Pith. sign in

REVIEW 4 minor 67 references

An auditor can publish a randomized audit schedule that gives every whistleblower report a formal per-report (0,δ)-differential-privacy guarantee, while the chance of auditing the wrong organization vanishes as reported gaps grow.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A new mechanism gives per-report (0,δ)-differential privacy for whistleblower audit transcripts, with selection error vanishing as report gaps grow faster than √log T.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection Solid theory paper: per-report DP for audit transcripts, a clean negative result for randomized response, and a correct reduction to continual counting; the few soft spots are stated limitations, not hidden flaws.

arxiv 2607.13928 v1 pith:7XNDCDV5 submitted 2026-07-15 cs.CR cs.LGstat.ML

Plausible Deniability Guarantees for Whistleblowers

Leo Richter , Matt J. Kusner This is my paper
classification cs.CR cs.LGstat.ML MSC 68P2768W20
keywords differential privacywhistleblower protectionplausible deniabilitycontinual countingauditingrandomized responseToeplitz factorization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that whistleblower protection can be made quantitative: an auditor can publish its audit decisions in a way that no audited organization can tell from the transcript whether any particular report was filed, even if that organization knows everything else and studies the mechanism. The authors formalize this as per-report (0,δ)-differential privacy on the transcript of audit selections. They prove that the classic randomized-response approach cannot deliver this: at any fixed privacy level it is forced to select almost uniformly at random, improving on uniform auditing by at most δ. They replace it with a counter-and-restart scheme in which each organization has its own private noisy report counter, and the auditor audits the organization with the highest noisy count. With a near-optimal continual-counting counter, the noise grows only as O(√log T) over T audit decisions, and the mis-selection error decays whenever the report gap to the leader grows faster than √log T.

Core claim

The central discovery is that private whistleblower auditing reduces to private continual counting: Algorithm 1 runs one (0,δ)-DP continual counter per audit target, feeds it only that target's new reports, and audits the target with the largest noisy active count, restarting the counter after each audit. Because a single report changes exactly one counter's input stream, the whole audit transcript inherits (0,δ)-DP by post-processing. Instantiating the reduction with the Toeplitz matrix-factorization counter gives per-report (0,δ)-DP with encoder sensitivity M_T = O(√log T). The utility theorem states that, conditional on history, the probability of not auditing the true leading target is a

What carries the argument

The counter-and-restart reduction (Algorithm 1) is the load-bearing object: it converts privacy of an underlying continual-counting mechanism into privacy of the audit transcript, while the restart enforces the report-reset condition that audits clear pending reports. The concrete instantiation uses a lower-triangular Toeplitz factorization of the prefix-sum matrix, whose encoder column-norm sensitivity is M_T = O(√log T); this factorization is the tightest among lower-triangular Toeplitz factorizations. The utility analysis rests on a fresh-noise decomposition (Lemma 18): at each step, only the newest Gaussian noise coordinate is unseen, so the noisy counts are exact Gaussians and the per-c

Load-bearing premise

The auditor is trusted to keep the private reports confidential; if the auditor is compromised or malicious, the (0,δ)-DP guarantee on the audit transcript is moot because the adversary could learn the report directly from the auditor's internal data.

