Carnegie Mellon University

A bright yellow graphic depicting cartoonish images of technology

May 14, 2024

INI Students Present Novel Research on Topics from Generative AI-Enhanced Cyberdefense to Deep Learning in Master's Thesis Projects

By Evan Lybrand

INI Communications Team

The programs offered by the Information Networking Institute (INI) provide flexible, interdisciplinary curricula, allowing students to build skills while working with our expert faculty and leading industry partners. For students interested in pursuing rigorous academic research, completing a Master's thesis can produce publishable work and demonstrate advanced technical knowledge. Each of the INI’s programs allow for an advanced option that concludes with a thesis.  

This year, 15 graduating INI students successfully defended a Master's thesis, representing four degree programs: MS in Information Networking; MS in Information Security; MS in Artificial Intelligence Engineering - Information Security; and MS in Mobile and IoT Engineering. These theses explore a wide range of topics, including enhancing reconnaissance drones, developing lunar rovers, improving the accuracy and efficiency of neural networks and harnessing generative AI to strengthen cyberdefense.

Explore the cutting-edge work being done by these INI students. 

Master's of Science in Information Networking (MSIN) 

This thesis introduces DroneDSL, a domain-specific language designed to enhance drone mission planning, especially for reconnaissance operations. The current drone mission planning has two main challenges. The first challenge stems from the complexity of integrating software and hardware functionalities, which complicates operations for pilots and limits the potential for code reusability among developers.  To address this, DroneDSL offers an abstraction layer, simplifying interaction with drone systems and promoting developer efficiency through reusable code modules.  The second challenge involves the need for dynamic mission execution capable of supporting complex logical operations such as looping and conditional execution, which surpass the capabilities of existing drone-specific languages designed for straight-forward, linear mission plans. By incorporating a finite state machine, DroneDSL enables drones to adapt their missions dynamically in response to changing conditions, significantly enhancing the expressiveness and effectiveness of drone operations in reconnaissance and beyond. 

The bidirectional reflectance distribution function (BRDF) is a compact representation for the way different materials interact with light. It is a key component for the photorealistic modeling of real-world materials, for example by physics-based rendering systems. There has been extensive research in optics, computer science, and computer graphics on modeling and measuring real-world BRDFs. In particular, data-driven BRDFs are directly acquired by measuring how light reflects off of samples of real-world materials. Well-acquired data-driven BRDFs can represent material details and types analytic BRDFs cannot express. This project explores the use of Markov chain Monte Carlo (MCMC) algorithm to directly draw samples from an unknown BRDF. MCMC can produces samples from any function, so long as we can evaluate it. In the case of BRDF acquisition, this means being able to measure the BRDF at any given incident and outgoing directions—this is indeed the main building block of any reflectometry procedure. This will enable our reflectometry procedure to adapt to the unknown BRDF and measure it at random locations distributed according to the unknown BRDF, without the need to first acquire the BRDF. The effectiveness of any MCMC sampling algorithm critically depends on the proposal distribution we use. We developed proposals that facilitate reflectometry, by exploring a few research directions: First, we explored custom proposals that take advantage of physical and empirical properties of real-world BRDFs—reciprocity, isotropy, bivariate symmetry, and so on. Second, we explored proposals that mimic existing physics-based analytical BRDF models—microfacet, PCA, mixture models. Third, we explored controlled MCMC algorithms, which adapt proposals to previously drawn samples as the MCMC iteration proceeds. Fourth, we explored the state-of-the-art technique of expressing probability density function using normalizing flows and how we can use it as proposal and pdf/BRDF of the material. 

The Sia scheduler represents an advancement in efficiently allocating cluster resources to Deep Learning Training (DLT) jobs in a heterogeneous GPU cluster, resulting in improved Job Completion Times (JCTs) and cluster utilization. However, its current implementation only addresses GPU heterogeneity in its allocation decisions, while eventually choosing to assign only homogeneous GPU resources for a DLT job. This limitation highlights the significance of exploring intra-job heterogeneous GPU allocation during scheduling decisions, which can unlock more cluster utilization and parallelism, thereby further optimizing average JCTs. The aim of the thesis is to address this limitation by extending a state-of-the-art DL execution engine to support automatic heterogeneity-aware hybrid parallelism, which is necessary to effectively utilize intra-job heterogeneous GPU allocations and obtain accurate physical DLT job profiles for Sia’s cluster simulations. This thesis also aims to study the effects of intra-job heterogeneous allocations through workload experiments in a small-scale research cluster, demonstrating their benefits on specific types of workloads. Lastly, the thesis will discuss a small independent case study of integrating the Sia Scheduler into an existing analytics platform for data science workloads, enabling Data Scientists/ML Practitioners to leverage the state-of-the-art capabilities of the Sia Scheduler.

