Patrick M. Mendoza

Hi there, I’m Patrick. I’m a 3rd year at UC Berkeley studying EECS and Applied Math. My interests lie within the intersection of Computer Science and Math such as Software Engineering, Machine Learning, and Cybersecurity. In particular, I’m very interested in Adversarial Machine Learning.

Currently, I’m an undergraduate researcher at Berkeley’s Artifical Intelligence Resarch (BAIR) Lab investigating the usage of latent diffusion models for latent-space represented speech under Cheol Jun Cho and Professor Gopala Anumanchipalli and with David Wagner’s Research Group, finding ways to mitigate safety alignment degradation of LLMs during fine-tuning.

In the past, I was an undergraduate researcher at UC Berkeley’s Algorithms & Computing for Education (ACE) Lab where I built anti-cheat detection software for online take-home assessments for the educational platform PraireLearn under Ruiwei Xiao and Professor Dan Garcia and at the University of Florida’s Reliable, Intelligent, Secure Internet-of-thINGs (RISING) Lab where I developed a generative trajectory algorithm for Connected and Autonomous Vehicles (CAVs) under malicious V2I communications under Tashfique Hasnine Choudhury and Professor Sandip Ray.

selected publications

Poster: Vehicle-to-Infrastructure Security for Reduced Speed Work Zone

Patrick M. Mendoza, Tashfique Hasnine Choudhury, and Sandip Ray

In Proceedings of the Twenty-Fourth International Symposium on Theory, Algorithmic Foundations, and Protocol Design for Mobile Networks and Mobile Computing, Washington, DC, USA, 2023

Abs DOI

We consider the cybersecurity challenges arising from communications between autonomous vehicles and smart infrastructures. In particular, we consider coordination between vehicles and Reduced Speed Work Zones (RSWZ). Malicious or tampered communications between these entities can have catastrophic consequences. We discuss methods for the analysis of such attacks. In particular, we show how to generate configurable, effective vehicular trajectories for exploring such attacks and how to utilize such trajectories in identifying impactful attacks and evaluating defenses.
Detecting Cheating in Online Take-Home Exams with Randomized Questions

Ruiwei Xiao, Eduardo Huerta-Mercado, and Dan Garcia

In Proceedings of the 54th ACM Technical Symposium on Computer Science Education V. 2, Toronto ON, Canada, 2023

Abs DOI

The last three years were a significant challenge for educational institutions, due to the loss of face-to-face instruction and exam proctoring. Many instructors turned to asynchronous, online exams as a replacement for standard pen-and-paper exams. It is no surprise that many tools aimed at delivering computer-based assessments have become popular and are centers of research and development. This poster discusses our attempt to build a post-exam cheating detection system for the PrairieLearn open-source platform that supports randomized question generators, to uncover irregularities in submissions. Our system compares all pairs of students using four rules: Times (did students take the exam synchronously?), Answers (did they have similar wrong answers?), Orders (did they answer the questions in the same order?), and Scores (did they achieve the same scores?). It adds one final individual rule, the Score-Time-Ratio, that measures how many "points per minute" a student has earned, to flag students who open the exam, copy in a perfect answer, and submit. We deliver a detailed report to the instructor, allowing them to sort their students based on these measures, providing a data-driven way for them to investigate.