What would settle it

Run Algorithm 1 with the Toeplitz counter on two adjacent report streams that differ by one report, estimate the total-variation distance between the transcript distributions by Monte Carlo, and check whether it exceeds δ; any violation of d_TV ≤ δ for a single adjacent pair refutes the claimed (0,δ)-DP guarantee. For the utility theorem, simulate a leader with gap Δ and verify whether the mis-selection probability matches the claimed exact Gaussian form Φ(−√2 κδ Δ/MT) for C=2; a systematic mismatch would indicate the fresh-noise decomposition fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Any future improvement in (0,δ)-DP continual counting immediately yields an improved private auditor, because the reduction is generic and privacy is preserved by post-processing.
  • Randomized response is effectively closed off as a basis for long-horizon whistleblower protection: its one-step error stays within δ of uniform auditing no matter how large the report gap.
  • With the Toeplitz counter, the privacy noise grows only logarithmically with the horizon, so long-running audit programs remain practical at fixed deniability levels.
  • The privacy guarantee holds for every prefix of the transcript, so an adversary watching audits unfold has no more power at intermediate times than at the end.
  • If a whistleblower files k reports, the person-level guarantee degrades to (0,kδ)-DP, so keeping k small is necessary for strong person-level deniability.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The log-horizon noise scale suggests the scheme could be adapted to open-ended audit pipelines, such as continuous compliance monitoring, where a fixed T would only set a certified horizon and resets refresh counters.
  • Because the utility theorem is stated in terms of the effective gap, an auditor could plausibly tune δ per reporting round or use the gap estimate to decide when an audit is 'safe' to announce, a practical lever the paper does not develop.
  • One natural stress-test: replace the trusted auditor with a local-randomization layer at the reporting channel (as the paper mentions for stronger threat models) and measure how much utility is lost; the negative result for randomized response hints the loss could be substantial.
  • The exactness of the per-challenger error formula suggests the bound could be inverted to design gap thresholds: given a target mis-selection rate, compute the required gap from Φ(−√2 κδ Δ̃/MT) before deploying.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

0 major / 4 minor

Summary. The paper formalizes whistleblower deniability as per-report (0,δ)-differential privacy on the public audit transcript, with the audited organization itself as the adversary. It proves that randomized-response auditing cannot improve over uniform random auditing by more than δ (Theorem 2), gives a generic reduction from private auditing to private continual counting (Proposition 3), instantiates the reduction with a Toeplitz matrix-factorization counter whose noise scale is O(√log T), and proves a conditional utility bound (Theorem 6) showing that the mis-selection probability decays as the effective leader–challenger gap grows faster than the noise scale. Simulations compare the mechanism against randomized response, uniform auditing, and the non-private greedy baseline under a static gap sweep and a dynamic online auditing setting.

Significance. If the claims hold, the paper supplies a principled formal guarantee for a real and under-formalized problem, and its reduction to continual counting is a clean contribution: any future improvement in (0,δ)-DP continual counting transfers directly to audit-transcript privacy. The negative result for randomized response is crisp and correctly identifies the horizon-independent failure mode. The privacy proof in Appendix D is a careful coupling argument; the utility proof in Appendix F is exact per challenger, with the union bound as the only inequality. The paper is appropriately explicit about its threat model and limitations (Section 9), and the experiments are reproducible in structure with seeds and full parameter tables. The O(√log T) bound is cited rather than proved, and the simulations lack error bars, but neither undermines the central claims.

minor comments (4)
  1. [Section 5.1 / Lemma 5] The headline O(√log T) noise scaling is imported from Dvijotham et al. (Lemma 2.1) without proof. Since this is the basis for the main utility and comparison claims, please either include a short derivation or state the cited result as a formal lemma in the appendix and refer to it explicitly in the main text.
  2. [Section 8] The simulations report point estimates from S=1000 trials but no error bars or confidence intervals. Adding standard errors or confidence bands, at least for Figure 2, would better support the claimed separation between TCA and randomized response.
  3. [Section 4] Typo in the contribution list: 'at any horizons' should be 'at any horizon'.
  4. [Section 7 / Remark F.1] Theorem 6 is stated conditionally on H_{t-1}, and the unconditional version appears only in Appendix Remark F.1. Since the abstract's 'error vanishes' claim is phrased unconditionally, consider moving a concise statement of the unconditional bound into the main text.

Circularity Check

0 steps flagged

No significant circularity; the derivation is a direct reduction to external continual-counting results.

full rationale

The paper's central privacy claim (Proposition 3) is a reduction: it assumes an arbitrary (0,δ)-DP continual counter and proves, by coupling and post-processing, that the audit transcript inherits the same guarantee. The proof identifies the single affected counter and couples all other randomness; no parameter is fitted to the quantity being predicted. The concrete instantiation uses an external Toeplitz factorization (Fichtenberger et al. 2023; Dvijotham et al. 2024) with explicit matrices and sequence (f_k), and the noise scale σ = M_T/(2κ_δ) is derived from the standard Gaussian total-variation lemma (Lemma 17), not from the desired output. The utility theorem (Theorem 6) is a direct Gaussian tail calculation conditional on a fixed history, with an exact per-challenger probability; the effective gap Δ̃ is a conditional quantity, and the bound is a mathematical consequence of the fresh-noise decomposition rather than a quantity defined in terms of the conclusion. The negative result for randomized response follows from a first-decision privacy constraint and is not derived from the statement it proves. All substantive citations are to external, published work; there are no self-citations or imported uniqueness theorems by the same authors. Stated limitations—trusted auditor, group-privacy degradation with k reports, and the absence of real-world data—are genuine scope boundaries and not instances of circular reasoning.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

