TL;DR
This paper argues that the difficulty in setting the privacy budget ($\varepsilon$) in differential privacy (DP) is not an inherent flaw of DP, but rather a reflection of the fundamental challenge of quantifying privacy risks in real-world contexts. The authors contend that focusing on minimizing $\varepsilon$ can lead to a false sense of security and hinder the adoption of DP, advocating instead for a more nuanced understanding of privacy guarantees.
What Problem Does This Paper Address?
Differential Privacy (DP) is a powerful framework for preserving privacy in machine learning, offering strong theoretical guarantees. However, a significant hurdle to its widespread adoption is the perceived difficulty in setting the "privacy budget" ($\varepsilon$). This budget quantifies the maximum privacy loss allowed. Critics often point to the challenge of interpreting $\varepsilon$ and determining "safe" values as a major limitation, sometimes using this as a reason to favor less rigorous privacy-preserving methods. This paper addresses the question of whether the difficulty in setting $\varepsilon$ is a fundamental problem with DP itself, or if it stems from a broader challenge in quantifying privacy, and how this perception impacts the field.
Methods
This paper is a position paper, meaning it primarily presents arguments and perspectives rather than introducing new algorithms or empirical results. The authors' "method" involves:
- Argumentation and Conceptual Analysis: They analyze the nature of privacy risk and the role of the privacy budget within the DP framework.
- Critique of Existing Perceptions: They challenge the common narrative that setting $\varepsilon$ is the primary obstacle to DP adoption.
- Re-framing the Problem: They argue that the difficulty lies in estimating privacy risks in context, a challenge faced by any privacy assessment method, not just DP.
- Highlighting DP's Strengths: They emphasize the desirable properties of DP, such as robustness to post-processing and composition, and the interpretability of $\varepsilon$ in simple settings.
- Identifying Harmful Practices: They criticize the trend of "privacy-washing" and the overemphasis on achieving small $\varepsilon$ values at the expense of genuine privacy.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Privacy Risk is Inherently Difficult to Quantify: The paper argues that the challenge of setting $\varepsilon$ is not unique to DP but is a fundamental problem in assessing privacy risks, especially in high-dimensional data. Historical failures of anonymization methods illustrate this difficulty.
- DP's $\varepsilon$ is a Feature, Not a Bug: The privacy budget, despite its perceived complexity, offers a standardized and interpretable way to communicate privacy guarantees, especially when compared to ad-hoc methods.
- Focusing on Small $\varepsilon$ is Misleading: The pursuit of minimal $\varepsilon$ values can lead to "privacy-washing" where algorithms appear more private than they are, or to the use of less robust privacy techniques. This can actively harm users by creating a false sense of security.
- Context is Crucial for Privacy: Privacy expectations are highly context-dependent. DP's worst-case guarantees, while sometimes pessimistic, provide a robust baseline.
- DP's Properties are Essential: Robustness to post-processing and composition are critical for long-term trust in privacy-preserving systems, and DP provides these guarantees, unlike many alternative methods.
- "No One Knows How to Set $\varepsilon$" is a Harmful Narrative: This perception can be used to justify avoiding DP altogether or adopting it improperly, hindering progress.
Key Results
This paper is not focused on quantitative results but rather on a conceptual argument. The "results" are the insights and arguments presented:
- The difficulty in setting $\varepsilon$ is a symptom of the broader challenge of quantifying privacy, not a flaw of DP itself.
- DP's $\varepsilon$ provides a valuable, albeit sometimes challenging, metric for privacy.
- The pursuit of small $\varepsilon$ values can be detrimental to actual privacy.
- Alternative privacy methods often lack the theoretical guarantees of DP and can provide a false sense of security.
Limitations and Future Work
- Limitations: The paper acknowledges that while DP offers theoretical guarantees, interpreting the privacy budget in complex, real-world scenarios remains challenging. It also notes that the "ideal" $\varepsilon$ value for a given task is still an open research question.
- Future Work: The authors suggest that the focus should shift from solely minimizing $\varepsilon$ to understanding the context of privacy risks and developing better methods for privacy accounting. They also advocate for clearer communication about DP's capabilities and limitations to practitioners and the public, encouraging broader adoption rather than focusing on perceived obstacles. They also suggest exploring alternative metrics that can be converted to standard $(\varepsilon, \delta)$-DP for comparison.
Summary updated