School of Electrical Engineering & Computer Science
Washington State University, Pullman, WA, USA
Research Experience
Real-time Scheduling of DNN Workloads in Secure Enclaves
Enabling timing guarantees for performing confidential deep inference in latency-critical learning-enabled systems.
A new scheduling framework and analytical model to determine the feasibility of deploying a given real-time DNN workload
on TrustZone enclaves.
A novel task fusion approach to further reduce TEE context-switch overheads while retaining real-time guarantees.
Exploring Covert Channel/Information Leakage in Real-time Systems
We propose an analytical model that finds the existence of an algorithmic covert channel for a set of multi-frame tasks.
We analyze the problem of information leakage in dual-mode fixed-priority real-time systems.
We propose a statistical analysis that allows low-priority tasks to infer execution patterns of a high-priority task.
Exploring the Feasibility of Deploying LLMs in Embedded Devices and Performance Analysis
Our study systematically investigates the performance of BERT-based language models on four off-the-shelf embedded platforms
(Raspberry Pi, Jetson, UP², and UDOO) with two different memory variants (2 GB and 4 GB RAMs). We analyzed the
trade-offs between complexity and accuracy across multiple NLP tasks.
We explore the feasibility of deploying complex NLP tasks on embedded systems and analyze them under three metrics:
(a) inference time, (b) memory usage, and (c) energy consumption. We developed a lookup table through empirical
observations that will be useful for system designers to decide suitable model configurations for the target platform.
Common vulnerabilities and exposures (CVE) dataset analysis: A collaboration with Pacific Northwest National Laboratory
Analyzed CVE datasets to identify patterns in vulnerability exploitability using signal processing techniques (e.g., FFT) and clustering methods (e.g., Mahalanobis and Euclidean distances).
Investigate co-exploitation behaviors, fragmentation patterns, and exploit density to understand coordinated vulnerability exploitation.
