I am happy to talk if you need my help. I am always looking for challenging, innovative work so feel free to contact me if you know any such opportunity. Email should be the best way to reach me.
Email ID: firstname.lastname@example.org×
I have a number of ongoing projects and I am always happy to collaborate on new exciting ideas.
This project aims to develop novel, efficient (less than 2x overhead compared to no security) to train and predict a deep neural network (DNN) over shared data. Our MPC protocol uses simple modular arithmetic to enable efficient protocols for training and inference. This approach offers a promising solution for a number of distributed ML use cases.
In this project, we synthesize emerging directions of research at the intersection of differential privacy and cryptography. Broadly, we explore the use of cryptographic primitives such as SMC, anonymous communication and oblivious computation to provide alternative deployment models for DP along with differentially private relaxations of these cryptographic pirmitives.
State-of-the-art mechanisms for oblivious RAM (ORAM) suffer from significant bandwidth overheads that impact the throughput and latency of memory accesses. The lack of low-bandwidth ORAMs, despite considerable efforts from the security community, is an undeniable indicator that we need a fundamentally new approach and this is what plan to achieve in this project.
The Trusted Cloud project at Microsoft Research aims to provide customers of cloud computing complete control over their data: no one should be able to access the data without the customer’s permission. Even if there are malicious employees in the cloud service provider, or hackers break into the data center, they still should not be able to get access to customer data.
Closely related to the Oblivious RAM and Oblivious Transfer, PIR is a cryptographic primitive to make private access to public data. Developing novel, exciting and practical protocol for PIR is what this project focuses on. More specifically, I am looking into developing efficient computational PIR techniques applying tools from cryptography to get performance improvements.
Machine-learning algorithms are increasingly utilized in privacy-sensitive applications and attacks have shown that such classifiers can contain sensitive information about the pool of users. This brings us to a natural question, how much information is contained in a classifier? In this work, we try to address this question. We rely on an information theoretic analysis for estimating the mutual information between the model and the data.
This project revolves around developing technologies for crowdsourcing statistics from users anonymously with provable privacy guarantees. We aim to build this system over the Ethereum platform combining from various ideas from applied cryptography with mechanics of the blockchain. The end goal is to enable large scale surveying of accurate yet anonymous statistics from a large set of users.