Distributed Network Tomography: Exact Recovery with Adversarial, Heterogeneous, and Sporadic Data

This lecture by Dr. Gugan Thoppe, Assistant Professor at IISc, explores distributed network tomography with a focus on exact recovery methods for adversarial, heterogeneous, and sporadic data. Learn about estimating the mean of random vectors in a distributed parameter-server–worker setting where workers may behave adversarially. Discover why traditional estimation strategies based on data encoding, robust aggregation, and homogenization fall short in network tomography applications, and explore a novel l1-minimization approach that achieves exact solutions even under challenging conditions. The presentation covers theoretical guarantees for the algorithm's convergence rate and includes empirical results demonstrating superior accuracy compared to existing methods, with potential extensions to tracking and general optimization scenarios. Dr. Thoppe, an IEEE Senior member with research experience from TIFR, Technion, and Duke University, specializes in reinforcement learning, federated learning, stochastic approximation, and random topology, and has received multiple awards for his research and teaching excellence.