data aggregation

Large language models (LLMs) often inadvertently disclose personally identifiable information (PII) when generating summaries of sensitive-domain texts (e.g., clinical or legal documents), posing serious privacy risks. Method: We conduct the first systematic, cross-model evaluation of PII leakage in privacy-preserving summarization across six closed- and open-source LLMs—spanning diverse architectures and parameter scales—and benchmark two privacy-control paradigms: prompt engineering and fine-tuning. Evaluation integrates quantitative metrics (PII recall rate, privacy score), qualitative analysis, and human assessment. Results: All tested LLMs exhibit significant PII leakage, with privacy protection substantially inferior to human summarizers; humans achieve near-zero PII exposure while preserving summary quality. This work reveals a critical privacy-security bottleneck in LLM-based sensitive-text summarization and establishes a rigorous, empirically grounded evaluation framework to advance trustworthy AI deployment in high-stakes domains.

This work addresses the severe degradation of privacy protection for high-influence records (e.g., top earners in income or corporate salary data) under traditional differential privacy mechanisms when applied to highly skewed datasets. We propose a novel per-record differential privacy framework wherein privacy loss decays logarithmically with record influence. Our method achieves graceful privacy degradation in unbounded statistical tasks (e.g., summation) via influence-adaptive noise scaling and theory-driven sensitivity control. Crucially, we break the conventional linear or quadratic decay barriers, establishing—for the first time—a provable logarithmic privacy loss decay guarantee. Experiments on real-world corporate salary data demonstrate that our approach significantly strengthens privacy protection for high-value records while reducing utility loss by over 40% compared to baseline methods. Moreover, it preserves estimator unbiasedness and satisfies strict differential privacy compliance.

This work addresses the robustness of mean estimation in statistical learning under three concurrent challenges: adversarial data contamination, heavy-tailed distributions, and differential privacy constraints. Methodologically, it unifies robust statistics, high-dimensional geometry, stochastic optimization, and differential privacy theory to establish the first conceptual and algorithmic bridge across distinct robustness paradigms. Key technical abstractions—including iterative filtering, covariance trimming, and fractional gradient descent—are identified as common algorithmic primitives. The paper proposes a suite of computationally efficient estimators achieving statistically optimal convergence rates; each attains the information-theoretic lower bound under all three constraint classes simultaneously. By reconciling theoretical tightness with practical efficiency, this framework advances robust mean estimation from ad hoc heuristics toward a principled, unified design paradigm.

This paper addresses critical inefficiencies in multi-party secure collaborative analytics—specifically, low efficiency of Join-GroupBy-Aggregate (JGA) queries, high redundant join overhead, and expensive dynamic updates. To this end, we propose the first workload-agnostic materialized view framework tailored for JGA. Our core contributions are threefold: (1) an MPC-free incremental view maintenance mechanism; (2) a suite of specialized secure GroupBy-Aggregate protocols that bypass generic MPC bottlenecks; and (3) precomputed joins with view reuse to eliminate redundant join operations. Evaluated on real-world query workloads, our framework achieves up to 308.9× speedup over state-of-the-art approaches, while significantly reducing both communication and computation costs. The design rigorously preserves security guarantees under standard cryptographic assumptions, and supports efficient dynamic data updates—thus jointly optimizing security, performance, and adaptability.

When multiple data owners cannot share raw data, conventional secure collaborative analytics (SCA) systems face three fundamental challenges: unbounded privacy loss, suboptimal query planning, and lossy analysis. To address these, this paper proposes the first private-summary-based collaborative analytics framework, which simultaneously achieves bounded privacy loss, high-level query planning, and lossless analysis. Our method introduces precomputed differentially private indexes and operation-cost estimation to eliminate runtime privacy overhead; employs a one-sided noise mechanism, private upper-bound estimation, and data-agnostic index construction; and integrates differential privacy–enhanced execution plan optimization. Experimental results demonstrate that, compared to state-of-the-art approaches, our framework accelerates queries by up to 80×, reduces memory consumption for complex queries by up to 900×, and decreases cumulative privacy loss by up to 89× in continuous analytical workloads.

This study addresses the risk that embeddings derived from large language models—such as final-token hidden states or mean-pooled representations—can inadvertently leak sensitive patient attributes like race, even when the original clinical text is restricted. The work reveals, for the first time, significant variation in the degree of privacy leakage across different embedding types and introduces SurfaceLoRA, a targeted privacy-preserving fine-tuning approach tailored to specific output representations. By integrating a gradient reversal discriminator with parameter-efficient fine-tuning, SurfaceLoRA effectively suppresses sensitive information in the target embeddings while preserving clinical summarization performance. Experimental results under five balanced probing protocols demonstrate that SurfaceLoRA reduces race recoverability from final-token embeddings to near-random levels, underscoring its efficacy in providing representation-specific privacy protection.

This study addresses the privacy risks posed by small-count cells in frequency tables released by statistical agencies. To mitigate these risks while preserving data utility, the authors propose information loss-bounded aggregation (iLBA), a novel approach that integrates small cell adjustment (SCA) with controlled tabular rounding. The method generates confidential summary tables under strict bounds on information loss, effectively balancing confidentiality and analytical usefulness. Notably, this work provides the first open-source R package implementing iLBA, enabling users to construct masked fine-grained tables, produce disclosure-controlled summary tables, and perform single-cell queries. The implementation supports efficient, reproducible disclosure control for tabular data, offering a practical solution for official statistics dissemination.

Protecting confidential data while preserving utility is particularly challenging when data sets contain outlying observations. Existing latent space anonymization methods, such as spectral anonymization (SA), rely on principal component analysis (PCA) and may therefore be vulnerable to contamination. We investigate anonymization in the presence of outliers and propose ICSA, a robust alternative to SA based on invariant coordinate selection (ICS). By replacing the PCA transformation with ICS, the robustness of the anonymization procedure can be regulated through the choice of scatter matrices. Alongside the methodological development, we derive a theoretical result showing that SA fails under sufficiently influential outliers. To assess the practical implications of this result, we compare the privacy-utility trade-off of ICSA and SA through simulation experiments under varying contamination settings and outlier severities. Our findings indicate that implementations of ICSA based on robust scatter matrices achieve stronger privacy protection than SA, while typically maintaining comparable, and in some cases improved, utility. We further examine the empirical performance of the proposed method using a benchmark clinical data set, where ICSA demonstrates superior overall privacy-utility efficiency relative to SA. These results suggest that explicitly accounting for outliers can materially improve anonymization performance and that robust latent space transformations offer a promising direction for privacy-preserving statistical data release.

This study addresses the limitations of traditional TCLUST in high-dimensional settings, where complex parameter estimation and sensitivity to outliers degrade performance. To overcome these issues, the authors propose tHHDC, a novel approach that integrates TCLUST’s robust trimming mechanism with the subspace modeling framework of High-Dimensional Data Clustering (HDDC). The method introduces eigenvalue constraints to enhance model stability and incorporates an effective parameter selection strategy. tHHDC successfully mitigates the shortcomings of existing methods regarding intersecting subspace structures and restrictive assumptions on error distributions. Extensive experiments on both simulated and real-world datasets demonstrate its superior robustness and accuracy in clustering high-dimensional noisy data. The work is further supported by a complete algorithmic implementation and rigorous theoretical analysis.