REVIEW 2 major objections 6 minor 41 references
Faster Rates for Private Adversarial Bandits
T0 review · 2 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read A simple batching trick converts any non-private bandit algorithm into a private one, cutting regret from O(√(KT log(KT))/ε) to O(√(KT)/√ε).
desk verdict A strong, likely-correct paper showing a clean batching conversion that improves private adversarial bandit regret to O(sqrt(KT)/sqrt(epsilon)); the main gaps are verification details, not errors. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the 'batched Laplace mechanism' conversion (Algorithm 1). It reduces the sensitivity of each released query by averaging over a batch of τ losses before adding Laplace noise of scale 1/(ετ). The overload-bearing property is that the base algorithm B must have low regret on losses that are unbiased but may be negative and unbounded—precisely the class of 'truncated non-negative, heavy-tailed' losses for which the HTINF algorithm of Huang et al. [2022] supplies a regret bound. This combination of batching with a heavy-tailed-regret guarantee is what converts a non-private bandit algorithm into a private one with only a √ε penalty in the regret.
What would settle it
Run the conversion (Algorithm 1) with a base algorithm B that is not known to handle unbounded negative losses—e.g., standard EXP3 without the heavy-tailed analysis—and check whether the empirical regret degrades on a constructed sequence where all losses are negative and unbounded. A concrete private test: simulate the batched conversion on a two-arm game with losses 0 and 1, but add Laplace noise of scale $10^{6}$ to the batched averages; if the base algorithm's regret grows polynomially in the noise scale, the conversion's utility lemma fails for that B, showing the assumption about the base algorithm is not automatic.
Extended reading notes
Core claim
The central claim is that privacy in adversarial bandits can be obtained 'for free' up to a √ε factor on top of the non-private regret, provided the underlying non-private algorithm can tolerate noisy, possibly negative, unbounded losses. The conversion (Algorithm 1) is simple: for each batch of τ rounds, play the same arm, compute the empirical average loss, add Laplace noise of scale 1/(ετ), and give only these noisy batched losses to the base algorithm B. Theorem 1 shows that the resulting algorithm is ε-differentially private and has regret at most τ times the regret of B on the batched noisy sequence, plus τ. Instantiating with the heavy-tailed FTRL algorithm HTINF gives the headline bound O(√(KT)/√ε + 1/ε) for all ε ∈ [1/T, 1], which is sublinear even when ε is as small as 1/√T. The same machinery, via a standard bandit-to-experts reduction, produces the first private algorithms for bandits with expert advice, with three incomparable rates covering different regimes of action count K, expert count N, and privacy ε.
Load-bearing premise
The entire O(√(KT)/√ε) upper bound rests on the external guarantee that the HTINF algorithm (or another non-private base algorithm) has low expected regret even when the observed losses are unbounded, negative, and heavy-tailed after adding Laplace noise; the paper does not prove that guarantee itself.
Editorial extensions
If this is right
- For adversarial bandits, ε-differential privacy costs only a √(1/ε) factor: regret O(√(KT)/√ε + 1/ε), improving the old O(√(KT log(KT))/ε) for every ε ≤ 1.
- Sublinear regret is achievable for every ε = ω(1/T), including ε ≤ 1/√T, which is impossible under local differential privacy; this gives the first central-versus-local separation for adversarial bandits.
- For bandits with expert advice, the paper gives the first differentially private algorithms, with regret O(√(NT)/√ε), O(√(KT log N log(KT))/ε), and O(N^{1/6}K^{1/2}T^{2/3} log(NT)/ε^{1/3} + N^{1/2} log(NT)/ε), covering high-privacy and large-N regimes.
- The conversion is generic: any future non-private bandit algorithm that handles unbounded negative losses with regret R would automatically yield a private algorithm with regret O(R/ε) via the same reduction.
- The paper also shows a lower bound for a natural class of algorithms (including EXP3 and batched variants): such algorithms cannot achieve additive separation between T and ε, suggesting the √ε penalty may be inherent for standard techniques.
Reading between the lines
- The technique suggests that heavy-tailed robustness and differential privacy are more closely related than previously recognized: any algorithm that is stable under the addition of unbounded zero-mean noise can be made private by batching, so improving heavy-tailed bandit algorithms directly improves private bandit algorithms.
