arXiv Machine Learning
Balancing Privacy, Utility, and Safety in LLM Alignment through Preference Optimization
Read the original on arXiv Machine Learning →The Flow has not summarised this story yet — read it at arXiv Machine Learning.
Related stories
arXiv Machine Learning
Jun 9
Benchmarking Empirical Privacy Protection for Adaptations of Large Language Models
arXiv:2606. 09401v1 Announce Type: new Abstract: Recent work has applied differential privacy (DP) to adapt large language models (LLMs) for sensitive applications, offering theoretical guarantees.
arXiv AI
Jun 30
Decomposing Memorization Reduction in Privacy-Preserving Fine-Tuning of SLMs for CSIRTs
arXiv:2606. 28479v1 Announce Type: cross Abstract: CSIRTs increasingly fine tune language models on vulnerability scan records, but these records expose internal network topology and create privacy risks under regulations such as GDPR and LGPD.
arXiv AI
Jun 10
Advancing the State-of-the-Art in Empirical Privacy Auditing
arXiv:2606. 10481v1 Announce Type: cross Abstract: Parameter-efficient fine-tuning of large language models (LLMs) can exhibit problematic memorization of individual training examples.
arXiv Computation and Language
Sep 2
Membership Inference in Fine-tuned Diffusion Language Models via Token-level Memorization Asymmetry
The paper introduces a new privacy vulnerability in diffusion language models (DLMs) called token‑level memorization asymmetry, derived from theoretical analysis of diffusion training dynamics. It proposes Q‑Skew, a quantile‑weighted skewness indicator, to perform membership inference on fine‑tuned DLMs, outperforming existing baselines across multiple datasets and models. Additionally, Q‑Skew can be used to extract personally identifiable information (PII), demonstrating a broader privacy attack surface.
arXiv AI
Sep 4
User Perceptions vs. Proxy LLM Judges: Privacy and Helpfulness in LLM Responses to Privacy-Sensitive Scenarios
The study investigates how users perceive the helpfulness and privacy-preservation of large language model (LLM) responses in privacy-sensitive scenarios. Using 94 participants and 90 PrivacyLens scenarios, researchers found that users’ evaluations of identical LLM outputs varied widely, whereas five proxy LLM judges showed high agreement but low correlation with user judgments. The results suggest that proxy LLMs cannot reliably estimate users’ diverse perceptions of utility and privacy, highlighting the need for more user-centered evaluation methods.
arXiv AI
Aug 11
CIDER: A Dataset of Contextual Disclosure Boundaries for Privacy Preference Alignment
arXiv:2608. 09164v1 Announce Type: new Abstract: Aligning large language models (LLMs) with human privacy preferences requires capturing individuals' disclosure boundaries beyond general privacy norms.