Optimality of Matrix Mechanism on $\ell_p^p$-metric

Jingcheng Liu,Jalaj Upadhyay,Zongrui Zou

2024-06-04

Abstract:In this paper, we introduce the $\ell_p^p$-error metric (for $p \geq 2$) when answering linear queries under the constraint of differential privacy. We characterize such an error under $(\epsilon,\delta)$-differential privacy. Before this paper, tight characterization in the hardness of privately answering linear queries was known under $\ell_2^2$-error metric (Edmonds et al., STOC 2020) and $\ell_p^2$-error metric for unbiased mechanisms (Nikolov and Tang, ITCS 2024). As a direct consequence of our results, we give tight bounds on answering prefix sum and parity queries under differential privacy for all constant $p$ in terms of the $\ell_p^p$ error, generalizing the bounds in Henzinger et al. (SODA 2023) for $p=2$.

Cryptography and Security,Machine Learning

What problem does this paper attempt to address?

The core problem that this paper attempts to solve is: under the constraint of maintaining differential privacy, when answering linear queries, how to characterize and optimize the performance based on the $\ell_p$ error metric (where $p \geq 2$). Specifically:

1. **Introducing a new error metric**: The paper introduces the $\ell_p^p$-error metric to evaluate the accuracy of answering linear queries under the differential privacy constraint. Previous research mainly focused on the $\ell_2^2$-error metric (Edmonds et al., STOC 2020) and the $\ell_{2p}$-error metric under the unbiased mechanism (Nikolov and Tang, ITCS 2024), and this paper extends these results to the broader $\ell_p^p$-error metric.

2. **Characterizing the error lower bound**: The authors provide a tight lower bound for the $\ell_p^p$-error under the $(ε, δ)$ - differential privacy mechanism. This enables them to provide strict error bounds for specific types of queries such as prefix sum queries and parity queries, thus generalizing the work of Henzinger et al. (SODA 2023).

3. **Optimality of the matrix mechanism**: The paper proves that the matrix mechanism is optimal under the $\ell_p^p$-error metric. This conclusion applies not only to the unbiased mechanism but also to the general differential privacy mechanism. The matrix mechanism protects privacy by decomposing the query matrix and adding appropriate Gaussian noise while ensuring the accuracy of the query results.

### Formula representation

The $\ell_p^p$-error metric defined in the paper is as follows:
\[ \text{err}_{\ell_p^p}(M, A) := \max_{x \in \mathbb{R}^n} \left( \mathbb{E} \left[ \| M(x) - Ax \|_p^p \right] \right)^{1/p} \]

In addition, for any query matrix $A \in \mathbb{R}^{m \times n}$, if a mechanism $M$ satisfies $(ε, δ)$ - differential privacy, its $\ell_p^p$-error satisfies the following inequality:
\[ \text{err}_{\ell_p^p}(M, A) \geq \frac{(1 - \tilde{\delta}) \gamma(p)(A)}{C' \varepsilon} \]
where $\tilde{\delta} = \frac{2 e^{2\varepsilon} (e^{1/2} - 1)}{e^\varepsilon - 1} \delta$, and $\gamma(p)(A)$ is a factorized norm related to the matrix $A$.

### Conclusion

By introducing the $\ell_p^p$-error metric and providing tight upper and lower bounds for it, this paper not only extends the results of existing research but also provides a more general and strict theoretical basis for linear queries under differential privacy. This helps to better understand how to design optimal privacy - protecting mechanisms under different error metrics.