CleanSight: Detection Strategies for Label-Flipping Data Poisoning Attacks on MNIST; The Defensive Plateau: Societal Impacts of the AI-Security Arms Race 3 views

Introduction This synthesis summarizes the findings from my technical project for developing and testing the data poisoning detection tool “CleanSight” and my STS research focused on the societal impact of Artificial Intelligence (AI) Security development. Both efforts are centered around AI Security, exploring the technical and societal. The CleanSight project explored the strengths and weaknesses of mathematical and algorithmic strategies against adversarial label-flipping, while the STS research aimed to identify global trends regarding the rapid developing AI security technologies while considering ethical and interest based perspectives. In the technical endeavor, my team and I developed CleanSight to automate the detection of label-flipping data poisoning attacks on an image dataset. The intended capability of CleanSight paralleled the dual-use concerns of AI addressed in my STS research, where I investigated the conflict at multiple societal levels and observed the impacts due to the exploitation of such. Project Summaries The technical portion of my thesis involved the design and testing of a code-based pipeline using Python for base language, a modular functional architecture for ease of future code refinement, global seed functions for results reproduction, a PyTorch defined MNIST image dataset, four unique poisoning methods, and four distinct detection strategies. Using such, my capstone group and I observed the strategies performance through the use of three common AI metrics, AUC, precision, and recall. We further identified the distinct strengths and weaknesses of the strategies to each unique poisoning attack. In my STS research, I observed that the societal impacts of AI security are redefining digital trust and geopolitical power dynamics at a fundamental level originating from the exploitation of AI’s dual-use vulnerabilities. Specifically, I identified the democratization of cyber warfare and the emergence of an “AI-security industrial complex,” which imposes severe economic burdens on organizations and states attempting to secure their models. Further, I observed an asymmetric, in capability, offensive-defensive AI arms race, where such burdens cannot be met, exacerbating the “security inequality" and hastens the societal shift towards “Zero-Trust” perceptual views and architecture. Conclusion Working on both CleanSight and the research of AI security societal impacts translated the scale AI development encompasses in variety and outcome and the influence researchers and frontier AI companies have over these AI systems. These projects deepened my technical proficiency and expanded my knowledge and ethical perspectives on AI development trends from the technological capability to the dynamics underpinning global scientific innovation and integration. The insights gained from these projects advocate for a multimodal epistemic approach to development and deployment, where societal stability and sustainment are not degraded by growing pressures from the AI security industry. Acknowledgements I acknowledge and thank the support and wisdom of my academic advisors, Dr. Richard Jacques and Dr. Hunter Moore.

BS (Bachelor of Science)

data poisoning; ai security; societal impacts of ai security; defensive plateau; label-flipping

School of Engineering and Applied Science Bachelor of Science in Systems Engineering Technical Advisor: Hunter Moore STS Advisor: Richard D. Jaques Technical Team Members: Devon Alexander, Eli Cook, Ashraf Ibrahim, Hunter Oakey ---- STS Research Paper - Generative AI Use Note: Google Gemini 3.1 Pro Deep Research and Google NotebookLM tools were strictly used to assist in research gathering and paper structure formulation.

English

All rights reserved by the author (no additional license for public reuse)

2026-05-09