differential-privacy

该论文提出SHIP（Structured and Hierarchical Insight with Privacy），将差分隐私洞察形式化为可复用的查询回答表示，通过自动构建层级混合分类法并发布DP合成表，实现隐私保护下的文本数据分析，性能优于现有私有洞察基线。

DP-MIND 提出一种记忆引导的框架，将差分隐私推理扩展到百万 token 长上下文场景，通过在扰动输入上用强化学习优化记忆构建策略并结合本地去噪模块，在长上下文基准上优于现有 DP 与长上下文推理基线。

论文提出 HXAI 框架，结合局部与分区两级模型，在分布式能源系统中实现隐私保护下的可解释 AI 分析，用于电网负荷管理，同时在差分隐私约束下保持解释的语义结构而非最小化数值误差。

Google Research introduces a new TEE-based federated learning system that provides fully verifiable and auditable privacy guarantees (remote attestation, confidentiality, integrity) with improved accuracy, now adopted by Gboard for faster compute.

The paper proposes clipping-aware objective conditioning for differentially private diffusion models applied to energy time-series imputation, using v-prediction and timestep-aware loss weighting to reduce pre-clipping gradient amplification and improve imputation utility over the epsilon-prediction baseline on five real-world energy datasets.

QuanText is a training-free and large-language-model-agnostic mechanism proposed to protect global secrets in textual datasets while preserving utility, using randomized quantization inspired by Statistic Maximal Leakage and achieving better empirical privacy-utility trade-offs than competing baselines.

Introduces DP-SPIN, a differentially private framework for aggregate measurement and summarization over semantic concepts in text data, with evaluations on CFPB, Amazon, and Yelp datasets.

This paper proposes a style-aware, prompt-driven approach for text anonymization using large language models to suppress identifiable stylistic markers while preserving meaning, reducing authorship attribution F1 by 60–70% and outperforming differential privacy-based methods.

The paper introduces α-split, a two-pool allocation method for differential privacy in federated learning, addressing cross-component budget collapse in speech-LLMs and improving utility and security against attacks.

RAPTOR is a role-aware framework for differentially private fine-tuning of Mixture-of-Experts models, addressing failure modes like clipping interference and noise dilution, with experimental improvements over standard DP baselines.

This paper demonstrates that causal fairness mechanisms, specifically edge cuts on causal graphs, are portable across various synthetic data generator families including GANs and diffusion models, with minimal impact on data fidelity and utility.

This paper examines privacy failures in split-LLM training, where returned gradients can nullify decoy mechanisms and enable gradient inversion attacks, undermining data protection.

The paper reveals a privacy-hallucination tradeoff in differentially private language models, where stricter privacy budgets increase hallucination risks and explores mitigation strategies.

The paper proposes a lightweight defense framework to enhance the empirical privacy of DP-SGD without theoretical privacy cost, validated through extensive audits across models, datasets, and threat models.

The paper investigates the geometric challenges in merging differentially private task models and introduces DP-Merging, a framework to enhance mergeability while maintaining privacy guarantees.

This paper introduces Private Best-of-N (PrivBoN) and Private Inference-Time Pessimism (PrivITP) methods that add calibrated noise to reward scores in inference-time alignment to achieve differential privacy and mitigate reward hashing, with minimal additional alignment cost.

The US Commerce Department is circulating a proposal to stop collecting census data on race and sexual orientation, and to exclude undocumented immigrants from the count for the 2030 census, which could reshape federal funding and political representation.

This paper introduces StraightDP, a geometry-aware differential privacy framework for text-conditioned rectified-flow transformers. It partitions the privacy budget to release class-conditional moments and use DP-SGD, improving accuracy and FID over uniform DP training at strong privacy levels.

Noisegate is an open-source differential-privacy gateway that allows untrusted AI agents to query sensitive datasets with mathematical guarantees against record leakage. It features built-in attack testing, validated noise mechanisms, and acts as an MCP server for Claude Desktop.

DP-FedSOFIM moves curvature estimation to the server in differentially private federated learning, achieving the same privacy guarantee as DP-FedGD with O(d) client memory and significant early-round accuracy gains.