The central mechanism relies on standard DP and published continual-counting results; no new entities or fitted parameters are introduced. The main modeling assumptions are the trusted auditor and the report-reset semantics.

axioms (5)
  • domain assumption The audit target (adversary) knows the mechanism and the entire report stream except the single report in question.
    Defines the strong-adversary threat model in Section 3; needed for the adjacent-stream distinguishing game.
  • domain assumption The auditor is trusted and will not leak report counts.
    Stated in Section 9; if false, the transcript guarantee is insufficient.
  • domain assumption After an audit, the target's active counts reset (Condition 2).
    Used to define active counts (Definition 3) and in the utility analysis; corresponds to an audit resolving pending concerns.
  • standard math The Toeplitz factorization (f_k) satisfies B C = A and M_T = O(√log T) (Fichtenberger et al. 2023, Dvijotham et al. 2024).
    External proven results used in Section 6.2 and Lemma 5; not proven in this paper.
  • standard math Standard differential privacy facts: TV characterization, group privacy, basic composition.
    Used in Proposition 7, Lemma 8, Corollary 9, Remark 5.1.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Plausible Deniability Guarantees for Whistleblowers." pith.science (2026). https://pith.science/paper/7XNDCDV5

@misc{pith2026260713928,
  author       = {Pith},
  title        = {Pith review of: Plausible Deniability Guarantees for Whistleblowers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7XNDCDV5}},
  note         = {Machine review of arXiv:2607.13928}
}
Share X Bluesky LinkedIn Reddit HN
abstract

Whistleblowers are a key safeguard against organizational wrongdoing, but the threat of retaliation deters reporting. Existing whistleblower-protection proposals lack formal privacy guarantees, and existing differential privacy mechanisms do not directly target the natural threat model -- one in which the audited organization itself observes auditor selection decisions and uses them to identify reporters. We formalize protection against a strong-adversary threat model as per-report $(0, \delta)$-differential privacy on the transcript of audit selections. Within this framework we prove that a natural approach -- randomized response applied at the selection step -- can never outperform uniform random auditing by more than $\delta$ at any horizon. We then give a generic mechanism that reduces private auditing to private continual counting: any $(0, \delta)$-DP continual counter plugs in by post-processing, and the audit transcript inherits the same per-report guarantee. Instantiating the reduction with a recent work in continual counting yields per-report $(0, \delta)$-DP with noise scaling as $O(\sqrt{\log T})$ across a horizon of $T$ audit decisions. A utility theorem shows that the selection error vanishes whenever the noisy report gap between the most-reported organization and the runner-up grows faster than $\sqrt{\log T}$. Simulations show a substantial improvement over randomized response.

Figures

Figures reproduced from arXiv: 2607.13928 by Leo Richter, Matt J. Kusner.

Figure 1
Figure 1. Figure 1: Threat model — two possible worlds. In World 1, the report in question is filed (filled red star), yielding stream RT ; in World 2, it is not (hollow icon), yielding R′ T , with all other reports identical. The auditor processes the stream confidentially via mechanism M and releases the audit transcript aT ∈ [C] T . The adversary — the audit target itself — observes only aT and applies a decision rule φ : … view at source ↗
Figure 2
Figure 2. Figure 2: Utility Simulation. (a) Static gap sweep: TCA’s mis-selection rate decreases as the leader–runner-up gap ∆ grows, while randomized response remains nearly gap-insensitive. (b) Dynamic auditing: TCA incurs substantially lower running active-count deficit than randomized response and uniform auditing, approaching the non-private greedy baseline. Remark 7.2 (Comparison with Randomized Response). Theorem 2 sho… view at source ↗
Figure 3
Figure 3. Figure 3: Mis-selection Rate vs. Leading Gap (C=20) [PITH_FULL_IMAGE:figures/full_fig_p037_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Mis-selection Rate vs. Leading Gap (C=50). Interpretation. Greedy auditing has zero active-count deficit by construction. Uniform random auditing ignores reports and therefore provides a lower utility baseline. Randomized response is expected to behave close to uniform random auditing in long horizons, because the fixed-horizon privacy calibration makes prand close to one. TCA should accumulate substantial… view at source ↗
Figure 5
Figure 5. Figure 5: Mis-selection Rate vs. Leading Gap (C=200). I Operationalization Our framework is defined by a small number of operational choices regarding the audit targets c ∈ [C], the audit horizon T and the target per-report deniability level δ. We describe these choices in a frontier-AI auditing setting, where confidential reports from employees or contractors may inform follow-up audits, model evaluations, or compl… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

