Teresa Anna Steiner

Sofya Raskhodnikova, Teresa Anna Steiner

We study differentially private algorithms for analyzing graphs in the challenging setting of continual release with fully dynamic updates, where edges are inserted and deleted over time, and the algorithm is required to update the solution at every time step. Previous work has presented differentially private algorithms for many graph problems that can handle insertions only or deletions only (called partially dynamic algorithms) and obtained some hardness results for the fully dynamic setting. The only algorithms in the latter setting were for the edge count, given by Fichtenberger, Henzinger, and Ost (ESA 21), and for releasing the values of all graph cuts, given by Fichtenberger, Henzinger, and Upadhyay (ICML 23). We provide the first differentially private and fully dynamic graph algorithms for several other fundamental graph statistics (including the triangle count, the number of connected components, the size of the maximum matching, and the degree histogram), analyze their error and show strong lower bounds on the error for all algorithms in this setting. We study two variants of edge differential privacy for fully dynamic graph algorithms: event-level and item-level. We give upper and lower bounds on the error of both event-level and item-level fully dynamic algorithms for several fundamental graph problems. No fully dynamic algorithms that are private at the item-level (the more stringent of the two notions) were known before. In the case of item-level privacy, for several problems, our algorithms match our lower bounds.

Giulia Bernardini, Philip Bille, Inge Li Gørtz et al.

For databases consisting of many text documents, one of the most fundamental data analysis tasks is counting (i) how often a pattern appears as a substring in the database (substring counting) and (ii) how many documents in the collection contain the pattern as a substring (document counting). If such a database contains sensitive data, it is crucial to protect the privacy of individuals in the database. Differential privacy is the gold standard for privacy in data analysis. It gives rigorous privacy guarantees, but comes at the cost of yielding less accurate results. In this paper, we carry out a theoretical study of substring and document counting under differential privacy. We propose a data structure storing $ε$-differentially private counts for all possible query patterns with a maximum additive error of $O(\ell\cdot\mathrm{polylog}(n\ell|Σ|))$, where $\ell$ is the maximum length of a document in the database, $n$ is the number of documents, and $|Σ|$ is the size of the alphabet. We also improve the error bound for document counting with $(ε, δ)$-differential privacy to $O(\sqrt{\ell}\cdot\mathrm{polylog}(n\ell|Σ|))$. We show that our additive errors for substring counting and document counting are optimal up to an $O(\mathrm{polylog}(n\ell))$ factor both for $ε$-differential privacy and $(ε, δ)$-differential privacy. Our data structures immediately lead to improved algorithms for related problems, such as privately mining frequent substrings and q-grams. Additionally, we develop a new technique of independent interest for differentially privately computing a general class of counting functions on trees.