose an efficient and scalable algorithm for solving multidimensional PDEs. We present two variants of our algorithm: the first leverages finite-difference method (FDM), classical-to-quantum (C2Q) encoding, and numerical instantiation, while the second employs FDM, C2Q, and column-by-column decomposition (CCD). Both variants are designed to enhance accuracy and scalability while reducing execution times. We have validated and evaluated our proposed concepts using a number of case studies including multidimensional Poisson equation, multidimensional heat equation, Black Scholes equation, and Navier-Stokes equation for computational fluid dynamics (CFD) achieving promising results. Our results demonstrate higher accuracy, higher scalability, and faster execution times compared to VQE-based solvers on noise-free and noisy quantum simulators from IBM. Additionally, we validated our approach on hardware emulators and actual quantum hardware, employing noise mitigation techniques. This work establishes a practical and effective approach for solving PDEs using quantum computing for engineering and scientific applications.

The escalating complexity inherent in modern computer architecture presents significant challenges for both professional hardware designers and students striving to gain foundational understanding. Historically, the steady improvement of computer systems was driven by transistor scaling, predictable performance increases, and relatively straightforward architectural paradigms. However, with the end of traditional scaling laws and the rise of heterogeneous and parallel architectures, designers now face unprecedented intricacies involving power management, thermal constraints, security considerations, and sophisticated software interactions. Prior tools and methodologies, often reliant on complex, command-line driven simulations, exacerbate these challenges by introducing steep learning curves, creating a critical need for more intuitive, accessible, and efficient solutions. To address these challenges, this thesis introduces two innovative, modern tools.

The first tool, SimScholar, provides an intuitive graphical user interface (GUI) built upon the widely-used gem5 simulator. SimScholar significantly simplifies the simulation process, enabling students and educators to more effectively engage with architectural concepts through a visually guided environment, both reducing complexity and enhancing conceptual understanding. Supporting SimScholar, the gem5 Extended Modules API (gEMA) offers streamlined backend integration with gem5, ensuring efficient communication, modularity, and maintainability.

The second contribution, gem5 Co-Pilot, delivers an advanced framework for architectural design space exploration (DSE). Co-Pilot integrates cycle-accurate simulation via gem5, detailed power and area modeling through McPAT, and intelligent optimization assisted by a large language model (LLM). Central to Co-Pilot is the Design Space Declarative Language (DSDL), a Python-based domain-specific language that facilitates structured, clear specification of design parameters and constraints.

Collectively, these tools constitute a comprehensive approach to taming complexity in computer architecture, offering powerful, user-friendly solutions tailored to both educational and professional settings.

The rapid development of mobile apps on modern smartphone platforms has raised critical concerns regarding user data privacy and the security of app-to-device communications, particularly with companion apps that interface with external IoT or cyber-physical systems (CPS). In this dissertation, we investigate two major aspects of mobile app security: the misuse of permission mechanisms and the security of app to device communication in automotive companion apps.

Mobile apps seek user consent for accessing sensitive information such as location and personal data. However, users often blindly accept these permission requests, allowing apps to abuse this mechanism. As long as a permission is requested, state-of-the-art security mechanisms typically treat it as legitimate. This raises a critical question: Are these permission requests always valid? To explore this, we validate permission requests using statistical analysis on permission sets extracted from groups of functionally similar apps. We identify mobile apps with abusive permission access and quantify the risk of information leakage posed by each app. Through a large-scale statistical analysis of permission sets from over 200,000 Android apps, our findings reveal that approximately 10% of the apps exhibit highly risky permission usage. 

Next, we present a comprehensive study of automotive companion apps, a rapidly growing yet underexplored category of mobile apps. These apps are used for vehicle diagnostics, telemetry, and remote control, and they often interface with in-vehicle networks via OBD-II dongles, exposing users to significant privacy and security risks. Using a hybrid methodology that combines static code analysis, dynamic runtime inspection, and network traffic monitoring, we analyze 154 publicly available Android automotive apps. Our findings uncover a broad range of critical vulnerabilities. Over 74% of the analyzed apps exhibit vulnerabilities that could lead to private information leakage, property theft, or even real-time safety risks while driving. Specifically, 18 apps were found to connect to open OBD-II dongles without requiring any authentication, accept arbitrary CAN bus commands from potentially malicious users, and transmit those commands to the vehicle without validation. 16 apps were found to store driving logs in external storage, enabling attackers to reconstruct trip histories and driving patterns. We demonstrate several real-world attack scenarios that illustrate how insecure data storage and communication practices can compromise user privacy and vehicular safety. Finally, we discuss mitigation strategies and detail the responsible disclosure process undertaken with the affected developers.

