I am a PhD candidate in the Computer Science department at the
University of Maryland, advised by Professors
Soheil Feizi and
Tom Goldstein. My research focus is on machine learning robustness with provable guarantees. Deep neural networks often malfunction due to minor changes in the input data, which poses a significant risk in safety-critical applications such as autonomous driving and medical diagnosis. My thesis research centers around designing machine learning algorithms with verifiable guarantees of robustness. I study robustness for a variety of deep-learning applications in computer vision, reinforcement learning, and streaming machine learning. I have also worked on robustness for distribution shifts and investigated the fundamental limitations of popular robustness techniques for high-dimensional data. The overarching goal of my research is to develop provable robustness techniques for real-world applications under different learning paradigms. I have led several research projects that have resulted in publications in reputable machine learning conferences like ICML, NeurIPS and ICLR.
In the past, I have also worked on topics in theoretical computer science like approximation algorithms and computational hardness of combinatorial optimization problems. Before joining UMD, I did my undergraduate studies at IIT Mandi and my master's at IIT Delhi. During my PhD, I have spent time as a research intern at Nokia Bell Labs and Amazon where I worked on network security-related machine learning applications and uncertainty estimation for action recognition. I have also served as a reviewer for machine learning conferences such as ICML and NeurIPS. For more on my research interests and contributions, please see my research statement.
Publications
Can AI-Generated Text be Reliably Detected?
Vinu Sankar Sadasivan, Aounon Kumar, Sriram Balasubramanian, Wenxiao Wang, Soheil Feizi
Preprint 2023, arXiv, Media Coverage: The Register, TechSpot.
Provable Robustness against Wasserstein Distribution Shifts via Input Randomization
Aounon Kumar, Alexander Levine, Tom Goldstein, Soheil Feizi
ICLR 2023, arXiv, code.
Policy Smoothing for Provably Robust Reinforcement Learning
Aounon Kumar, Alexander Levine, Soheil Feizi
ICLR 2022 [OpenReview], arXiv, code.
Center Smoothing: Provable Robustness for Functions with Structured Outputs
Aounon Kumar, Tom Goldstein
NeurIPS 2021 [link], arXiv, code.
Curse of Dimensionality on Randomized Smoothing for Certifiable Robustness
Aounon Kumar, Alexander Levine, Tom Goldstein, Soheil Feizi
ICML 2020 [link], arXiv, code.
Certifying Confidence via Randomized Smoothing
Aounon Kumar, Alexander Levine, Soheil Feizi, Tom Goldstein
NeurIPS 2020 [link], arXiv, code.
Detection as Regression: Certified Object Detection by Median Smoothing
Ping-yeh Chiang, Michael J. Curry, Ahmed Abdelkader, Aounon Kumar, John Dickerson, Tom Goldstein
NeurIPS 2020 [link], arXiv.
Tight Second-Order Certificates for Randomized Smoothing
Alexander Levine, Aounon Kumar, Thomas Goldstein, Soheil Feizi
Preprint 2020, arXiv.
On the cost of essentially fair clusterings
Ioana O. Bercea, Martin Groß, Samir Khuller, Aounon Kumar, Clemens Rösner, Daniel R. Schmidt, Melanie Schmidt
APPROX 2019 [link], arXiv.
Capacitated k-Center Problem with Vertex Weights
Aounon Kumar
FSTTCS 2016 [link].
Work Experience
Select Projects
Provable Robustness against Wasserstein Shifts: I developed a robustness certificate for the performance of machine learning models under shifts in the input distribution such as RGB shifts, hue shifts, and brightness/saturation changes. It is an efficient technique that can certify neural networks that are several layers deep.
Certified Reinforcement Learning: In this work, presented at ICLR 2022, I proved robustness guarantees for an RL agent that randomizes its observations of the environment before passing them through the policy network. The robustness certificate guarantees that the total reward obtained by the agent under an adversarial attack remains above a certain threshold.
Certified Robustness beyond Classification: One of the objectives of my research is to extend provable robustness beyond classifier outputs to more complex outputs like images, segmentation masks, and abstract latent representations. In NeurIPS 2021, I presented a procedure for certifying such structured outputs under several commonly used distance metrics such as LPIPS, cosine distance, and intersection-over-union. In another work, presented at NeurIPS 2020, I develop a procedure for certifying the confidence score produced by conventional neural networks which is often used to estimate the uncertainty in their predictions.
Curse of Dimensionality: In this work, presented at ICML 2020, I studied the limitations of a popular certified robustness technique called randomized smoothing that obtains good certificates against L1 and L2-norm bounded adversaries. My work shows that it suffers from the curse of dimensionality for higher norms such as the L∞-norm. The theoretical results prove that the best possible L∞-certificate decays at the rate of O(1/√d) with the input dimensionality d, regardless of the choice of the smoothing distribution.
Contact
Department of Computer Science
IRB 2116
University of Maryland, College Park, MD 20742.
Email: aounon at umd dot edu
