Abstract
As the presence of artificial intelligence (AI) grows in societies around the world, many countries are concerned that its rapid development is outpacing the enactment of ethical frameworks that safeguard the rights of the citizenry. Despite these concerns, the United States (US), a global leader in AI development, has already begun to integrate AI into its legislative practices. Thus, my STS paper seeks to answer the following research question: “To what extent has AI become a standard tool for American policymakers, and with what consequences?” After gathering statistics and case studies to analyze how commonplace has become American policymaking, I found that AI is not a standard tool for American policymakers despite the fact that AI use has increased in federal agencies within the past few years. I also found that the lack of AI use policies and training initiatives for American legislators might explain why a portion of them have not adopted it into their work.
Unlike the state governments, which are individually making strides toward governing AI use in their respective legislatures, the federal government seeks to remove AI regulation altogether in favor of increased marketability. Without proper guardrails, however, the incorporation of AI into policymaking will give the private sector undue influence over governance, endangering the rights of minorities and posing a significant threat to national security. To address these concerns, I recommend that all policymakers complete an AI training program and familiarize themselves with international documents such as the Kuala Lumpur Declaration and the Maturity Framework for AI in Parliaments, which provide important recommendations for the adoption of AI tools into legislatures.
I also recommend that the US government invest in AI security projects like my Systems Engineering technical capstone, CleanSight, which researches the effectiveness of four automated detection strategies (Gaussian Mixture Models (GMMs), k-means clustering, Mahalanobis distance, and Wasserstein distance) in identifying label-flipping attacks used on the MNIST image dataset. The importance of my capstone is underscored by the fact that AI is migrating into critical decision-making domains as shown in my STS project. It is also underscored by the novelty of the detection strategies; unlike most state-of-the-art detection strategies, none of these make assumptions about the attributes of the attack that they are trying to detect.
My capstone team and I chose to focus on detecting label-flipping attacks because they are a common type of cyberattack in which a bad actor changes the classification labels of the training images to reduce model performance. We also chose to evaluate how the detection strategies work against label-flipping attacks on the MNIST dataset because the dataset is publicly-accessible and contains tens of thousands of grayscale images of digits in varying writing styles. We tested our strategies by developing an automated pipeline that flips the labels of the images, compresses their data, and inputs the data into each strategy, which outputs a score representing the probability that the label of the image had been altered.
While we found that the k-means clustering strategy identified the most label-flipped images overall, we also found that other strategies outperformed it in certain evaluation metrics. Most notably, the GMM strategy had the highest precision of all the strategies, meaning that a high percentage of the images it identified were label-flipped. Moreover, the k-means strategy was outperformed by other strategies on certain subcategories of label-flipping attacks. The Mahalanobis distance strategy, for example, had the highest overall performance on untargeted label-flipping attacks, which occur when images from random image classes within the dataset are label-flipped. In the future, we would like to test whether the effectiveness of these strategies holds for more complex datasets such as CiFAR-10, which contains colored-images of living things, such as cats and dogs, as well as non-living things, such as airplanes and trucks.
Finally, I would recommend that the government invest in research that studies the incorporation of software accessibility in higher education as disabled audiences are often neglected in the development of new technologies. One example of such research would be my Computer Science technical report, which discusses how CS3240: Software Engineering is a Computer Science course offered at the University of Virginia that lacks material on software accessibility even though it requires students to build accessible web applications. The proposed solution to this problem is to incorporate design concepts from SYS3023: Human-Machine Interface in future CS3240 curriculums into two software accessibility lectures taught by an SYS3023 professor and into an assignment where students use online tools to evaluate whether a webpage they created is appropriate for a disabled audience. The anticipated effect of the proposed solution is that CS3240 students will become more aware of how software design impacts a disabled user’s interaction with certain types of software. In the future, the proposal would benefit from feedback from both current former CS3240 students and professors because they can share their personal experience with the course and how it has attempted to address software accessibility and provide insight on what kinds of lectures and assignments would be most effective in introducing CS3240 students to new concepts.
Notes
School of Engineering and Applied Science
Bachelor of Science in Systems Engineering (primary major) and Computer Science (secondary major)
Technical Advisors: Hunter Moore, Rosanne Vrugtman, Briana Morrison
STS Advisor: Karina Ripley
Technical Team Members: Eli Cook, Adam Fridley, Ashraf Ibrahim, Hunter Oakey