Nonlinear Transmission Lines (NLTLs) have gained significant interest due to their ability to generate ultra-short, high-power RF pulses, which are valuable in applications such as ultrawideband radar, space vehicles, and battlefield communication disruption. The waveforms generated by NLTLs offer frequency diversity not typically observed in High-Power Microwave (HPM) sources based on electron beams. Nonlinearity in lumped element transmission lines is usually introduced using voltage-dependent capacitors due to their simplicity and widespread availability. The periodic structure of these lines introduces dispersion, which broadens pulses. In contrast, nonlinearity causes higher-amplitude regions to propagate faster. The interaction of these effects results in the formation of stable, self-localized waveforms known as solitons.
Soliton propagation in NLTLs can be described by the Korteweg-de Vries (KdV) equation. In this thesis, the Bäcklund Transformation (BT) method has been used to derive both single and two-soliton solutions of the KdV equation. This method links two different partial differential equations (PDEs) and their solutions to produce solutions for nonlinear PDEs. The two-soliton solution is obtained from the single soliton solution using a nonlinear superposition principle known as Bianchi’s Permutability Theorem (BPT). Although the KdV model is suitable for NLTLs where the capacitance-voltage relationship follows that of a reverse-biased p-n junction, it cannot generally represent arbitrary nonlinear capacitance characteristics.
To address this limitation, a Finite Difference Time Domain (FDTD) method has been developed to numerically solve the NLTL equation for soliton propagation. To demonstrate the pulse sharpening and RF generation capability of a varactor-loaded NLTL, a 12-section lumped element circuit has been designed and simulated using LTspice and verified with the calculated result. In airborne radar systems, operational constraints such as range, accuracy, data rate, environment, and target type require flexible waveform design, including variation in pulse widths and pulse
repetition frequencies. A gradient descent optimization technique has been employed to generate pulses with varying amplitudes and frequencies by optimizing the NLTL parameters. This work provides a theoretical analysis and numerical simulation to study soliton propagation in NLTLs and demonstrates the generation of tunable RF pulses through optimized circuit design.

In recent years, the intersection of Artificial Intelligence and healthcare has paved the way for intelligent dietary assistance. NutriBot is an AI-powered chatbot developed using the Rasa framework to deliver personalized nutrition recommendations based on user preferences, diet types, and nutritional goals. This full-stack system integrates Rasa NLU, a Flask backend, the Nutritionix API for real-time food data, and a React.js + Tailwind CSS frontend for seamless interaction. The system is containerized using Docker and deployable on cloud platforms like GCP.

The chatbot supports multi-turn conversations, slot-filling, and remembers user preferences such as dietary restrictions or nutrient focus (e.g., high protein). Evaluation of the system showed perfect intent and entity recognition accuracy, fast API response times, and user-friendly fallback handling. While NutriBot currently lacks persistent user profiles and multilingual support, it offers a highly accurate, scalable framework for future extensions such as fitness tracker integration, multilingual capabilities, and smart assistant deployment.

Accurate classification of brain tumors from MRI scans plays a vital role in assisting clinical diagnosis and treatment planning. This project investigates and compares three deep learning-based classification approaches designed to evaluate the effectiveness of integrating recurrent layers into conventional convolutional architectures. Specifically, a CNN-LSTM model, a CNN-RNN model with GRU units, and a baseline CNN classifier using EfficientNetB0 are developed and assessed on a curated MRI dataset.

The CNN-LSTM model uses ResNet50 as a feature extractor, with spatial features reshaped and passed through stacked LSTM layers to explore sequential learning on static medical images. The CNN-RNN model implements TimeDistributed convolutional layers followed by GRUs, examining the potential benefits of GRU-based modeling. The EfficientNetB0-based CNN model, trained end-to-end without recurrent components, serves as the performance baseline.

All three models are evaluated using training accuracy, validation loss, confusion matrices, and class-wise performance metrics. Results show that the CNN-LSTM architecture provides the most balanced performance across tumor types, while the CNN-RNN model suffers from mild overfitting. The EfficientNetB0 baseline offers stable and efficient classification for general benchmarking.

