TL;DR
This paper introduces a formal framework for protecting whistleblowers from retaliation by ensuring "plausible deniability" through Differential Privacy (DP). The authors prove that standard randomized selection methods fail over long time horizons and propose a new mechanism—Toeplitz Continual Auditing (TCA)—that maintains high auditing utility while providing strong, provable privacy guarantees for reporters.
What Problem Does This Paper Address?
Whistleblowers are essential for organizational accountability, yet they face a massive deterrent: the fear of retaliation. Even if a report is submitted anonymously to an independent auditor, the timing of an audit can be a "dead giveaway." If an organization is audited immediately after a specific internal event known only to a few people, the organization can easily infer that a report was filed and narrow down the list of suspects.
Current whistleblower protections often lack formal technical guarantees. While "Randomized Response" (flipping a biased coin to decide whether to follow a signal or act randomly) is a classic tool for privacy, the authors show it has a fatal flaw in this context: to maintain a fixed level of privacy over a long period, the auditor must eventually become so random that they ignore reports entirely, rendering the audit process useless.
The challenge is to design an auditing policy that is responsive (audits the most-reported organizations) but private (the organization cannot tell if a specific audit was triggered by a whistleblower or just by random noise).
Methods
The authors formalize the problem using $(0, \delta)$-Differential Privacy on the public transcript of audit decisions. They assume a "strong adversary" model where the organization being audited knows its own internal state and is actively trying to identify whistleblowers from the audit history.
The Core Innovation: From Selection to Counting
Instead of adding randomness to the selection step (like Randomized Response), the authors propose privatizing the report counts themselves.
- Continual Counting: Each organization has a running counter of reports. The mechanism adds carefully calibrated Gaussian noise to these counts at every timestep.
- Toeplitz Factorization: To keep the noise as low as possible over time, they use a state-of-the-art "matrix mechanism" based on Toeplitz factorization. This ensures that the noise only grows at a rate of $O(\sqrt{\log T})$, where $T$ is the time horizon.
- The Selection Rule: The auditor simply selects the organization with the highest noisy count.
- Report Reset: Once an organization is audited, its counter is reset. The authors provide a novel proof showing that this "reset" logic still preserves the overall privacy guarantee through a complex "coupling" argument.
Datasets
Because real-world whistleblower data is highly sensitive and rarely public, the authors evaluated their approach using synthetic simulations:
- Static Gap Sweep: A controlled environment where one "guilty" organization has a fixed number of reports ($\Delta$) and others have zero. This measures how many reports are needed to "break through" the noise.
- Dynamic Online Auditing: A streaming report process where reports are generated via independent Poisson models (with a higher arrival rate for a "lead" organization). This simulates a realistic, long-term auditing environment.
Key Insights and Ideas
- The "Impossibility" of Randomized Response: The paper proves that for any fixed privacy level, Randomized Response eventually "degenerates" into uniform random guessing as the number of audit decisions grows. This effectively kills the utility of long-term auditing programs.
- Post-Processing is Free: By privatizing the counts (the input to the decision) rather than the decision itself, the audit selection becomes a "post-processing" step of a DP output. This allows the auditor to use the best available DP counting algorithms "off-the-shelf."
- Vanishing Error: The authors provide a utility theorem showing that as the gap between a "guilty" organization and its peers grows, the probability of the auditor making a mistake (mis-selection) vanishes, even while maintaining strict privacy.
Key Results
- Noise Scaling: The proposed TCA mechanism achieves noise scaling of $O(\sqrt{\log T})$, a massive improvement over the linear scaling required by simpler methods.
- Utility Gains: In simulations, TCA significantly outperformed Randomized Response. In the "static gap" test, TCA’s error rate dropped sharply as the report gap increased, while Randomized Response remained stuck at high error levels regardless of how many reports were filed.
- Efficiency: In dynamic settings, TCA stayed close to the "greedy" (non-private) baseline, meaning it successfully identified the most-reported organizations without sacrificing whistleblower safety.
Limitations and Future Work
- Linear Privacy Loss: Currently, if a single whistleblower files multiple reports, their privacy protection degrades linearly. The authors suggest that moving to a different DP framework (like $f$-DP) might provide tighter bounds for "frequent" whistleblowers.
- The Trusted Auditor: The model assumes the auditor is a "trusted third party" who won't leak the raw, non-noisy counts. Strengthening the model to handle an untrusted auditor (Local DP) would be a significant next step.
- Real-World Complexity: The paper doesn't account for "side-channel" attacks, such as an organization monitoring who visits the auditor's office or leaks from non-anonymous reporting channels (e.g., email metadata).
- Scope of Audit: The guarantee only covers the decision to audit, not the content or outcome of the audit itself, which could also leak information.
Summary updated