sid0nair@research— incoming MS Privacy Eng. @ CMU
~/ research

Research

My work centers on second-order federated optimization with differential-privacy guarantees, making distributed training converge faster while keeping formal privacy budgets tight. In plain terms, I'm building ways for institutions like hospitals, banks and data centres to train shared AI models together without ever handing over the private data behind them, so that as machine learning scales, personal privacy doesn't have to be the price. Below: peer-reviewed work and preprints, then the research appointments behind them.

01

Focus Areas

DP
Differential privacy — Rényi-DP accounting, subsampling amplification, (ε,δ) guarantees
FL
Federated learning over non-IID, heterogeneous clients
2nd-order
Newton-CG, L-BFGS, Fisher-information & low-rank curvature
Secure AI
Privacy-preserving deployment & domain adaptation
02

Publications & Preprints

04
Under reviewApr 2025

Accelerated Training of Federated Learning via Second-Order Methods

Under review at IEEE Transactions on Pattern Analysis and Machine Intelligence (T-PAMI)

Co-authored work detailing the convergence behaviour and communication efficiency of second-order federated-learning methods — benchmarking Newton-CG, L-BFGS and quasi-Newton schemes against first-order baselines on heterogeneous clients, with Hessian-free and low-rank approximations to cut communication overhead.

Target TMLR 2026Jan 2026

DP-FedSOFIM: Differentially Private Federated Stochastic Optimization using Regularized Fisher Information Matrix

Submitted to Transactions on Machine Learning Research (TMLR)

Privacy-preserving second-order federated optimization with (ε−δ) differential-privacy guarantees for heterogeneous data distributions. A regularized Fisher Information Matrix preconditions updates while a federated privacy accountant tracks the budget under subsampling amplification.

Under reviewAug 2025

FedDAF: Federated Domain Adaptation using Model Functional Distance

Under review — IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)

A novel federated domain-adaptation approach addressing limited target data and client distribution heterogeneity, introducing functional-distance-based aggregation that aligns target and source models via mean gradient fields — 15–20% accuracy gains over state-of-the-art FL, PFL and FDA methods on CIFAR-10, PACS and Caltech-10.

PublishedMar 2025

Towards Support-Free Printing in Extrusion-Based Additive Manufacturing: Path Planning & Process Control

Presented at ASME International Mechanical Engineering Congress & Exposition (IMECE) 2025

Optimized computational object slicing and multi-axis G-code design for support-free extrusion printing — part of an industry–university collaboration with SONY on India's first 5-axis FDM printer.

03

Research Appointments

Indian Statistical Institute (ISI), Kolkata
Jun 2025 — Present
Prof. Tanmay Sen · Statistical Quality Control & OR Unit
  • Designing second-order federated optimizers with momentum-corrected Newton steps for faster convergence.
  • Implementing a federated privacy accountant with Rényi-DP and subsampling amplification, ensuring (ε, δ)-DP.
  • Building a dimension-aware gradient-clipping framework for ResNet, stabilizing updates under noisy Hessians.
  • Achieved 8× noise reduction via subsampling — 70% accuracy, outperforming first-order baselines.
Indian Institute of Technology (IIT), Hyderabad
Oct 2024 — Apr 2025
Prof. C. Mohan · Dept. of CSE (Former Dean, Public & Corporate Relations)
  • Benchmarked convergence of Newton-CG, L-BFGS and quasi-Newton methods on heterogeneous federated clients.
  • Implemented Hessian-free and low-rank approximation schemes, reducing communication overhead by 40%+.
  • Tested diagonal and Nyström-based second-order updates for scalability in large-scale distributed optimization.
  • Developed preconditioning strategies for distributed second-order optimization under non-IID heterogeneity.
Policy-Gradient Methods in Parameterized RL
Aug 2025 — Present
B.Tech Project · Prof. Amber Srivastava, Mech. Eng., IIT Delhi
  • Extending the parameterized-MDP framework to continuous action spaces via policy-gradient methods vs. Q-learning.
  • Implementing the maximum-entropy principle with deterministic annealing in an actor-critic architecture for exploration control.