67 extracted references · 4 canonical work pages

  1. [1]

    Theory of cryptography conference , pages=

    Calibrating noise to sensitivity in private data analysis , author=. Theory of cryptography conference , pages=. 2006 , organization=

    2006

  2. [2]

    arXiv preprint arXiv:2202.11205 , year=

    Constant matters: Fine-grained complexity of differentially private continual observation , author=. arXiv preprint arXiv:2202.11205 , year=

    Pith/arXiv arXiv

  3. [3]

    2024 IEEE 65th Annual Symposium on Foundations of Computer Science (FOCS) , pages=

    Efficient and near-optimal noise generation for streaming differential privacy , author=. 2024 IEEE 65th Annual Symposium on Foundations of Computer Science (FOCS) , pages=. 2024 , organization=

    2024

  4. [4]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    Unifying Re-Identification, Attribute Inference, and Data Reconstruction Risks in Differential Privacy , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

  5. [5]

    Proceedings of the National Academy of Sciences , volume=

    Towards formalizing the GDPR’s notion of singling out , author=. Proceedings of the National Academy of Sciences , volume=. 2020 , publisher=

    2020

  6. [6]

    Contemporary Accounting Research , volume=

    Whistleblowing allegations, audit fees, and internal control deficiencies , author=. Contemporary Accounting Research , volume=. 2021 , publisher=

    2021

  7. [7]

    Foundations and trends

    The algorithmic foundations of differential privacy , author=. Foundations and trends. 2014 , publisher=

    2014

  8. [8]

    Proceedings on Privacy Enhancing Technologies , year=

    SoK: Managing risks of linkage attacks on data privacy , author=. Proceedings on Privacy Enhancing Technologies , year=

  9. [9]

    Journal of business ethics , volume=

    Prediction of whistleblowing or non-reporting observation: The role of personal and situational factors , author=. Journal of business ethics , volume=. 2013 , publisher=

    2013

  10. [10]

    Review of Law & Economics , volume =

    The Hidden Costs of Whistleblower Protection , author =. Review of Law & Economics , volume =. 2025 , publisher =

    2025

  11. [11]

    Journal of business ethics , volume=

    Whistleblowing in organizations: An examination of correlates of whistleblowing intentions, actions, and retaliation , author=. Journal of business ethics , volume=. 2005 , publisher=

    2005

  12. [12]

    Philosophy & Social Criticism , volume=

    Whistleblowing as civil disobedience: The case of Edward Snowden , author=. Philosophy & Social Criticism , volume=. 2014 , publisher=

    2014

  13. [13]

    International Conference on Machine Learning , pages=

    Constant matters: Fine-grained error bound on differentially private continual observation , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  14. [14]

    The Thirty Seventh Annual Conference on Learning Theory , pages=

    Lower bounds for differential privacy under continual observation and online threshold queries , author=. The Thirty Seventh Annual Conference on Learning Theory , pages=. 2024 , organization=

    2024

  15. [15]

    International Conference on Machine Learning , pages=

    The price of differential privacy under continual observation , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  16. [16]

    International Conference on Artificial Intelligence and Statistics , pages=

    Differentially private histograms under continual observation: Streaming selection into the unknown , author=. International Conference on Artificial Intelligence and Statistics , pages=. 2022 , organization=

