Differential Privacy (DP) has emerged as the gold standard for protecting individual data in machine learning. However, its mathematical frameworks—ranging from $(\epsilon, \delta)$-DP to Rényi DP and $f$-DP—often feel disconnected and difficult to communicate to non-experts. This paper introduces the sensitivity index ($\psi$), a single, parsimonious parameter that unifies these interpretations and provides a more intuitive, geometric understanding of the Gaussian Mechanism (GM).
The Sensitivity Index ($\psi$)
The Gaussian Mechanism is the primary tool for adding noise to high-dimensional queries, such as gradients in deep learning (DP-SGD). The authors define the sensitivity index as:
$$\psi = \frac{\Delta}{\sigma}$$
where $\Delta$ represents the $L_2$ sensitivity of the query and $\sigma$ is the standard deviation of the added noise. This single value acts as a "singular parameter" that fully characterizes the privacy guarantees of the mechanism, simplifying the need to juggle multiple variables like $\epsilon$, $\delta$, and $\alpha$.
Privacy as Binary Classification
A core contribution of this work is the geometric interpretation of privacy. If an adversary attempts to determine whether a specific individual's data was included in a dataset, they are essentially acting as a binary classifier. The privacy of the Gaussian Mechanism can be viewed as the adversary's inability to distinguish between two neighboring datasets.
In this framework, the privacy profile is perfectly described by a binormal Receiver Operating Characteristic (ROC) curve. The authors prove that the ROC curve for a Gaussian Mechanism with index $\psi$ follows the formula $R(x) = \Phi(\psi + \Phi^{-1}(x))$, where $\Phi$ is the Gaussian cumulative distribution function.
Key Results and Unification
By establishing $\psi$ as the central parameter, the authors provide exact mathematical mappings between various DP definitions:
- Area Under the Curve (AUC): The privacy level can be expressed as a single percentage-like value: $AUC = \Phi(\psi / \sqrt{2})$.
- Rényi DP Equivalence: A Gaussian Mechanism with index $\psi$ satisfies $(\alpha, \frac{\alpha}{2}\psi^2)$-Rényi DP.
- Superior Conversion: The paper demonstrates that using the "privacy profile" based on $\psi$ allows for tighter, more accurate conversions to $(\epsilon, \delta)$-DP compared to standard, often "lossy," Rényi DP methods.
Conclusion
The sensitivity index $\psi$ serves as a "Rosetta Stone" for the Gaussian Mechanism, bridging the gap between complex theoretical frameworks and practical engineering. By linking privacy to the familiar language of ROC curves and classification metrics, this work democratizes the understanding of differential privacy and provides tools for more precise privacy accounting in machine learning pipelines. While currently limited to the Gaussian Mechanism, this approach paves the way for similar parsimonious descriptions of other privacy-enhancing technologies.
Summary updated