I am 3rd-year PhD candidate at NYU Centre for Cybersecurity, New York University and advised by Prof. Brandon Reagen. I work at the intersection of deep learning, applied cryptography (precisely, homomorphic encryption and multiparty computation), and computer architecture as a part of DPRIVE and ADA projects. My primary research focus is to co-design deep neural networks and cryptographic primitives to achieve real-time inference on encrypted inputs.
Before joining ECE@NYU, I completed M.Tech. (RA) from CSE IIT Hyderabad where I worked on co-optimization for efficient deep learning. Before this, I worked as an electrical design engineer on a team of solid-state drive development at Seagate Technology Bangalore (INDIA), for two years. Prior to this, I was a project research assistant at IIT Bombay and completed B.Tech. from ECE@NIT Surat.
Ph.D. in Privacy-preserving Deep Learning, 2020 - present
New York University
M.Tech. (Research Assistant) in Computer Science and Engineering, 2017 - 2020
Indian Institute of Technology Hyderabad
B.Tech. in Electronics and Communication Engineering, 2009 - 2013
National Institute of Technology Surat
We co-design Garbled circuit and stochastic ReLUs exploiting the fault tolerance of DNNs for fast private inference
We expose the pitfalls of current practice in private inference optimization and showed that most of them applicable only to zero arrival request and failed even at a moderate arrival
We investigate the “escaping activations” problem in low-degree polynomial substitution of ReLUs for faster private inference and proposed Quadratic Imitation Learning
DeepReDuce is the first work which introduces criticality based ReLU dropping for fast private inference
We proposed a data-type aware arithmetic intensity for estimating energy efficiency of DNNs at design time
We demonstrated how side-channel attacks can be used to decipher the architecture of basic building blocks in DNNs
We proposed DNN co-optimization techniques to circumvent the low PE utilization in depthwise convolution
We proposed light-weight subspace attention module for compact DNNs
We proposed group convolution techniques for energy-efficient and accurate DNNs
We showed the design ramifications of FLOPs and parameter optimized DNNs
Circa reduces the runtime overhead of ReLU operation by 1.9x by decoupling the sign evaluation and multiplication steps in the Garbled circuit with no loss in accuracy. Further, it achieves a total of 4.7x runtime reduction by employing the sign approximation in the Garbled circuit by leveraging the error-tolerant properties of neural networks within a 1% accuracy margin.
DeepReDuce is a set of optimizations for the judicious removal of ReLUs to reduce private inference latency by leveraging the ReLUs heterogeneity in classical networks. DeepReDuce strategically drops ReLUs upto 4.9x (on CIFAR-100) and 5.7x (on TinyImageNet) for ResNet18 with no loss in accuracy. Compared to the state-of-the-art for private inference DeepReDuce improves accuracy and reduces ReLU count by up to 3.5% (iso-ReLU) and 3.5×(iso-accuracy), respectively.