    2022

  17. [17]

    arXiv preprint arXiv:2408.07021 , year=

    Count on your elders: Laplace vs Gaussian noise , author=. arXiv preprint arXiv:2408.07021 , year=

    Pith/arXiv arXiv

  18. [18]

    Proceedings of the 2025 Annual ACM-SIAM Symposium on Discrete Algorithms (SODA) , pages=

    Improved differentially private continual observation using group algebra , author=. Proceedings of the 2025 Annual ACM-SIAM Symposium on Discrete Algorithms (SODA) , pages=. 2025 , organization=

    2025

  19. [19]

    Proceedings of the 2023 Annual ACM-SIAM Symposium on Discrete Algorithms (SODA) , pages=

    Almost tight error bounds on differentially private continual counting , author=. Proceedings of the 2023 Annual ACM-SIAM Symposium on Discrete Algorithms (SODA) , pages=. 2023 , organization=

    2023

  20. [20]

    Journal of the American statistical association , volume=

    Randomized response: A survey technique for eliminating evasive answer bias , author=. Journal of the American statistical association , volume=. 1965 , publisher=

    1965

  21. [21]

    Journal of Cybersecurity , volume=

    Pseudonymization and reporters’ protection by design in the EU whistleblower directive , author=. Journal of Cybersecurity , volume=. 2025 , publisher=

    2025

  22. [22]

    Indian Journal of Corporate Governance , volume=

    Comparison of whistle-blower protection mechanism of select countries , author=. Indian Journal of Corporate Governance , volume=. 2018 , publisher=

    2018

  23. [23]

    Data & Policy , volume=

    Evaluating the effectiveness of whistleblower protection: A new index , author=. Data & Policy , volume=

  24. [24]

    2020 , howpublished =

    2020

  25. [25]

    The Journal of Law, Economics, & Organization , volume=

    An economic model of whistle-blower policy , author=. The Journal of Law, Economics, & Organization , volume=. 2009 , publisher=

    2009

  26. [26]

    The Review of Economic Studies , volume=

    Crime, intimidation, and whistleblowing: A theory of inference from unverifiable reports , author=. The Review of Economic Studies , volume=. 2019 , publisher=

    2019

  27. [27]

    Proceedings of the forty-second ACM symposium on Theory of computing , pages=

    Differential privacy under continual observation , author=. Proceedings of the forty-second ACM symposium on Theory of computing , pages=

  28. [28]

    ACM Transactions on Information and System Security (TISSEC) , volume=

    Private and continual release of statistics , author=. ACM Transactions on Information and System Security (TISSEC) , volume=

  29. [29]

    Journal of Political Economy , volume=

    Making corruption harder: Asymmetric information, collusion, and crime , author=. Journal of Political Economy , volume=. 2018 , publisher=

    2018

  30. [30]

    , title =

    Dwork, Cynthia and Naor, Moni and Pitassi, Toniann and Rothblum, Guy N. , title =. Proceedings of the Forty-Second ACM Symposium on Theory of Computing , pages =. 2010 , isbn =. doi:10.1145/1806689.1806787 , abstract =

    arXiv 2010

  31. [31]

    Hubert and Shi, Elaine and Song, Dawn , title =

    Chan, T.-H. Hubert and Shi, Elaine and Song, Dawn , title =. ACM Trans. Inf. Syst. Secur. , month = nov, articleno =. 2011 , issue_date =. doi:10.1145/2043621.2043626 , abstract =

    arXiv 2011

  32. [32]

    Continual Counting with Gradual Privacy Expiration , url =

    Andersson, Joel Daniel and Henzinger, Monika and Pagh, Rasmus and Steiner, Teresa Anna and Upadhyay, Jalaj , booktitle =. Continual Counting with Gradual Privacy Expiration , url =. doi:10.52202/079017-0299 , editor =

    work page doi:10.52202/079017-0299

  33. [33]

    2025 , eprint=

    Streaming Private Continual Counting via Binning , author=. 2025 , eprint=

    2025

  34. [34]

    2024 , eprint=

    Lower Bounds for Differential Privacy Under Continual Observation and Online Threshold Queries , author=. 2024 , eprint=

    2024

  35. [35]

    Proceedings of the 48th Annual IEEE Symposium on Foundations of Computer Science , pages =