- The central-versus-local separation for adversarial bandits mirrors analogous separations in other online learning settings and hints that the privacy cost in central DP may be fundamentally smaller than in the local model for sequential decision problems, not just for statistical estimation.
- A natural testable extension would be to instantiate the conversion with the best-known non-private minimax-optimal bandit algorithms (for the adversarially corrupted or heavy-tailed settings) to see whether the √ε dependence can be sharpened or matched by a lower bound for central DP.
- The expert-advice results suggest that the batching idea can be combined with the EXP4-type reductions to handle large expert classes; the open regime N ≥ T at ε ≤ 1/√T may be approachable with privacy-preserving aggregation of expert weights, which the current batching does not directly address.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper studies differentially private adversarial bandits and bandits with expert advice. The main technical tool is a black-box conversion (Theorem 1): any non-private bandit algorithm B with bounded regret on unbounded, negative, heavy-tailed losses can be converted into an ε-differentially private algorithm by playing in batches of size τ and feeding B the batched average loss plus Laplace noise of scale 1/(ετ); the privacy proof (Lemma 3) is a direct pairwise indistinguishability argument, and the utility proof (Lemma 4) applies a batched-regret transfer. Instantiating the conversion with HTINF yields the advertised O(√(TK)/√ε) regret (Corollary 2), and with EXP3 and FTPL gives log-factor variants (Corollaries 22 and 24). For bandits with expert advice, the paper gives three algorithms: an HTINF-based conversion with O(√(TN)/√ε) (Corollary 8), a locally private EXP4 with O(√(TK log N) log(KT)/ε) (Theorem 9), and a batched EXP4 with O(N^{1/6}K^{1/2}T^{2/3}/ε^{1/3} + N^{1/2}/ε) (Theorem 10). Section 5 proves a lower bound for a class of private algorithms that quickly concentrate on good arms, suggesting Ω(√(T/ε)) for such algorithms.
Significance. If the two concerns below are resolved, the paper makes a significant contribution: it establishes the first separation between central and local differential privacy for adversarial bandits, shows sublinear regret for ε as small as ω(1/T), and provides the first differentially private algorithms for bandits with expert advice in several parameter regimes. The conversion argument is conceptually clean and likely reusable. Strengths include a direct privacy proof instead of a loose composition bound, fully worked EXP3 and FTPL instantiations with explicit parameters and constants, and an honest discussion of an unverified prior claim in Appendix G.1. The main caveats are the reliance on an external heavy-tailed regret theorem for the headline log-free rate and an under-specified privacy analysis in Theorem 10.
major comments (2)
- [Appendix D, Theorem 21 and Corollary 2] The headline log-free rate O(√(KT)/√ε) in Corollary 2 rests entirely on Theorem 21, an external regret bound for HTINF attributed to Huang et al. [2022] that is neither proved nor reproduced in the manuscript. The hypotheses of the original Theorem 4.1 in Huang et al. apparently concern heavy-tailed stochastic bandits, and the manuscript does not verify that the adversarial, non-stationary batched noisy-loss sequence fed to B by Algorithm 1 satisfies the original theorem's definition of regret, in particular whether the comparison is to the expected best arm Σ_t E[\tildeℓ_t(i)] or to the realized best arm. If the original guarantee does not hold for arbitrary sequences of random losses with per-round truncated non-negativity and (α,σ)-heavy-tailed moments, the derivation of Corollary 2 collapses. Because this corollary is the source of the claimed improvement over the prior O(√(KT log(KT))/ε) bound, the authors must either provide a self-contained proof of Theorem 21 for the exact loss model used by Algorithm 1, or quote the precise original statement and demonstrate that its hypotheses and comparator match the application. The internally proved EXP3 and FTPL conversions (Corollaries 22 and 24) give the same √ε scaling up to logarithmic factors, so the central/local separation could still survive, but the exact advertised rate would not be established.