This project presents the design, implementation, and evaluation of a novel web- based Integrated Development Environment (IDE) tailored for executing both single-file or multi-directory based projects. This project aims to leverage the latest advancements in containerization and distributed task queues systems to provide a scalable, robust, and ecient code running environment for software development. By harnessing Docker for containerization, our platform isolates user projects in separate execution environments, thus ensuring security and dependency management without sacrificing performance. This architecture not only supports concurrent handling of multiple user requests but also optimizes computational resources by distributing workloads across multiple servers effectively. Our platform also has a user-friendly interface that allows for choosing from a variety of programming languages, real-time code-editing, and an output terminal displaying real-time results, enhancing productivity and facilitating a seamless development experience. 
Moreover, our platform integrates the most cutting-edge large language models to provide comprehensive, in-depth analysis and feedback on user projects in a straight- forward dialogue manner. Users are able to ask for recommendations, analysis, and code-completion suggestions in our interface. Our evaluation metrics demonstrate the ability to handle a large volume of con- current workloads while maintaining low latency and high availability, making it an ideal platform for both educational and professional usages. The result suggests the success of our ambition in creating a fully cloud-based software development environment and proposes a partial solution to obstacles faced in modern IDEs.

Master’s of Science in Information Security (MSIS) 

The lunar landscape contains pits caused by volcanism and impact cratering. Lunar pits are sites that sample lunar stratigraphy, potentially including a series of distinct lava flows, paleoregolith layers, crater ejecta, and impact melt deposits. These details provide valuable information about the moon’s evolution.The mission is to rove around the perimeter of these volcanic pits stopping intermittently to view the interiors with four high resolution cameras that image from the visible to the thermal infrared. These images stitch together to form a 3-dimensional model of the pit that can then be used for scientific purposes. My goal for the project is to help the PITMAGIC team build the rover capable of accomplishing the mission, especially aiding in the embedded and computer systems aspects of the rover.

Ensuring the integrity of software components during system boot is crucial for establishing trust in a computing system. Various technologies, including UEFI Secure Boot, Trusted Boot, and Measured Boot, focus on securing the bootstrapping process, creating a chain of trust from the firmware to the kernel and device drivers. However, this chain of trust typically ends once the operating system is running. Techniques like Integrity Measurement Architecture (IMA) and Policy-reduced Integrity Measurement Architecture (PRIMA) aim to extend this trust to the application layer, measuring the integrity of user-mode applications after the bootstrapping process on Linux/SELinux platforms. However, there exists a gap in the chain of trust in the Windows operating system. This thesis proposes a novel model for measuring the integrity of user-mode processes on Windows and explores how Remote Attestation can be utilized for the application layer.

Garbled Circuits (GC) originally proposed by Yao is the cryptographic technique that is traditionally used to perform secure two party computation (2PC). However, circuits by their nature are static, and if the two parties want to securely compute a RAM program involving dynamic memory accesses, then converting the RAM program to a circuit would incur a quadratic cost as all memory locations will need to be scanned on each access. To avoid this quadratic blowup, the abstraction of Garbled RAM (GRAM) was introduced by Lu and Ostrovsky to garble RAM programs directly without converting to a circuit. In this paper, we use techniques of tri-state circuits (TSC) proposed by Heath et. al. to present a new Garbled RAM scheme which matches the best known asymptomatic overhead of O(log^3(N) * loglog(N)) gates per access while providing significant improvements in constant factors. In our scheme we propose a new construction for the garbled stack data structure with both an average-case and worst-case cost of O(logN). Our new construction is an improvement over the existing garbled stack construction used in all previous works which has a worst-case cost of O(N). Finally, we use our new scheme to create the first ever practical implementation of a GRAM.

Network encryption can protect its user’s privacy and identity, but encrypted traffic is still vulnerable to side-channel attacks. A VPN or TOR can provide added layers of protection, but even with these protections, leaking information through side-channel analysis is still possible. Side-channel attacks can lead to private or confidential information being revealed. Dynamic Adaptive Streaming over HTTPS is one example of a protocol that is vulnerable to side-channel attacks. This paper examines two machine-learning models for identifying encrypted videos streamed using DASH. The models include a convolutional neural network and a random forest classifier. Different features and capture lengths were examined and compared for each machine-learning model. The video traffic includes videos streamed using SSL, VPN, and TOR connections. Different streaming qualities (240p, 360p, 540p, and auto) were used when streaming the videos and collecting the traffic data. The data was divided into two categories: TOR only and mixed data (mix of SSL, VPN, and TOR). Each method has different benefits and tradeoffs, including accuracy, scalability, and complexity. Achieving an accuracy over 90\% in some test cases, these methods demonstrate a vulnerability in the DASH protocol that allows videos to be uniquely identified from encrypted traffic.