    McSherry, Frank and Talwar, Kunal , title =. Proceedings of the 48th Annual IEEE Symposium on Foundations of Computer Science , pages =. 2007 , isbn =. doi:10.1109/FOCS.2007.41 , abstract =

    doi:10.1109/focs.2007.41 2007

  36. [36]

    Permute-and-Flip: A new mechanism for differentially private selection , url =

    McKenna, Ryan and Sheldon, Daniel R , booktitle =. Permute-and-Flip: A new mechanism for differentially private selection , url =

  37. [37]

    Saghaian N

    Zeyu Ding and Daniel Kifer and Sayed M. Saghaian N. E. and Thomas Steinke and Yuxin Wang and Yingtai Xiao and Danfeng Zhang , title =. CoRR , volume =. 2021 , url =. 2105.07260 , timestamp =

    Pith/arXiv arXiv 2021

  38. [38]

    Proceedings of the 51st Annual ACM SIGACT Symposium on Theory of Computing , pages =

    Liu, Jingcheng and Talwar, Kunal , title =. Proceedings of the 51st Annual ACM SIGACT Symposium on Theory of Computing , pages =. 2019 , isbn =. doi:10.1145/3313276.3316377 , abstract =

    arXiv 2019

  39. [39]

    Practical Differentially Private Top-k Selection with Pay-what-you-get Composition , url =

    Durfee, David and Rogers, Ryan M , booktitle =. Practical Differentially Private Top-k Selection with Pay-what-you-get Composition , url =

  40. [40]

    Lower Bounds on Anonymous Whistleblowing

    Quach, Willy and Tyner, LaKyah and Wichs, Daniel. Lower Bounds on Anonymous Whistleblowing. Theory of Cryptography. 2023

    2023

  41. [41]

    Anonymous Whistleblowing over Authenticated Channels

    Agrikola, Thomas and Couteau, Geoffroy and Maier, Sven. Anonymous Whistleblowing over Authenticated Channels. Theory of Cryptography. 2022

    2022

  42. [42]

    Riposte: An Anonymous Messaging System Handling Millions of Users , year =

    Corrigan-Gibbs, Henry and Boneh, Dan and Mazi\`. Riposte: An Anonymous Messaging System Handling Millions of Users , year =. Proceedings of the 2015 IEEE Symposium on Security and Privacy , pages =. doi:10.1109/SP.2015.27 , abstract =

    work page doi:10.1109/sp.2015.27 2015

  43. [43]

    Advances in Neural Information Processing Systems , volume=

    Privacy odometers and filters: Pay-as-you-go composition , author=. Advances in Neural Information Processing Systems , volume=

  44. [44]

    SIAM Journal on Computing , volume=

    What can we learn privately? , author=. SIAM Journal on Computing , volume=. 2011 , publisher=

    2011

  45. [45]

    Digital Journalism , volume =

    Philip Di Salvo , title =. Digital Journalism , volume =. 2021 , publisher =. doi:10.1080/21670811.2021.1889384 , URL =

    Pith/arXiv arXiv 2021

  46. [46]

    Privacy Auditing with One (1) Training Run , url =

    Steinke, Thomas and Nasr, Milad and Jagielski, Matthew , booktitle =. Privacy Auditing with One (1) Training Run , url =

  47. [47]

    Su , title =

    Jinshuo Dong and Aaron Roth and Weijie J. Su , title =. CoRR , volume =. 2019 , url =. 1905.02383 , timestamp =

    Pith/arXiv arXiv 2019

  48. [48]

    CoRR , volume =

    Ilya Mironov , title =. CoRR , volume =. 2017 , url =. 1702.07476 , timestamp =

    Pith/arXiv arXiv 2017

  49. [49]

    and Mitchell, Margaret and Gebru, Timnit and Hutchinson, Ben and Smith-Loud, Jamila and Theron, Daniel and Barnes, Parker , title =

    Raji, Inioluwa Deborah and Smart, Andrew and White, Rebecca N. and Mitchell, Margaret and Gebru, Timnit and Hutchinson, Ben and Smith-Loud, Jamila and Theron, Daniel and Barnes, Parker , title =. Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency , pages =. 2020 , isbn =. doi:10.1145/3351095.3372873 , abstract =

    arXiv 2020

  50. [50]