- [Appendix F.4, privacy proof of Theorem 10] The privacy analysis for Algorithm 2 is not justified as written. The proof states that each coordinate of the batched query has sensitivity at most K/(γτ) and then concludes from the Laplace mechanism and advanced composition that the vector output is (ε,δ)-differentially private. However, the vector's L1 sensitivity is N·K/(γτ), while the Laplace scale used, λ = 3K√(N log(1/δ))/(γτε), is a factor of √(N/log(1/δ)) too small for a standard L1-vector Laplace mechanism, and no advanced-composition argument over the N coordinates is supplied. If the intended mechanism is the L2-sensitivity Laplace or Gaussian mechanism, the proof must state the relevant sensitivity and tail bound explicitly and verify the constants. As written, this step does not establish the stated (ε,δ) guarantee, and since Theorem 10 is one of the three advertised expert-advice results, this is a load-bearing gap.
minor comments (6)
- [Algorithm 6 (Theorem 9)] Line 7 of Algorithm 6 defines ℓ'_t(i) for every i∈[K], but the algorithm only observes ℓ_t(I_t). The pseudocode should define ℓ'_t(I_t) := ℓ_t(I_t) + Z^{I_t}_t, and the proof's estimator should be written accordingly; as written, the algorithm appears to require access to unobserved losses.
- [Appendix A] Lemma 14 duplicates Lemma 13: the 'Basic Composition' statement appears twice with identical content.
- [Throughout] There are several typos: 'worse-case' in Corollaries 2 and 24, 'an action action' in Definition 2, and 'the our notion' in Section 2.4. These should be corrected.
- [Section 5.2] The verification that EXP3 and its batched variants satisfy conditions (1) and (2) of Lemma 11 is only sketched; adding a formal argument with explicit constants, or explicitly labeling the claim as heuristic, would make the barrier section self-contained.
- [Abstract and Corollary 2] The abstract's phrase 'improving upon the existing upper bound ... for all ε ≤ 1' is only literally true for the HTINF instantiation of Corollary 2; the internally proved EXP3 and FTPL conversions give the improvement only up to logarithmic factors. If Theorem 21 is not reproved, the abstract should be adjusted to reflect the log-factor variants.
- [Section 2.4] The statement that the privacy guarantee holds 'against adaptive adversaries' while utility is only for oblivious adversaries is interesting and should be made precise with a definition of adaptive-adversary privacy, since Definition 3 is stated only for oblivious loss sequences.
Circularity Check
No circularity: the main upper bound follows from the paper's own batched-Laplace conversion plus an external heavy-tailed regret bound for HTINF; self-citations are limited to secondary comparison corollaries.
full rationale
The derivation chain is self-contained against external benchmarks. Theorem 1 is proven from a batched-regret lemma (Appendix C) using Arora et al.'s batching theorem and a Laplace-mechanism privacy argument; neither assumes the target regret bound. Corollary 2 instantiates Theorem 1 with HTINF, and the only load-bearing heavy-tailed guarantee is Theorem 21, quoted from Huang et al. [2022], an independent prior work by different authors. The paper internally verifies the two required properties of the noisy losses: truncated non-negativity (Lemma 19) and (2, sqrt(2+4 lambda^2))-heavy-tailedness (Appendix D). No parameter is fitted to data, and no regret quantity is used both as input and as output. The self-citations (Asi et al. 2023 and 2024) are used for the adaptive-adversary separation corollary and for comparisons; they are not used to prove the main private-bandit upper bound, and the central central-vs-local privacy separation rests on the independent lower bound of Basu et al. [2019]. Therefore no circular step is present; the unverified status of Huang et al.'s Theorem 4.1 is a correctness or verification risk, not circularity.
Assumptions & free parameters
free parameters (3)
- batch size tau =
ceil(1/epsilon)
- learning rate eta (EXP3 conversion) =
sqrt(log(K) / (22 epsilon K T log^2(epsilon K T)))
- mixing parameter gamma (EXP3 conversion) =
4 eta K log(epsilon K T)
assumptions (5)
- domain assumption HTINF regret bound (Huang et al. 2022, Theorem 4.1): regret <= 30 sigma K^{1-1/alpha} (T+1)^{1/alpha} for truncated non-negative heavy-tailed losses.
- domain assumption Arora et al. 2012 Theorem 2: batching a bandit algorithm with batch size tau increases regret by at most tau R_B(T/tau) + tau.
- domain assumption Local differential privacy lower bound: any epsilon-LDP bandit algorithm has regret Omega(T) when epsilon < 1/sqrt(T) (Basu et al. 2019).
- standard math Standard differential privacy composition theorems and Laplace mechanism (Dwork et al. 2014).
