Dependency Triad: A Metric to Quantify the Dependencies Between Attributes for Local Differential Privacy

CoP: Coordinated Perturbation for Controlled Disclosure Under Local Differential Privacy

Collecting multidimensional user data is essential for personalized services, yet it poses significant privacy risks. While privacy regulations like the GDPR and CPRA advocate for data minimization, attribute correlations can inadvertently amplify unintentional information disclosure, leading to correlation-induced information leakage (CIL). Although data collectors often possess rich prior knowledge of these correlations, existing Local Differential Privacy (LDP) mechanisms are inadequate for effectively leveraging this information to reduce CIL. In this paper, we propose CoP, a coordinated perturbation mechanism designed to mitigate CIL in multidimensional data collection while preserving utility. Unlike traditional LDP approaches, CoP explicitly incorporates prior distribution knowledge to coordinate the perturbation process across attributes. By optimizing the perturbation strategy based on known correlations, CoP achieves a better privacy-utility trade-off. Extensive evaluations across both synthetic and real-world datasets demonstrate that CoP significantly outperforms state-of-the-art LDP mechanisms in reducing disclosure while preserving analytical accuracy.

From Privacy-Utility Trade-Offs to Policies: Optimized Anonymization Recommendations for Data Trustees in Data Spaces

: As data spaces emerge to facilitate sovereign data exchange, geospatial data has become a critical resource across various domains. However, sharing sensitive geodata, such as cadastral property records, presents a privacy-utility trade-off. While the European Data Governance Act (DGA) establishes data trustees as the technical intermediaries authorized to perform necessary anonymization, a gap remains between theoretical spatial privacy algorithms and their practical, policy-driven application. This research addresses this gap by presenting an empirical evaluation of 11 spatial anonymization methods applied to complex polygon geometries. Using a dataset of 2,147 forested cadastral parcels, we quantify the trade-offs using privacy metrics ( k -anonymity, differential privacy ε ) and utility measures (Hausdorff distance, area deviation, centroid shift). Our results identify algorithmic failure modes, such as the “sparse forest” phenomenon, and demonstrate the improved performance of combined hybrid approaches. To operationalize these findings, we integrate our results into the architecture of data trustees operating within various domain-specific data spaces. We demonstrate how the empirically derived, use-case-specific anonymization guidelines can be translated into machine-interpretable Open Digital Rights Language (ODRL) policies. By mapping spatial transformation algorithms to ODRL constraints, this research provides data trustees with an approach to automatically enforce spatial privacy and sovereignty within data spaces.

Privacy-Preserving Data Drift Detection and Recovery for Large-Scale LLM Applications via Proxy Representations

LLM applications deployed at scale face a fundamental challenge: privacy constraints prevent direct inspection of user interactions, making it difficult to obtain any representative evaluation dataset or to track the ongoing evolution of production traffic. We present ProxyDrift, a framework that (i) identifies and measures drift between production traffic and offline evaluation sets, and (ii) constructs and refreshes those evaluation sets accordingly; all without access to raw user data. Our approach operates entirely on non-PII proxy representations: structured, multi-dimensional descriptors derived from LLM-based classification of user interactions. We introduce (1) a chance-calibrated, redundancy-aware (RA) alignment score that aggregates per-dimension drift measurements via mutual information; (2) a conditional sampler that generates synthetic proxies respecting inter-dimensional dependencies; (3) a roundtrip consistency analysis that exposes generator/classifier disagreements and guides proxy taxonomy refinement; and (4) a feedback-linkage analysis that ties per-dimension and per-value proxy distributions to user satisfaction, surfacing actionable failure and success modes. Serving hundreds of millions of users, ProxyDrift enables continuous drift monitoring and targeted synthetic data generation without exposing sensitive user data. Experiments confirm strong roundtrip consistency, discriminator-level indistinguishability of synthetic queries from human queries, and tight end-to-end alignment (RA~0.9) with production.

Beyond Direct Identifiers: Probabilistic Privacy Risk Estimation for Privacy-Conscious LLM Query Delegation

Recent work on protecting privacy during user-LLM interactions often focuses on direct, explicit identifiers: the personally-identifiable information (PII) captured by standard detectors. One such approach is Privacy-Conscious Delegation (PCD), where a local LLM acts as an intermediary. However, privacy risk does not stem solely from explicit identifiers but also PII-free self-disclosures, leaving users identifiable through combinations of quasi-identifying traits. We investigate a probabilistic variant of PCD, where we augment its objectives with an LLM-driven probabilistic estimation of k-anonymity. To facilitate this, we first create the PUPA-SD dataset, which contains naturalistic user queries with self-disclosure. Our preliminary results indicate that optimizing PAPILLON on PUPA-SD improves quality on unseen conversations across a variety of local models and produces the best privacy-utility balance for Llama-3.2-3B, while smaller models struggle to jointly optimize quality and privacy. We propose k-anonymity as a useful auxiliary metric for tackling PCD.

DPEI: Privacy Budget Savings for Edge Information Protection in Synthetic Graph Publishing

Can We Explain What We Anonymize? On the Impact of Data Anonymization on Post-hoc Model Explanations

Privacy-preserving data publishing and explainable artificial intelligence (XAI) are both essential for trustworthy machine learning, yet their interaction remains largely underexplored. In practice, models are often trained on anonymized datasets, but little is known about how classical anonymization techniques affect post-hoc explanations. In this paper, we provide a systematic empirical study of how feature attribution rankings change under widely used anonymization models, including k-anonymity, $\ell$-diversity, t closeness, and $(\alpha, k)$-anonymity. Across multiple real-world datasets and classifiers, we compare explanations generated by SHAP and LIME and quantify their stability using rank correlation and hypothesis testing. Our findings reveal a fundamental trade-off: explainable privacy-preserving models are feasible under mild privacy constraints, but strict anonymization requirements often lead to unstable explanations and severe utility degradation.