Atul Sharma
PhD, Electrical & Computer Engineering, Purdue University
Federated and decentralized machine learning: robustness, privacy, personalization
I advance collaborative machine learning where many nodes jointly train models without sharing raw data. I was advised by Prof. Saurabh Bagchi and Prof. Somali Chaterji, with theoretical collaboration from Prof. Chaoyue Liu.
NeurIPS 2026Tiered Gossip Learning, accepted
US Patent ApplicationCollaborative federated learning, published Jan 2026
IEEE S&P · CVPR · AsiaCCSPeer-reviewed security and ML venues
ICML 2026 Silver ReviewerTop-tier reviewer designation
Research
My research focuses on collaborative machine learning, where multiple nodes jointly train models without sharing raw data. My early work studied Byzantine-robust aggregation in federated learning, where a central server coordinates collaboration. I then explored improving model mixing and convergence under communication constraints in distributed settings. My current work investigates fully decentralized learning protocols that let nodes autonomously discover and collaborate with peers with similar data distributions, improving personalization in non-IID environments.
Personalization in Decentralized Learning Under review, ICLR 2026
- Designed a decentralized P2P protocol that learns the latent statistical topology of non-IID networks via social triangulation, where nodes share ranked neighbor lists and train per-node embeddings with a ladder triplet loss.
- Matches validation-based baselines without a server, validation data, or cluster priors, using a four-slot explore–exploit sampler and a tent-weighted warmup anchor for stable affinity estimation.
Tiered Gossip Learning (TGL) Accepted, NeurIPS 2026
- Introduced a hybrid two-layer push-gossip-pull topology that achieves dense global model mixing and scalability by offloading aggregation to a dynamic relay backbone.
- Showed through theory and experiments that TGL matches the convergence rate of decentralized local SGD while cutting communication cost by up to 80% on non-IID vision and language tasks.
- Developed an aggregation defense against model poisoning based on a novel statistic, the flip-score, to identify malicious clients.
- Evaluated against multiple models and datasets and against an adaptive whitebox attacker with full knowledge of the defense.
- Built a framework that adapts non-directional Transformers to configure genomic error-correction tools, with an 18x inference speedup over state-of-the-art RNN models.
- Removed the need for a reference genome by using perplexity as a fitness measure, staying within 0.31% of the theoretical optimum alignment rate on Illumina and Nanopore datasets.
Publications & Patent
- PatentSystem and Methods for Collaborative Federated Learning
Saurabh Bagchi, Somali Chaterji, Atul Sharma
U.S. Patent Application No. 19/251,710 (US20260004192A1), published Jan 2026 · Google Patents - 2024LOKI: Large-scale Data Reconstruction Attack against Federated Learning through Model Manipulation
Joshua C. Zhao, Ahmed Roushdy Elkordy, Atul Sharma, Yahya H. Ezzeldin, Salman Avestimehr, Saurabh Bagchi
IEEE Symposium on Security and Privacy (S&P) · link - 2024Leak and Learn: An Attacker's Cookbook to Train Using Leaked Data from Federated Learning
Joshua C. Zhao, Ahaan Dabholkar, Atul Sharma, Saurabh Bagchi
IEEE/CVF CVPR · pp. 12247–12256 · link - 2023How to Learn Collaboratively: Federated Learning to Peer-to-Peer Learning and What's at Stake
Atul Sharma, Wei Chen, Joshua Zhao, Qiang Qiu, Saurabh Bagchi, Somali Chaterji
IEEE/IFIP DSN, Supplemental Volume · link · PDF - 2023FLAIR: Defense against Model Poisoning Attack in Federated Learning
Atul Sharma, Wei Chen, Joshua Zhao, Qiang Qiu, Saurabh Bagchi, Somali Chaterji
ACM Asia Conference on Computer and Communications Security (AsiaCCS) · link · code - 2023The Resource Problem of Using Linear Layer Leakage Attack in Federated Learning
Joshua C. Zhao, Ahmed Roushdy Elkordy, Atul Sharma, Yahya H. Ezzeldin, Salman Avestimehr, Saurabh Bagchi
IEEE/CVF CVPR · pp. 3974–3983 · link - 2022Lerna: transformer architectures for configuring error correction tools for short- and long-read genome sequencing
Atul Sharma, Pranjal Jain, Ashraf Mahgoub, Zihan Zhou, Kanak Mahadik, Somali Chaterji
BMC Bioinformatics 23(1) · link · code - 2021Power-efficient spike sorting scheme using analog spiking neural network classifier
AK Mukhopadhyay, Atul Sharma, Indrajit Chakrabarti, Arindam Basu, Mrigank Sharad
ACM Journal on Emerging Technologies in Computing Systems (JETC) 17(2) · link
Industry Research & Engineering
Google · Software Engineering Intern
- Trained ML models on telemetry to estimate SSD flash wear-out, extending hardware lifespan in Google data centers.
- Built an end-to-end data pipeline reaching 80% precision and 90% recall in predicting critical bad-block growth.
NVIDIA · Applied Power Architecture Intern
- Implemented Implicit Q-Learning to automate CPU/GPU operating-point selection under power limits.
- Achieved a 16x inference speedup with no accuracy loss by replacing existing ML models with an optimized RL agent.
NVIDIA · Applied Power Architecture Intern
- Extended Dynamic Boost 2.0 to non-graphics workloads to optimize power allocation between the CPU and GPU.
- Developed a sequence alignment algorithm to identify identical workloads across operating points, building a high-fidelity dataset for performance prediction.
NVIDIA · ASIC Engineer Intern
- Built a generic sequence monitor and custom testbench to verify power-gating features in GPUs, then integrated it into the full-chip testbench.
Education
Purdue University · Ph.D., Electrical and Computer Engineering
IIT Kharagpur · B.Tech., Electronics & Electrical Communication Engineering (Honors)
Data Security & PrivacySecurity AnalyticsDistributed Computing Systems
Computer Network SystemsFault Tolerant Computer System Design
Artificial IntelligenceOperating SystemsMachine LearningDeep Learning