Hello!

I am currently a PhD candidate in the Department of Computer Science at the University of Maryland at College Park, where I am advised by Prof. Tom Goldstein and Prof. David Jacobs. In May 2019, I earned my bachelor's degree in Computer Science and Mathematics from Harvey Mudd College. I am broadly interested in computer vision and deep learning. My research has explored vulnerabilities of neural networks including adversarial examples and dataset poisoning.

In my free time, I like to run and build stuff.

Publications

1. Sandoval-Segura, Wang, Panda, Goldblum, Basri, Goldstein, Jacobs. "Using Attention Sinks to Identify and Evaluate Dormant Heads in Pretrained LLMs". Under Review. [arXiv:2504.03889]

2. Singla, Sandoval-Segura, Goldblum, Geiping, Goldstein. "A Simple and Efficient Baseline for Data Attribution on Images". NeurIPS Workshop on Attributing Model Behavior at Scale (ATTRIB), 2023. [arXiv:2311.03386][Code]

3. Sandoval-Segura, Singla, Geiping, Goldblum, Goldstein. "What Can We Learn from Unlearnable Datasets?". Advances in Neural Information Processing Systems 37 (NeurIPS), 2023. [Proceedings][arXiv:2305.19254][Code]

4. Sandoval-Segura, Singla, Geiping, Goldblum, Goldstein, Jacobs. "Autoregressive Perturbations for Data Poisoning". Advances in Neural Information Processing Systems 36 (NeurIPS), 2022. [Proceedings][arXiv:2206.03693][Code]

5. Sandoval-Segura, Singla, Fowl, Geiping, Goldblum, Jacobs, Goldstein. "Poisons that are learned faster are more effective". In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2022. [CVF Open Access]

6. Bashir, Montañez, Sehra, Sandoval-Segura, Lauw. "An Information-Theoretic Perspective on Overfitting and Underfitting". In Australasian Joint Conference on Artificial Intelligence (AJCAI), 2020. [arXiv:2010.06076]

7. Sandoval-Segura, Lauw, Bashir, Shah, Sehra, Macias, Montañez. "The Labeling Distribution Matrix (LDM): A Tool for Estimating Machine Learning Algorithm Capacity". 12th International Conference on Agents and Artificial Intelligence (ICAART 2020), 2020. [arXiv:1912.10597]

8. Drissi, Sandoval, Ojha, Medero. "Harvey Mudd College at SemEval-2019 Task 4: The Clint-Buchanan Hyperpartisan News Detector". In Proceedings of The 13th International Workshop on Semantic Evaluation (SemEval), 2019. [arXiv:1905.01962]

9. Drissi, Watkins, Khant, Ojha, Sandoval, Segev, Weiner, Keller. “Programming Language Translation using a Grammar-Driven Tree-to-Tree Model”. ICML Workshop on Neural Abstract Machines and Program Induction v2 (NAMPI), 2018. [arXiv:1807.01784]

Other Writing

• Our work on breaking image protections against AI editing using JPEG was featured by New Scientist!

• My Master's paper: "Adversarially Robust Segmentation Models Learn Perceptually-aligned Gradients" [arXiv:2204.01099]

• A project from summer 2021: "AutoProtoNet: Interpretability for Prototypical Networks" [arXiv:2204.00929]

• What is an Adversarial Example and Why Do They Matter?

• How to Train a Neural Network Classifier on ImageNet using Tensorflow 2

• A summary about an error in the proof of Lemma 1 of Monte-Carlo Planning in Large POMDPs.

Teaching

CMSC 421: Introduction to Artificial Intelligence (Spring 2021), Graduate Teaching Assistant, University of Maryland at College Park

CMSC 421: Introduction to Artificial Intelligence (Fall 2020), Graduate Teaching Assistant, University of Maryland at College Park

CMSC 436: Programming Handheld Systems (Fall 2019), Graduate Teaching Assistant, University of Maryland at College Park

CS 81: Computability and Logic (Spring 2019), Grader and Tutor, Harvey Mudd College

MATH 187: Operations Research (Spring 2019), Grader, Harvey Mudd College

Past Projects

• An Alexa skill + Flask app running on Raspberry Pi that opens/closes my blinds:

• Automated tapping the screen on an iPhone by using a relay to "click" an RF remote control page turner. Required breaking open the remote and soldering wires to integrate the relay. Note to self: there are much easier ways to do this!

• An implementation of the GD-UAP algorithm in PyTorch. This code can be used to generate UAPs which can fool a variety of image classifiers with high accuracy.

• I developed experimental software to turn AR-enabled iPhones into 3D-scanners. Here's a point cloud of a full-size pickup truck captured with my iPhone 7:

  

• TapTunes: an iOS app which brings album dashboards to augmented reality. The project won 2nd Place in the Advanced category of the 5C Hackathon in 2017. Here's what a Spotify playlist album tower looks like on top of my college desk:

  

• Arithmetical: a mix of math games for iOS where users race against a timer to keep Spotify music playing. The app accumulated over 3000 downloads. Due to yearly Apple Developer membership fees, the app is no longer available through the App Store, but it can still be found on GitHub.

  

• I wired and wrote software for a 32x32 LED matrix that displays time, local weather, and a recent news headline. It uses an Arduino Mega outfitted with an ethernet shield:

  

Other

Contact

psando AT cs DOT umd DOT edu