arXiv subjects

Sophie Taylor

Publications and source records attributed to Sophie Taylor.

3 recordsLinked to original sources

arXiv2026

Realisation-Level Privacy Filtering

We study differentially private data release, where a database is accessed through successive, possibly adaptive queries and mechanisms. Existing composition theorems and privacy filters almost always combine worst case per-round privacy parameters, leaving room for more refined accounting based on realised leakage, which we term realisation-level accounting. We present a realisation-level filtering approach to determine stopping times for data releases, and design one such filter. Despite technical challenges arising from conditioning on realisations and stopping time, we prove that the filter guarantees $(ε, δ)$-differential privacy, with $ε$ and $δ$ chosen by the data handler. Through numerical evidence, we demonstrate that realisation-level filtering provides a path to better utility beyond mechanism-level methods. Furthermore, our proposed filter applies to arbitrary mechanisms, including those that are badly behaved under Rényi differential privacy.

cs.CR↗

arXiv2026

Adaptive Privacy of Sequential Data Releases Under Collusion

The fundamental trade-off between privacy and utility remains an active area of research. Our contribution is motivated by two observations. First, privacy mechanisms developed for one-time data release cannot straightforwardly be extended to sequential releases. Second, practical databases are likely to be useful to multiple distinct parties. Furthermore, we can not rule out the possibility of data sharing between parties. With utility in mind, we formulate a new privacy-utility trade-off problem to adaptively tackle sequential data requests made by different, potentially colluding entities. We consider both expected distortion and mutual information as measures to quantify utility, and use mutual information to measure privacy. We assume an attack model whereby illicit data sharing, which we call collusion, can occur between data receivers. We develop an adaptive algorithm for data releases that makes use of a Blahut-Arimoto-style algorithm. We show that the resulting data releases are optimal when expected distortion quantifies utility, and locally optimal when mutual information quantifies utility. Numerical experiments on real data demonstrate that the proposed adaptive algorithm can exploit previously released information to reduce cumulative leakage under collusion without sacrificing much, if any utility. Finally, we discuss how our findings may extend to applications in machine learning.

cs.IT↗

arXiv2025

The Asymptotic Behaviour of Information Leakage Metrics

Information theoretic leakage metrics quantify the amount of information about a private random variable $X$ that is leaked through a correlated revealed variable $Y$. They can be used to evaluate the privacy of a system in which an adversary, from whom we want to keep $X$ private, is given access to $Y$. Global information theoretic leakage metrics quantify the overall amount of information leaked upon observing $Y$, whilst their pointwise counterparts define leakage as a function of the particular realisation $Y=y$ that the adversary sees, and thus can be viewed as random variables. We consider an adversary who observes a large number of independent identically distributed realisations of $Y$. We formalise the essential asymptotic behaviour of an information theoretic leakage metric, considering in turn what this means for pointwise and global metrics. With the resulting requirements in mind, we take an axiomatic approach to defining a set of pointwise leakage metrics, as well as a set of global leakage metrics that are constructed from them. The global set encompasses many known measures including mutual information, Sibson mutual information, Arimoto mutual information, maximal leakage, min entropy leakage, $f$-divergence metrics, and g-leakage. We prove that both sets follow the desired asymptotic behaviour. Finally, we derive composition theorems which quantify the rate of privacy degradation as an adversary is given access to a large number of independent observations of $Y$. It is found that, for both pointwise and global metrics, privacy degrades exponentially with increasing observations for the adversary, at a rate governed by the minimum Chernoff information between distinct conditional channel distributions. This extends the work of Wu et al. (2024), who have previously found this to be true for certain known metrics, including some that fall into our more general set.

cs.IT↗