A
Anvith Thudi
Researcher at University of Toronto
Publications - 11
Citations - 68
Anvith Thudi is an academic researcher from University of Toronto. The author has contributed to research in topics: Computer science & Stochastic gradient descent. The author has an hindex of 1, co-authored 5 publications receiving 7 citations.
Papers
More filters
Proceedings ArticleDOI
Proof-of-Learning: Definitions and Practice
Hengrui Jia,Mohammad Yaghini,Christopher A. Choquette-Choo,Natalie Dullerud,Anvith Thudi,Varun Chandrasekaran,Nicolas Papernot +6 more
TL;DR: In this paper, the authors introduce the concept of proof-of-learning in machine learning and demonstrate how a seminal training algorithm accumulates secret information due to its stochasticity.
Journal ArticleDOI
Selective Classification Via Neural Network Training Dynamics
TL;DR: This work instantiates a method that tracks when the label predicted during training stops disagreeing with the final predicted label, and achieves state-of-the-art accuracy/coverage trade-offs on typical selective classification benchmarks.
Journal Article
Bounding Membership Inference
TL;DR: This paper provides a tighter bound on the positive accuracy of any MI adversary when a training algorithm provides (cid:15) -DP or ( (cID:15), δ )-DP, and informs the design of a novel privacy amplification scheme, where an effective training set is sub-sampled from a larger set prior to the beginning of training, to greatly reduce the bound on MI accuracy.
Posted Content
Unrolling SGD: Understanding Factors Influencing Machine Unlearning
TL;DR: In this paper, the authors taxonomize approaches and metrics of approximate unlearning and identify verification error, i.e., the L 2 difference between the weights of an approximately unlearned and a naively retrained model, as a metric approximate un learning should optimize for as it implies a large class of other metrics.
Journal ArticleDOI
On the Fundamental Limits of Formally (Dis)Proving Robustness in Proof-of-Learning
Cong Fang,He Jia,Anvith Thudi,Mohammad Yaghini,Christopher A. Choquette-Choo,Natalie Dullerud,Varun Chandrasekaran,Nicolas Papernot +7 more
TL;DR: It is shown that, until the aforementioned open problems are addressed, relying more heavily on cryptography is likely needed to formulate a new class of PoL protocols with formal robustness guarantees, and that establishing precedence robustly also reduces to an open problem in learning theory.