- domain assumption Oblivious adversary assumption: the loss sequence is chosen before the game begins.
Cite this review
Pith. "Pith review of Faster Rates for Private Adversarial Bandits." pith.science (2026). https://pith.science/paper/AGX46EE7
@misc{pith2026250521790,
author = {Pith},
title = {Pith review of: Faster Rates for Private Adversarial Bandits},
year = {2026},
howpublished = {\url{https://pith.science/paper/AGX46EE7}},
note = {Machine review of arXiv:2505.21790}
}
abstract
We design new differentially private algorithms for the problems of adversarial bandits and bandits with expert advice. For adversarial bandits, we give a simple and efficient conversion of any non-private bandit algorithm to a private bandit algorithm. Instantiating our conversion with existing non-private bandit algorithms gives a regret upper bound of $O\left(\frac{\sqrt{KT}}{\sqrt{\epsilon}}\right)$, improving upon the existing upper bound $O\left(\frac{\sqrt{KT \log(KT)}}{\epsilon}\right)$ for all $\epsilon \leq 1$. In particular, our algorithms allow for sublinear expected regret even when $\epsilon \leq \frac{1}{\sqrt{T}}$, establishing the first known separation between central and local differential privacy for this problem. For bandits with expert advice, we give the first differentially private algorithms, with expected regret $O\left(\frac{\sqrt{NT}}{\sqrt{\epsilon}}\right), O\left(\frac{\sqrt{KT\log(N)}\log(KT)}{\epsilon}\right)$, and $\tilde{O}\left(\frac{N^{1/6}K^{1/2}T^{2/3}\log(NT)}{\epsilon ^{1/3}} + \frac{N^{1/2}\log(NT)}{\epsilon}\right)$, where $K$ and $N$ are the number of actions and experts respectively. These rates allow us to get sublinear regret for different combinations of small and large $K, N$ and $\epsilon.$
Figures
Reference graph
Works this paper leans on
-
[1]
Fighting bandits with a new kind of smoothness
Jacob D Abernethy, Chansoo Lee, and Ambuj Tewari. Fighting bandits with a new kind of smoothness. Advances in Neural Information Processing Systems, 28, 2015
2015
-
[2]
The price of differential privacy for online learning
Naman Agarwal and Karan Singh. The price of differential privacy for online learning. In International Conference on Machine Learning, pages 32--40. PMLR, 2017
work page 2017
-
[3]
Online bandit learning against an adaptive adversary: from regret to policy regret
Raman Arora, Ofer Dekel, and Ambuj Tewari. Online bandit learning against an adaptive adversary: from regret to policy regret. arXiv preprint arXiv:1206.6400, 2012
arXiv 2012
-
[4]
Private online prediction from experts: Separations and faster rates
Hilal Asi, Vitaly Feldman, Tomer Koren, and Kunal Talwar. Private online prediction from experts: Separations and faster rates. In The Thirty Sixth Annual Conference on Learning Theory, pages 674--699. PMLR, 2023
work page 2023
-
[5]
Private Online Learning via Lazy Algorithms
Hilal Asi, Tomer Koren, Daogao Liu, and Kunal Talwar. Private online learning via lazy algorithms. arXiv preprint arXiv:2406.03620, 2024
work page Pith review arXiv 2024
-
[6]
Minimax policies for adversarial and stochastic bandits
Jean-Yves Audibert and S \'e bastien Bubeck. Minimax policies for adversarial and stochastic bandits. In COLT, pages 217--226, 2009
2009
-
[7]
The nonstochastic multiarmed bandit problem
Peter Auer, Nicolo Cesa-Bianchi, Yoav Freund, and Robert E Schapire. The nonstochastic multiarmed bandit problem. SIAM journal on computing, 32 0 (1): 0 48--77, 2002
2002
-
[8]
When privacy meets partial information: A refined analysis of differentially private bandits
Achraf Azize and Debabrota Basu. When privacy meets partial information: A refined analysis of differentially private bandits. Advances in Neural Information Processing Systems, 35: 0 32199--32210, 2022
work page 2022
Show all 41 references
-
[9]
Differential privacy for multi-armed bandits: What is it and what is its cost? arXiv preprint arXiv:1905.12298, 2019
Debabrota Basu, Christos Dimitrakakis, and Aristide Tossou. Differential privacy for multi-armed bandits: What is it and what is its cost? arXiv preprint arXiv:1905.12298, 2019
