About
Call for PapersThe REU workshop seeks original papers that focus on the area(s) of machine learning and cryptography. This workshop allows undergraduates the opportunity to showcase their work over the past ten weeks in these areas. We accept submissions of original research papers discussing these following subjects.
Topics of Interest include (but are not limited to):
- Natural Language Processing
- Image and/or Text Classification
- Security and Cryptography
- Machine Learning Models and Techniques
Review ProcessEach group will review each other submitted paper. These reviews must be submitted to Maximilian Tiao or Jonathan Wu by August 12, 2021 11:59 PM Eastern time.
Paper FormatPapers must be written in LaTeX using the IEEE manuscript template and submitted as a PDF. They should be 8.5 by 11 inches and at least 5 pages in length. Papers must be submitted to Maximilian Tiao or Jonathan Wu by August 11, 2021 11:59 PM Eastern time. Posters must be submitted to Sidi Diawo by the same date and time.
Committee
General Co-Chair
Melissa Perez
Website Chairs
William Silva and Nathan Swearingen
Technical Program Chairs
Maximilian Tiao and Jonathan Wu
Poster Chairs
Sidi Diawo and Sanjay Gorur
Publications
Trigger Patch Analysis for Backdoor Attacks in Federated Learning
Maximilian Tiao: Department of Computer Science and Engineering
University of California, Santa Cruz
Melissa Perez: College of Engineering, EECS
Oregon State University
Sidi Diawo: Electrical and Computer Engineering
University of Memphis
Abstract: In federated learning, a centralized server collects and aggregates machine learning models sent by a network of edge devices or clients. Although there are many benefits, federated learning is vulnerable to trigger backdoor attacks. In this attack, adversaries inject a set of poisoned images into the clean training dataset to misclassify them to another category the adversary chooses. In this paper, we are analyzing key parameters that affect the trigger backdoor attack success rate in a simple implementation of federated learning. The parameters we decide to focus on are patch size, patch location, poisoning rate, percentage of malicious clients, and image similarity. We perform several experiments manipulating these parameters to better understand what makes these backdoor attacks effective and successful. In addition, we also perform experiments with different machine learning models to demonstrate that trigger backdoor attacks can be applied in any scenario. Read More
A Standardized Protocol for Secure E-Voting
Nathan Swearingen: School of Science
IUPUI
William Silva: School of Engineering
UConn
Abstract: With many e-voting schemes in the literature to choose from, election officials have many options to satisfy various security requirements. However, in some cases, it is important that participants have multiple implementations to choose from, or that they can create and use their own implementations. Such requires a communication protocol that all parties can understand. In this paper, we present a potential communication protocol for one particular e-voting scheme. Read More
Predicting Patient Mortality Using Graph Convolutional Neural Network Training
Jonathan Wu: CS at The Georgia Institute of Technology
Sanjay Gorur: Electrical and Computer Engineering
University of Texas at Austin
Abstract: Electronic health records contain valuable information in regards to patients’ features and can be manipulated to observe similarities where patients overlap to predict diagnoses. In order to detect such patient risk, a graph is constructed by extracting relevant features to relate patients in the EHRs and thus create a homogeneous graph to assess patient similarity. A novel GCN training algorithm alike Cluster-GCN is further used to predict mortality risk of patients by exploiting the graph clustering structure through node classification. Read More