    Who Audits the Auditors? Recommendations from a field scan of the algorithmic auditing ecosystem , url=

    Costanza-Chock, Sasha and Raji, Inioluwa Deborah and Buolamwini, Joy , year=. Who Audits the Auditors? Recommendations from a field scan of the algorithmic auditing ecosystem , url=. doi:10.1145/3531146.3533213 , booktitle=

    arXiv

  51. [51]

    2023 , eprint=

    Frontier AI Regulation: Managing Emerging Risks to Public Safety , author=. 2023 , eprint=

    2023

  52. [52]

    Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society , author=

    Responsible Reporting for Frontier AI Development , volume=. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society , author=. 2024 , month=. doi:10.1609/aies.v7i1.31678 , abstractNote=

    doi:10.1609/aies.v7i1.31678 2024

  53. [53]

    2025 , eprint=

    Audit Cards: Contextualizing AI Evaluations , author=. 2025 , eprint=

    2025

  54. [54]

    2022 , eprint=

    When Privacy Meets Partial Information: A Refined Analysis of Differentially Private Bandits , author=. 2022 , eprint=

    2022

  55. [55]

    Proceedings of the AAAI Conference on Artificial Intelligence , author=

    Algorithms for Differentially Private Multi-Armed Bandits , volume=. Proceedings of the AAAI Conference on Artificial Intelligence , author=. 2016 , month=. doi:10.1609/aaai.v30i1.10212 , abstractNote=

    work page doi:10.1609/aaai.v30i1.10212 2016

  56. [56]

    2016 , publisher=

    Committing to effective whistleblower protection , author=. 2016 , publisher=

    2016

  57. [57]

    She paid a price

    She blew the whistle to protect seniors at the Rosslyn. She paid a price. , author=

  58. [58]

    The Journal of Finance , volume =

    Dyck, Alexander and Morse, Adair and Zingales, Luigi , title =. The Journal of Finance , volume =. doi:https://doi.org/10.1111/j.1540-6261.2010.01614.x , url =. https://onlinelibrary.wiley.com/doi/pdf/10.1111/j.1540-6261.2010.01614.x , abstract =

    arXiv 2010

  59. [59]

    and Welch, Kyle T

    Stubben, Stephen R. and Welch, Kyle T. , title =. Journal of Accounting Research , volume =. doi:https://doi.org/10.1111/1475-679X.12303 , url =. https://onlinelibrary.wiley.com/doi/pdf/10.1111/1475-679X.12303 , abstract =

    work page doi:10.1111/1475-679x.12303

  60. [60]

    Bowen and Andrew C

    Robert M. Bowen and Andrew C. Call and Shiva Rajgopal , journal =. Whistle-Blowing: Target Firm Characteristics and Economic Consequences , urldate =

  61. [61]

    Global Survey of Business Ethics 2022-2024: teaching, research and training , pages=

    Global survey on business ethics 2022-2024-the Netherlands , author=. Global Survey of Business Ethics 2022-2024: teaching, research and training , pages=. 2024 , publisher=

    2022

  62. [62]

    2021 , school=

    Who gets stitches? The effects of rewarding whistleblowers and protecting their identity on subsequent cooperation , author=. 2021 , school=

    2021

  63. [63]

    Public Personnel Management , volume=

    Whistleblower retaliation in the public sector , author=. Public Personnel Management , volume=. 2011 , publisher=

    2011

  64. [64]

    , author=

    Retaliation against whistle blowers: Predictors and effects. , author=. Journal of applied psychology , volume=. 1986 , publisher=

    1986

  65. [65]

    2019 , month = oct, number =

    2019

  66. [66]

    2024 , month =

    Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence and amending Regulations (EC) No 300/2008, (EU) No 167/2013, (EU) No 168/2013, (EU) 2018/858, (EU) 2018/1139 and (EU) 2019/2144 and Directives 2014/90/EU, (EU) 2016/797 and (EU) 2020/1828 (Artificial Intelligence ...

    2024

  67. [67]

    2024 , howpublished =

    Artificial Intelligence Act , subtitle=. 2024 , howpublished =

    2024

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.