1905 arXiv
-
[10]
Regret analysis of stochastic and nonstochastic multi-armed bandit problems
S \'e bastien Bubeck, Nicolo Cesa-Bianchi, et al. Regret analysis of stochastic and nonstochastic multi-armed bandit problems. Foundations and Trends in Machine Learning , 5 0 (1): 0 1--122, 2012
2012
-
[11]
Prediction, learning, and games
Nicolo Cesa-Bianchi and G \'a bor Lugosi. Prediction, learning, and games. Cambridge university press, 2006
2006
-
[12]
Follow-the-perturbed-leader for adversarial bandits: Heavy tails, robustness, and privacy
Duo Cheng, Xingyu Zhou, and Bo Ji. Follow-the-perturbed-leader for adversarial bandits: Heavy tails, robustness, and privacy
-
[13]
The price of bandit information in multiclass online classification
Amit Daniely and Tom Helbertal. The price of bandit information in multiclass online classification. In Conference on Learning Theory, pages 93--104. PMLR, 2013
2013
-
[14]
Local privacy, data processing inequalities, and statistical minimax rates
John C Duchi, Michael I Jordan, and Martin J Wainwright. Local privacy, data processing inequalities, and statistical minimax rates. arXiv preprint arXiv:1302.3203, 2013
2013 arXiv
-
[15]
Differential privacy
Cynthia Dwork. Differential privacy. In International colloquium on automata, languages, and programming, pages 1--12. Springer, 2006
2006
-
[16]
Differential privacy under continual observation
Cynthia Dwork, Moni Naor, Toniann Pitassi, and Guy N Rothblum. Differential privacy under continual observation. In Proceedings of the forty-second ACM symposium on Theory of computing, pages 715--724, 2010 a
2010
-
[17]
Boosting and differential privacy
Cynthia Dwork, Guy N Rothblum, and Salil Vadhan. Boosting and differential privacy. In 2010 IEEE 51st Annual Symposium on Foundations of Computer Science, pages 51--60. IEEE, 2010 b
2010
-
[18]
The algorithmic foundations of differential privacy
Cynthia Dwork, Aaron Roth, et al. The algorithmic foundations of differential privacy. Foundations and Trends in Theoretical Computer Science , 9 0 (3--4): 0 211--407, 2014
2014
-
[19]
(nearly) optimal algorithms for private online learning in full-information and bandit settings
Abhradeep Guha Thakurta and Adam Smith. (nearly) optimal algorithms for private online learning in full-information and bandit settings. Advances in Neural Information Processing Systems, 26, 2013
2013
-
[20]
Follow-the-perturbed-leader achieves best-of-both-worlds for bandit problems
Junya Honda, Shinji Ito, and Taira Tsuchiya. Follow-the-perturbed-leader achieves best-of-both-worlds for bandit problems. In International Conference on Algorithmic Learning Theory, pages 726--754. PMLR, 2023
2023
-
[21]
Optimal algorithms for private online learning in a stochastic environment
Bingshan Hu, Zhiming Huang, and Nishant A Mehta. Optimal algorithms for private online learning in a stochastic environment. arXiv preprint arXiv:2102.07929, 2021
2021 arXiv
-
[22]
Adaptive best-of-both-worlds algorithm for heavy-tailed multi-armed bandits
Jiatai Huang, Yan Dai, and Longbo Huang. Adaptive best-of-both-worlds algorithm for heavy-tailed multi-armed bandits. In international conference on machine learning, pages 9173--9200. PMLR, 2022
2022
-
[23]
(near) dimension independent risk bounds for differentially private learning
Prateek Jain and Abhradeep Guha Thakurta. (near) dimension independent risk bounds for differentially private learning. In International Conference on Machine Learning, pages 476--484. PMLR, 2014
2014
-
[24]
Differentially private online learning
Prateek Jain, Pravesh Kothari, and Abhradeep Thakurta. Differentially private online learning. In Conference on Learning Theory, pages 24--1. JMLR Workshop and Conference Proceedings, 2012
2012
-
[25]
The composition theorem for differential privacy
Peter Kairouz, Sewoong Oh, and Pramod Viswanath. The composition theorem for differential privacy. In International conference on machine learning, pages 1376--1385. PMLR, 2015
2015
-
[26]