UAVs have emerged as a transformative technology for various applications, including emergency services, delivery, and video streaming. Among these, video streaming services in areas with limited physical infrastructure, such as disaster-affected areas, play a crucial role in public safety. UAVs can be rapidly deployed in search and rescue operations to efficiently cover large areas and provide live video feeds, enabling quick decision-making and resource allocation strategies. However, ensuring reliable and robust UAV communication in such scenarios is challenging, particularly in unlicensed spectrum bands, where interference from other nodes is a significant concern. To address this issue, developing a distributed transmission control and video streaming is essential to maintaining a high quality of service, especially for UAV networks that rely on delay-sensitive data.

In this MSc thesis, we study the problem of distributed transmission control and video streaming optimization for UAVs operating in unlicensed spectrum bands. We develop a cross-layer framework that jointly considers three inter-dependent factors: (i) in-band interference introduced by ground-aerial nodes at the physical layer, (ii) limited-size queues with delay-constrained packet arrival at the MAC layer, and (iii) video encoding rate at the application layer. This framework is designed to optimize the average throughput and PSNR by adjusting fading thresholds and video encoding rates for an integrated aerial-ground network in unlicensed spectrum bands. Using consensus-based distributed algorithm and coordinate descent optimization, we develop two algorithms: (i) Distributed Transmission Control (DTC) that dynamically adjusts fading thresholds to maximize the average throughput by mitigating trade-offs between low-SINR transmission errors and queue packet losses, and (ii) Joint Distributed Video Transmission and Encoder Control (JDVT-EC) that optimally balances packet loss probabilities and video distortions by jointly adjusting fading thresholds and video encoding rates. Through extensive numerical analysis, we demonstrate the efficacy of the proposed algorithms under various scenarios.

Combinational and sequential locking methods are promising solutions for protecting hardware intellectual property (IP) from piracy, reverse engineering, and malicious modifications by locking the functionality of the IP based on a secret key. To improve their security, researchers are developing attack methods to extract the secret key.  

While the attacks on combinational locking are mostly inapplicable for sequential designs without access to the scan chain, the limited applicable attacks are generally evaluated against the basic random insertion of key gates. On the other hand, attacks on sequential locking techniques suffer from scalability issues and evaluation of improperly locked designs. Finally, while most attacks provide an approximately correct key, they do not indicate which specific key bits are undetermined. This thesis proposes an oracle-guided attack that applies to both combinational and sequential locking without scan chain access. The attack applies light-weight design modifications that represent the oracle using a finite state machine and applies an assertion-based query of the unlocking key. We have analyzed the effectiveness of our attack against 46 sequential designs locked with various classes of combinational locking including random, strong, logic cone-based, and anti-SAT based. We further evaluated against a sequential locking technique using 46 designs with various key sequence lengths and widths. Finally, we expand our framework to identify undetermined key bits, enabling complementary attacks on the smaller remaining key space.

In the era of digital commerce, personalized recommendations are pivotal for enhancing user experience and boosting engagement. This project presents a comprehensive recommendation system integrated into an e-commerce web application, designed using Flask and powered by collaborative filtering via Singular Value Decomposition (SVD). The system intelligently predicts and personalizes product suggestions for users based on implicit feedback such as purchases, cart additions, and search behavior.

The foundation of the recommendation engine is built on user-item interaction data, derived from the Brazilian e-commerce Olist dataset. Ratings are simulated using weighted scores for purchases and cart additions, reflecting varying degrees of user intent. These interactions are transformed into a user-product matrix and decomposed using SVD, yielding latent user and product features. The model leverages these latent factors to predict user interest in unseen products, enabling precise and scalable recommendation generation.

To further enhance personalization, the system incorporates real-time user activity. Recent search history is stored in an SQLite database and used to prioritize recommendations that align with the user’s current interests. A diversity constraint is also applied to avoid redundancy, limiting the number of recommended products per category.

The web application supports robust user authentication, product exploration by category, cart management, and checkout simulations. It features a visually driven interface with dynamic visualizations for product insights and user interactions. The home page adapts to individual preferences, showing tailored product recommendations and enabling users to explore categories and details.

In summary, this project demonstrates the practical implementation of a hybrid recommendation strategy combining matrix factorization with contextual user behavior. It showcases the importance of latent factor modeling, data preprocessing, and user-centric design in delivering an intelligent retail experience.