In the current landscape of reverse engineering tools, the Unified Extensible Firmware Interface (UEFI) remains largely under-developed; a significant portion of analysis time is spent sifting through the underlying architecture and identifying the relevant components for individual analysis with traditional executable reverse engineering techniques. In this paper, we introduce EFIdra, a tool which we built to address these issues by identifying executables within an image, allowing analysts to follow references between executables, and supporting extensibility for adding new architectures and formats. We then conduct a case study on analyzing a UEFI image using widely available tools and techniques, and compare it with our analysis using EFIdra to demonstrate the gaps that our tool fills and the ways in which it reduces the barrier to entry for UEFI reverse engineering.

Master’s of Science in Artificial Intelligence Engineering - Information Security (MSAIE-IS) 

Recent decades have seen the unprecedented success of Artificial Intelligence (AI), with its impact resonating beyond the confines of the technology sector to influence fields as diverse as finance, healthcare, social media, and education. Despite achieving remarkable service advancements, the security implications of AI technologies have not been adequately explored. This oversight raises critical concerns, particularly as our reliance on digital infrastructure grows. In this context, cybersecurity emerges as a pivotal element, ensuring the integrity, confidentiality, and availability of information against an increasingly sophisticated threat landscape. My thesis aims to address the dual facets of the secure AI challenge: firstly, assessing the vulnerability of current AI systems to cyber threats, and secondly, exploring the potential of AI technologies as proactive defense mechanisms to bolster system security. Specifically, this talk  will delve into federated learning—a security-conscious learning paradigm—and generative AI, the state-of-the-art application, emphasizing their roles in advancing novel attack and defense strategies within the cybersecurity domains. 

We propose a mask pretraining method for Graph Neural Networks (GNNs) to improve their performance on fitting potential energy surfaces, particularly in water systems. GNNs are pretrained by recovering spatial information related to masked-out atoms from molecules, then transferred and finetuned on atomic forcefields. Through such pretraining, GNNs learn meaningful prior about structural and underlying physical information of molecule systems that are useful for downstream tasks. From comprehensive experiments and ablation studies, we show that the proposed method improves the accuracy and convergence speed compared to GNNs trained from scratch or using other pretraining techniques such as denoising. On the other hand, our pretraining method is suitable for both energy-centric and force-centric GNNs. This approach showcases its potential to enhance the performance and data efficiency of GNNs in fitting molecular force fields.

Master’s of Science in Mobile and IoT Engineering (MSMITE) 

As Application Programming Interfaces (APIs) continue to proliferate, driving digital transformation across industries, the need for robust API security becomes paramount. However, existing API security tools rely heavily on clear specifications to conduct effective assessments. Without defined endpoints and input types, these security checks lack efficacy, leaving systems vulnerable to attacks. This research addresses the critical challenge of inferring the exposed API attack surface by identifying endpoints, arguments, and HTTP methods. By obtaining the API specification, tools such as Mayhem for API, StackHawk, and ZAP can conduct thorough security assessments, effectively identifying and addressing vulnerabilities. By providing a holistic solution to API inference and security challenges, this research aims to fortify API security and mitigate potential risks comprehensively. Through innovative methodologies and advanced techniques, this study contributes to the ongoing efforts to strengthen the security posture of modern software systems. 

This Master’s thesis addresses the need to enhance activity recognition for monitoring and guidance in health and wellness applications. Despite the growing expectations, the challenge lies within effectively gathering user activity data. This study aims at assessing the ability of RF (Radio Frequency) data available in smart wireless devices and or wearable devices to improve HAR (Human Activity Recognition) by leveraging the RSSI (Received Signal Strength Indicator) information. The study was performed using on-body wireless sensor nodes provided by the Bosch Research Team. Unlike previous approaches that have utilized RF data to determine coarse grain location tracking, this study emphasizes on leveraging the RF data from wireless protocols such as BLE (Bluetooth Low Energy) to give valuable insights into fine grain relative positioning that help to understand precise human biomechanical motions. This novel approach focuses on utilizing the existing infrastructure present in wireless devices to improve human activity recognition which could be helpful for fitness and wellness related applications. It also has the potential to reduce power usage by enabling motion sensors to be used less. The preliminary results conducted by 25-40 participants prove that there is valuable information given by RF data resulting in a 30% reduction in errors.

Colorizing images is a challenging problem due to its multi-modal uncertainty and high degree of ill-posedness. Directly training a deep neural network often results in inaccurate semantic colors and a lack of color richness. Although recent methods offer improved outcomes, the capability to extract semantic information and the quality of image generation are still limited. To further improve the results, we propose Color Diffuser, a comprehensive approach to recovering vivid colors by leveraging the rich and diverse color priors encapsulated in the pre-trained diffusion model. Specifically, our method comprises a pixel decoder and a query-based color decoder. The pixel decoder integrates the semantic features obtained from the diffusion inversion process. The color decoder utilizes rich semantic features to refine color queries, thus eliminating the need for manually designed priors. Through cross-attention, our two decoders collaborate to establish correlations between color and multi-scale semantic representations, significantly mitigating the color bleeding effect. Extensive experiments demonstrate that Color Diffuser outperforms existing state-of-the-art methods qualitatively.