Abstract
We introduce a family of information leakage measures called maximal $(\alpha,\beta)$-leakage (M$\alpha$beL), parameterized by real numbers $\alpha$ and $\beta$ greater than or equal to 1. The measure is formalized via an operational definition involving an adversary guessing an unknown (randomized) function of the data given the released data. We obtain a simplified computable expression for the measure and show that it satisfies several basic properties such as monotonicity in $\beta$ for a fixed $\alpha$, non-negativity, data processing inequalities, and additivity over independent releases. We highlight the relevance of this family by showing that it bridges several known leakage measures, including maximal $\alpha$-leakage $(\beta=1)$, maximal leakage $(\alpha=\infty,\beta=1)$, local differential privacy (LDP) $(\alpha=\infty,\beta=\infty)$, and local Renyi differential privacy (LRDP) $(\alpha=\beta)$, thereby giving an operational interpretation to local Renyi differential privacy. We also study a conditional version of M$\alpha$beL on leveraging which we recover differential privacy and Renyi differential privacy. A new variant of LRDP, which we call maximal Renyi leakage, appears as a special case of M$\alpha$beL for $\alpha=\infty$ that smoothly tunes between maximal leakage ($\beta=1$) and LDP ($\beta=\infty$). Finally, we show that a vector form of the maximal Renyi leakage relaxes differential privacy under Gaussian and Laplacian mechanisms.
Introduction: Navigating the Privacy Zoo
The privacy research community has historically been divided by its choice of metrics. Computer scientists often favor Differential Privacy (DP) for its robust worst-case guarantees, while information theorists lean toward measures like Mutual Information or Maximal Leakage (MaxL), which offer clear "operational" meanings—such as quantifying how much an adversary's ability to guess sensitive data improves after an observation.
This fragmentation creates a "zoo" of disconnected measures. System designers often struggle to translate a guarantee in one framework (like Renyi DP) into a practical understanding of adversarial gain. This paper addresses this gap by introducing Maximal $(\alpha, \beta)$-Leakage ($M\alpha beL$), a unified family of measures that bridges the gap between information theory and computer science. By tuning just two parameters, $M\alpha beL$ provides a single, operationally meaningful framework for quantifying sensitive information leakage.
The MαbeL Framework: An Operational Approach
The core of $M\alpha beL$ is an operational "guessing" game. Consider an adversary who wishes to guess some randomized function of sensitive data $X$ after observing a released version $Y$. The measure is parameterized by two real numbers:
- $\alpha$: Relates to the power function applied to the probability of a correct guess (the "loss function").
- $\beta$: Tunes the measure between an average-case (averaging over all possible released values $Y$) and a worst-case (focusing on the single most revealing value of $Y$).
As shown in the figure above, $M\alpha beL$ acts as a coordinate system for privacy. In the local setting (a), setting $\beta=1$ yields Maximal $\alpha$-Leakage, while letting $\alpha, \beta \to \infty$ recovers Local Differential Privacy (LDP). Crucially, this framework provides an operational interpretation for Renyi Differential Privacy (RDP), showing it is a specific point in the $M\alpha beL$ family. The authors also introduce a "vector form" (b) to move from local user protection to global dataset protection, successfully recovering standard DP and RDP.
Theoretical Properties and Composition
The authors derive a simplified, computable expression for $M\alpha beL$ and prove it satisfies the essential "axiomatic" properties required of any robust leakage measure:
- Non-negativity: Leakage is never negative.
- Data Processing Inequality: Post-processing released data cannot increase the information leaked about the original source.
- Additivity: The measure is additive over independent data releases, supporting adaptive composition.
To ensure mathematical continuity across all parameter ranges, the authors also introduce a $(\alpha, \tau)$ reparameterization, which fixes potential "edges" where the original $(\alpha, \beta)$ form might behave unpredictably.
Mechanism Analysis: Laplacian and Gaussian
While theoretical, the paper evaluates $M\alpha beL$ using standard mathematical privacy mechanisms rather than empirical datasets:
- Laplacian Mechanism: Typically used for $\epsilon$-differential privacy.
- Gaussian Mechanism: Standard for $(\epsilon, \delta)$-differential privacy or Renyi DP.
The analysis demonstrates that $M\alpha beL$ provides a "smoother" and more flexible privacy guarantee than traditional DP. For these mechanisms, $M\alpha beL$ can offer a more nuanced view of privacy that avoids the overly conservative nature of worst-case DP bounds while maintaining rigorous protection.
Discussion and Future Directions
While $M\alpha beL$ provides a powerful unified map, some challenges remain. Calculating the supremum over all possible data distributions ($P_X$) can be computationally intensive for high-dimensional data. Furthermore, while the paper maps the landscape, determining the "optimal" $\alpha$ and $\beta$ for specific real-world applications—such as medical databases or census data—remains an open question for practitioners.
Future work may explore extending this framework beyond "guessing" adversaries to other tasks, such as membership inference or reconstruction attacks in deep learning, further solidifying $M\alpha beL$ as a foundational tool for privacy-preserving system design.
Summary updated