Three Professional Computer Science Master's Projects Win 2021 Innovation Prize
Three projects from the master's program in professional computer science have won the Innovation Prize for this year. Selected students were chosen by their instructors to present their project to program's advisory board members, who chose winners based on the following criteria: innovation, potential impact and quality of implementation.
The Innovation Prize fund was established in 2019 with the goal of supporting SFU computing science students in creating innovative projects. Each winning team received $2,500 in recognition of their exceptional work. Students who competed but did not win received a smaller award honouring their contributions to the competition.
"This year is special because students had to work on projects remotely with their team members,” says computing science professor Jiannan Wang, director of the program.
“This certainly created more challenges in terms of team communication and project management. I am proud to see that the three winning teams quickly adapted to this change and delivered high-quality projects, which greatly impressed our advisory board members."
Below is information on each of the winning projects, as well as some testimonials from the students:
Prediction on Antibiotics Resistance and Length of Stay for ICU Patients
Project Presenters: Kyoun Huh, Bilal Hussain, Sakina Patanwala, Hon Wing Eric Chan
Abstract: In this project, we have used the MIMIC-III openly available dataset developed by the MIT Lab for Computational Physiology to predict if an organism is resistant or sensitive towards an antibiotic measured by interpreting a culture test as well as to predict a patient's length of stay and mortality rate using their demographics and first 24-hour vital signs. We have specifically selected several features from the dataset after in-depth advice from medical professionals, medical literature review, and algorithms in some instances to train numerous supervised learning algorithms. We have designed a general dashboard with processed and saved data and a Flask application for delivering real-time prediction and have deployed both applications on production using Docker, which is registered in the Docker registry images on a GitHub registry space to facilitate the release of products. Simultaneously, we have utilized an AWS EC2 to host the docker containers for enhanced availability. Our web application allows a caregiver to enter patient-specific details to know the antibiotic effectiveness. A medical professional can also enter a patient's ICU Stay ID to determine if the patient has a high mortality rate or length of stay in ICU through our system based on vital signs to help them make critical healthcare decisions.
Student testimonials:
“The biggest takeaway would be the experience of handling complex, real-world and critical data in the medical area, coupled with struggling to find an appropriate design with cloud-based services.” - Kyoun Huh
“I am deeply honored to receive this reward. It was a very interesting experience to apply everything we have learned in this professional master's program to solve health care problems that can save lives." - Hon Wing Eric Chan
Conditional IMLE based Multimodal Shape Completion
Project presenter: Himanshu Arora
Abstract: Shape completion is the problem of completing partial scans of shapes. This problem has vital applications in computer vision and robotics due to issues like occlusion or sparsity in real-world data. However, most of the existing research related to shape completion has been focused on completing shapes by learning a one-to-one mapping which limits the creative facets of this problem. In this work, we propose a novel Multimodal Shape Completion technique that is effectively able to learn a one-to-many mapping and generate diverse completed shapes. Our approach is built on the recently proposed Conditional Implicit Maximum Likelihood Estimation (IMLE) technique wherein we condition out inputs on partial 3D point clouds. We extensively evaluate our approach by comparing to various baselines both quantitatively and qualitatively. Further, we demonstrate state-of-the-art results in terms of completeness and diversity in the generated shapes.
Student testimonial:
“I myself proposed this project idea and even though it seemed ambitious initially, my professor encouraged me to take it up. I had a great learning curve during this project and I'm now working to publish it in a top conference.” - Himanshu Arora
eBPF-IDS - A high-performance network intrusion detection engine
Project presenters: Uwe Helm, Tsz Tat (Dickson) Lee, Sameer Ahmed, Aayush Chopra
Abstract: Detecting anomalies and potential attacks within network traffic requires observation and processing of all network data on a connected device. The application of detection rules to this network data is often inefficient and leads to high resource usage. We implement a low-level network anomaly detection engine with the goal of maximum throughput at minimal performance overhead. Our detection engine collects and processes data in the context of an eBPF program attached to a Linux kernel. eBPF allows sandboxed bytecode to be loaded and executed directly in kernel-space and is ideal for high-throughput applications that require continuous access to network data. To minimize processing overhead further, inbound modules utilize XDP hardware offloading and and can process packets before any memory allocation in the kernel takes place.
The detection engine is modular and extendable and currently implements detection rules for ARP spoofing, TCP SYN flooding, TCP/UDP port scans and ICMP tunneling. Relevant packet information is extracted within the eBPF context and passed to a console-based userspace frontend through a shared hash map. The frontend allows users to activate desired detection modules and displays alerts immediately upon detection.