I am a PhD candidate in Electrical Engineering at Princeton University. I build trustworthy machine learning systems, especially using generative models.
I am advised by Prateek Mittal and Mung Chiang. Before coming to Princeton, I completed my undergraduate at IIT Kharagpur, India. I have previously interned at Microsoft Research and Meta AI. I have also been fortunate to receive Qualcomm Innovation Fellowship.
Major update: I am leading the organization of virtual seminar series on Security & Privacy in Machine Learning (SPML).
I am on the job market looking for research-scientist/post-doc positions.
Extracting Training Data from Diffusion Models
We show that modern diffusion models, such as Stable-diffusion and ImageN, memorize certain training images, which can be extracted by an adversary during sampling.
A Light Recipe to Train Robust Vision Transformers
Contrary to the conventional wisdom of using heavy data augmentation in ViTs, we show that a lighter data augmentation (along with other bag-of-tricks) achieves state-of-the-art performance with ViTs adversarial training.
Generating High Fidelity Data from Low-density Regions using Diffusion Models
We improve the sampling process of diffusion models to generate high fidelity hard, i.e., from low-density regions, synthetic images.
Understanding Robust Learning through the Lens of Representation Similarities
Using representation similarity metrics, such as CKA, we demonstrate multiple interesting characteristics of adversarially robust networks compared to non-robust networks.
Robust Learning Meets Generative Models: Can Proxy Distributions Improve Adversarial Robustness?
We show that synthetic data from diffusion model provides a termendous boost in generalization performance of robust training.
Lower Bounds on Cross-Entropy Loss in the Presence of Test-time Adversaries
We provide lower-bounds on cross-entropy loss in persence of adversarial attacks on basic vision datasets.
SSD: A Unified Framework for Self-Supervised outlier detection
ICLR 2021, Short version accepted at NeurIPS SSL workshop, 2020
Using only unlabeled data, we develop a highly succesful framework to detect outliers or out-of-distribution samples.
RobustBench: A Standardized Adversarial Robustness Benchmark
Project-page / Pdf / Code
We provide a leaderboard to track progress + a library for unified access to SOTA defenses against adversarial examples.
Time for a Background Check! Uncovering the impact of Background Features on Deep Neural Networks
ICML workshop on Object-Oriented Learning, 2020
Project-page / Pdf / Bibtex / Slides / Video
We investigate background invariance and influence over 32 deep neural networks on ImageNet dataset.
On Separability of Self-Supervised Representations
ICML workshop on Uncertainty & Robustness in Deep Learning, 2020
We compare the representations learned by several self-supervised methods with supervised networks.
HYDRA: Pruning Adversarially Robust Neural Networks
NeurIPS 2020, Short paper in ICLR workshop on Trustworthy Machine Learning, 2020
Project-page / Pdf / Bibtex / Code / Slides / Video
We achieve state-of-the-art accuracy and robustness for pruned networks (pruning up to 100x).
PatchGuard: Provable Defense against Adversarial Patches Using Masks on Small Receptive Fields
Project-page / Pdf / Bibtex
A general defense framework to acheive provable robustness against adversrial patches.
Fast-Convergent Federated Learning
To appear in IEEE Journal on Selected Areas in Communications (J-SAC) - Series on Machine Learning for Communications and Networks
Project-page / Pdf / Bibtex
We proposed a fast-convergent federated learning algorithm, called FOLB, which improves convergence speed by an intelligent sampling of devices in each round.
A Critical Evaluation of Open-World Machine Learning
ICML Workshop on Uncertainty & Robustness in Deep Learning , 2020
We discover a conflict between the objective of open-world machine learning and adversarial robustness.
Analyzing the Robustness of Open-World Machine Learning
ACM Workshop on Artificial Intelligence and Security (AISec), 2019
We demonstrate the vulnerability of open-world ML to adversarial examples and proposed a defense.
Research Work in Undergraduate
Teaching and Mentoring