By batching losses and using heavy-tailed bandit algorithms, any non-private adversarial bandit algorithm can be made epsilon-differentially private with regret O(sqrt(KT)/sqrt(epsilon)), and the first private expert-advice algorithms are given.
Private online prediction from experts: Separations and faster rates
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Faster Rates for Private Adversarial Bandits
By batching losses and using heavy-tailed bandit algorithms, any non-private adversarial bandit algorithm can be made epsilon-differentially private with regret O(sqrt(KT)/sqrt(epsilon)), and the first private expert-advice algorithms are given.