Regret bounds for sleeping experts and bandits
Robert Kleinberg, Alexandru Niculescu-Mizil, and Yogeshwer Sharma. Regret bounds for sleeping experts and bandits. Machine learning, 80 0 (2): 0 245--272, 2010
2010
-
[27]
The weighted majority algorithm
Nick Littlestone and Manfred K Warmuth. The weighted majority algorithm. Information and computation, 108 0 (2): 0 212--261, 1994
1994
-
[28]
Bandit algorithms for precision medicine
Yangyi Lu, Ziping Xu, and Ambuj Tewari. Bandit algorithms for precision medicine. arXiv preprint arXiv:2108.04782, 2021
2021 arXiv
-
[29]
(nearly) optimal differentially private stochastic multi-arm bandits
Nikita Mishra and Abhradeep Thakurta. (nearly) optimal differentially private stochastic multi-arm bandits. In Proceedings of the Thirty-First Conference on Uncertainty in Artificial Intelligence, pages 592--601, 2015
2015
-
[30]
Mitigating bias in adaptive data gathering via differential privacy
Seth Neel and Aaron Roth. Mitigating bias in adaptive data gathering via differential privacy. In International Conference on Machine Learning, pages 3720--3729. PMLR, 2018
2018
-
[31]
Importance weighting without importance weights: An efficient algorithm for combinatorial semi-bandits
Gergely Neu and G \'a bor Bart \'o k. Importance weighting without importance weights: An efficient algorithm for combinatorial semi-bandits. Journal of Machine Learning Research, 17 0 (154): 0 1--21, 2016
2016
-
[32]
Multiclass online learnability under bandit feedback
Ananth Raman, Vinod Raman, Unique Subedi, Idan Mehalel, and Ambuj Tewari. Multiclass online learnability under bandit feedback. In International Conference on Algorithmic Learning Theory, pages 997--1012. PMLR, 2024
2024
-
[33]
Multi-armed bandits with local differential privacy
Wenbo Ren, Xingyu Zhou, Jia Liu, and Ness B Shroff. Multi-armed bandits with local differential privacy. arXiv preprint arXiv:2007.03121, 2020
2007 arXiv
-
[34]
Estimation of regression coefficients when some regressors are not always observed
James M Robins, Andrea Rotnitzky, and Lue Ping Zhao. Estimation of regression coefficients when some regressors are not always observed. Journal of the American statistical Association, 89 0 (427): 0 846--866, 1994
1994
-
[35]
An optimal private stochastic-mab algorithm based on optimal private stopping rule
Touqir Sajed and Or Sheffet. An optimal private stochastic-mab algorithm based on optimal private stopping rule. In International Conference on Machine Learning, pages 5579--5588. PMLR, 2019
2019
-
[36]
A lower bound for multi-armed bandits with expert advice
Yevgeny Seldin and G \'a bor Lugosi. A lower bound for multi-armed bandits with expert advice. In 13th European Workshop on Reinforcement Learning (EWRL), volume 2, page 7, 2016
2016
-
[37]
Differentially private contextual linear bandits
Roshan Shariff and Or Sheffet. Differentially private contextual linear bandits. Advances in Neural Information Processing Systems, 31, 2018
2018
-
[38]
Differentially private multi-armed bandits in the shuffle model
Jay Tenenbaum, Haim Kaplan, Yishay Mansour, and Uri Stemmer. Differentially private multi-armed bandits in the shuffle model. Advances in Neural Information Processing Systems, 34: 0 24956--24967, 2021
2021
-
[39]
Algorithms for differentially private multi-armed bandits
Aristide Tossou and Christos Dimitrakakis. Algorithms for differentially private multi-armed bandits. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 30, 2016
2016
-
[40]
Achieving privacy in the adversarial multi-armed bandit
Aristide Tossou and Christos Dimitrakakis. Achieving privacy in the adversarial multi-armed bandit. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 31, 2017
2017
-
[41]
Locally differentially private (contextual) bandits learning
Kai Zheng, Tianle Cai, Weiran Huang, Zhenguo Li, and Liwei Wang. Locally differentially private (contextual) bandits learning. Advances in Neural Information Processing Systems, 33: 0 12300--12310, 2020
2020
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.