Human-in-the-Loop Training to Enhance Object Detection Models; Regulating Facial Recognition: A Balance of Privacy and Security

The relationship between the STS Research paper and the computer science capstone project is image recognition. The capstone project discusses using machine learning and machine learning techniques to better train a model to detect and classify ships from satellite imagery. While the STS research paper discusses the ethics and problems facing facial recognition, which is another form of image recognition, like that used in detecting ships. With both of these papers an understanding of how these image recognition models technically work are understood along with the moral impacts of the same type of image recognition and how it affects a society is also explored and debated providing the reader with a strong understanding of the technology as a whole.
The Capstone project covers image recognition, but more specifically image recognition of ships from satellite imagery. The paper goes into my internship experience at the Naval Information Warfare Center where a model had to be created to accurately classify and locate different types of ships. The technical challenge discussed in the report is training the machine learning model quickly and efficiently to achieve acceptable accuracy. The report goes into how a human in the loop method was implemented. The human-the-loop process involves creating a partially trained model from hand labelled data. Even though hand labelling data takes extra time it is needed to get a baseline model to further detect images. With the model trained on these hand-labelled it was then possible for the model to label its own data as long as a human was in the loop verifying what the model labelled was correct. If it was not correct, then the person in the loop would manually relabel it. As this process iteratively repeated the model would get more accurate and thus leading to less need for the person overseeing the model to intervene and change the labelled data. The HITL process in the end leads to a model with a large training set and more accurate parameters which also leads to a more accurate model. The application of this project is also included within the technical report along with how courses within the University of Virginia helped prepare me for the type of work performed and what could be changed with the courses to better prepare me.
The STS paper discusses the advent of facial recognition technology in law enforcement and private company settings and how that is affecting both the privacy and security of U.S citizens. Facial recognition is currently being deployed into increasingly more cities across the country and is being used as a tool for law enforcement to identify individuals. Facial recognition raises many serious concerns over the breach of privacy of many people being identified by these systems and whether that breach of privacy is worth the additional safety facial recognition provides. The overarching research question that will be answered throughout the paper is what sort of regulation the United States could enact to better protect the privacy of its citizens while mitigating the security facial recognition provides. The problem of the tradeoff between privacy and security is inherently a wicked problem due to having more information on individuals’ leads being able to stop malicious acts or identify individuals or committed them. This paper will hope to uncover not a solution to the problem but a viable tradeoff to a problem that has not seen any form of regulation to help control it. There will never be a completely correct solution to this debate but there are basic laws and regulation that could be enacted to stopping abuse of facial recognition technology by both government and private institutions.
Working on both of these projects at the same time gave me a unique opportunity to both deeply explore a technical aspect while considering its ethical impacts and how I as an engineer contribute to a technology that does have ethical problems. Having this strong technical understanding of the underlying technology of image recognition also helped in understanding the problems of facial recognition at a deeper level. This allowed a very informed discussion of feasible solutions to the problem of rising use of facial recognition in law enforcement and private companies. While on the other side realizing the ethical concerns of image recognition helps create safer and more ethical machine learning models, which as an engineer, is something that I can personally control and influence heavily. Overall, both projects led to me becoming a more informed engineer to the technology I was helping create which in turn will improve the work that I may continue to do within the field.

School of Engineering and Applied Science

Bachelor of Science in Computer Science

Technical Advisor: Rosanne Vrugtman

STS Advisor: Bryn Seabrook