🛰️ Daily AI Frontier
‹ back to 2026-07-16

Plausible Deniability Guarantees for Whistleblowers

Research Other

Merged summary

TL;DR — This is a security/privacy theory paper that gives whistleblowers formal plausible-deniability guarantees by framing audit-selection privacy as differential privacy; it's not about AI/ML advancement, so no listed topic fits.

  • Formalizes protection under a strong-adversary threat model (the audited org observes selection decisions) as per-report $(0, δ)$-differential privacy over the transcript of audit selections.
  • Proves a negative result: randomized response at the selection step can never beat uniform random auditing by more than $δ$ at any horizon.
  • Provides a generic reduction from private auditing to private continual counting; any $(0, δ)$-DP counter plugs in via post-processing, yielding noise scaling of $O(\sqrt{\log T})$ over $T$ decisions.
  • Utility theorem: selection error vanishes when the noisy report gap between top and runner-up organizations grows faster than $\sqrt{\log T}$; simulations beat randomized response.

Sources (1)

Plausible Deniability Guarantees for Whistleblowers

arXiv cs.CR Leo Richter, Matt J. Kusner 2026-07-15 arXiv:2607.13928

TL;DR — This is a security/privacy theory paper that gives whistleblowers formal plausible-deniability guarantees by framing audit-selection privacy as differential privacy; it's not about AI/ML advancement, so no listed topic fits.

  • Formalizes protection under a strong-adversary threat model (the audited org observes selection decisions) as per-report $(0, δ)$-differential privacy over the transcript of audit selections.
  • Proves a negative result: randomized response at the selection step can never beat uniform random auditing by more than $δ$ at any horizon.
  • Provides a generic reduction from private auditing to private continual counting; any $(0, δ)$-DP counter plugs in via post-processing, yielding noise scaling of $O(\sqrt{\log T})$ over $T$ decisions.
  • Utility theorem: selection error vanishes when the noisy report gap between top and runner-up organizations grows faster than $\sqrt{\log T}$; simulations beat randomized response.
item →