Research Projects
Optimized Matrix Data structure for Sparse Matrix Multiplication (SpMM)
Modern applications like machine learning often exhibit 10-50% sparsity, yet current sparse computing implementations only achieve performance benefits at extremely low sparsity levels (1-3% nonzeros). This creates a significant gap where potentially beneficial sparse optimizations remain unused. The project aims to construct new data structures and algorithms specifically designed for low-sparsity applications, using the looplet abstraction to implement novel intersection strategies that can efficiently handle higher nonzero densities.
I am benchmarking sparse-dense matrix multiplication across varying sparsity levels using Julia, comparing performance against dense implementations to identify optimization opportunities. My work involves implementing and evaluating new data structures (bitmap-based formats and structured sparsity schemes), and developing corresponding iteration algorithms and code generation strategies.
Fixed-Parameter Algorithms for Partial Graph Complementation
Partial complementation (k-flips) of graphs has emerged as an important area with applications in quantum computing, yet algorithmic research remains limited. This field lacks comprehensive study of complexity, approximation algorithms, and fixed-parameter tractability (FPT) for finding specific graph properties under partial complementation operations, presenting opportunities for theoretical advancement.
I am developing algorithms to solve the optimization problem of finding the minimum k such that a graph G has a partial complement whose largest component contains k vertices. My approach focuses on creating FPT algorithms while analyzing scaling behavior with approximation ratios and exploring kernelization techniques.
Intrinsics within LLVM Clang
Compiler intrinsics are essential for high-performance computing as they leverage built-in optimizer knowledge to generate faster code than equivalent inline assembly. While optimizing a cryptographic multi-precision modular multiplication function in LLVM Clang, we discovered that reordering intrinsics could significantly improve runtime performance. However, minimal research exists on how compilers handle intrinsic ordering and what sequences produce optimal results.
I conducted an in-depth study of LLVM's intermediate representation (IR) and compiler passes to understand how intrinsic code is processed within LLVM Clang. Through systematic experimentation, I analyzed how different orderings of intrinsic instructions affect optimization outcomes.
Side-channel Attack (SCA) preventive Analog to Digital Converter (ADC)
Integrating Analog to Digital Converters (ADCs) are widely used for data measurement but remain vulnerable to Trojan Hardware and power side-channel attacks (SCA). This project developed a security solution using Rancode implementation and internal dummy counters to prevent attackers from extracting sensitive information through power consumption analysis, enhancing the overall security of ADC designs.
I created and led a 4-member research team under the Secure Hardware Vertically Integrated Projects (VIP) group, organizing work across analog circuit design and SCA testing subgroups. My technical work included developing expertise in power side-channel attacks using ChipWhisperer tools, formulating and executing differential power analysis experiments on real target devices, and designing attack methodologies specifically tailored to evaluate our secure ADC design with dummy counter countermeasures. I delivered monthly progress presentations to the VIP group and created comprehensive technical documentation to support future students continuing this work. (Technical Document, Slides)
Use of Out of Band (OOB) in Bluetooth
Bluetooth communication remains vulnerable to Man-in-the-Middle (MitM) attacks during the initial pairing phase, despite advances in wireless security research. This project explored combining Near Field Communication (NFC) with Bluetooth's Out-of-Band (OOB) pairing protocol to create a more secure connection method. By leveraging NFC's requirement for physical proximity, this approach offers a promising solution to deter remote attacks during the critical pairing process.
I independently led this research project under the Secure Hardware VIP group. My work included studying and testing Bluetooth 5.3 standards, conducting practical MitM attack demonstrations on Bluetooth devices, and evaluating the security benefits of NFC-assisted pairing. I delivered monthly progress presentations to the VIP group and created comprehensive technical documentation to support future students continuing this work. (Technical Document, Slides)