How Private are Language Models in Abstractive Summarization?
Not on the map. Why?
Abstract (arXiv)
In sensitive domains such as medical and legal, protecting sensitive information is critical, with protective laws strictly prohibiting the disclosure of personal data. This poses challenges for sharing valuable data such as medical reports and legal cases summaries. While language models (LMs) have shown strong performance in text summarization, it is still an open question to what extent they can provide privacy-preserving summaries from non-private source documents. In this paper, we perform a comprehensive study of privacy risks in LM-based summarization across two closed- and four open-weight models of different sizes and families. We experiment with both prompting and fine-tuning strategies for privacy-preservation across a range of summarization datasets including medical and legal domains. Our quantitative and qualitative analysis, including human evaluation, shows that LMs frequently leak personally identifiable information in their summaries, in contrast to human-generated privacy-preserving summaries, which demonstrate significantly higher privacy protection levels. These findings highlight a substantial gap between current LM capabilities and expert human expert performance in privacy-sensitive summarization tasks.
Key facts. Cited by 2 papers in the map (data through 13 Oct 2026). Builds on 41 of its 94 references that are in the map. First posted on arXiv 16 Dec 2024.
Builds on
41 of 94 references.
References · 41
- 2024The Llama 3 Herd of Modelscited by 14,486
- 2019BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehensioncited by 5,884
- 2019BERTScore: Evaluating Text Generation with BERTcited by 5,480
- 2016Deep Learning with Differential Privacycited by 2,718
- 2019PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarizationcited by 1,100
- 2016Abstractive Text Summarization Using Sequence-to-Sequence RNNs and Beyondcited by 810
- 2019Evaluating the Factual Consistency of Abstractive Text Summarizationcited by 580
- 2023Benchmarking Large Language Models for News Summarizationcited by 444
- 2023Analyzing Leakage of Personally Identifiable Information in Language Modelscited by 284
- 2023Adapted Large Language Models Can Outperform Medical Experts in Clinical Text Summarizationcited by 246
- 2023Can LLMs Keep a Secret? Testing Privacy Implications of Language Models via Contextual Integrity Theorycited by 223
- 2022What Does it Mean for a Language Model to Preserve Privacy?cited by 195
29 more
- 2019Privacy- and Utility-Preserving Textual Analysis via Calibrated Multivariate Perturbationscited by 118
- 2022Preventing Verbatim Memorization in Language Models Gives a False Sense of Privacycited by 99
- 2022The Text Anonymization Benchmark (TAB): A Dedicated Corpus and Evaluation Framework for Text Anonymizationcited by 95
- 2021Does BERT Pretrained on Clinical Notes Reveal Sensitive Data?cited by 93
- 2022Evaluating the Factual Consistency of Large Language Models Through News Summarizationcited by 91
- 2023TrueTeacher: Learning Factual Consistency Evaluation with Large Language Modelscited by 82
- 2023On Learning to Summarize with Large Language Models as Referencescited by 80
- 2023Element-aware Summarization with Large Language Models: Expert-aligned Evaluation and Chain-of-Thought Methodcited by 76
- 2024Follow My Instruction and Spill the Beans: Scalable Data Extraction from Retrieval-Augmented Generation Systemscited by 76
- 2021Selective Differential Privacy for Language Modelingcited by 61
- 2020Dr. Summarize: Global Summarization of Medical Dialogue by Exploiting Local Structurescited by 49
- 2024AgentsCourt: Building Judicial Decision-Making Agents with Court Debate Simulation and Legal Knowledge Augmentationcited by 39
- 2024GoldCoin: Grounding Large Language Models in Privacy Laws via Contextual Integrity Theorycited by 36
- 2019Scoring Sentence Singletons and Pairs for Abstractive Summarizationcited by 34
- 2023Large Language Models for Healthcare Data Augmentation: An Example on Patient-Trial Matchingcited by 30
- 2019Plain English Summarization of Contractscited by 28
- 2022Legal Case Document Summarization: Extractive and Abstractive Methods and their Evaluationcited by 23
- 2024FactPICO: Factuality Evaluation for Plain Language Summarization of Medical Evidencecited by 18
- 2024A Survey on Safe Multi-Modal Learning Systemcited by 15
- 2020PHICON: Improving Generalization of Clinical Text De-identification Models via Data Augmentationcited by 14
- 2024TriSum: Learning Summarization Ability from Large Language Models with Structured Rationalecited by 10
- 2024MetaSumPerceiver: Multimodal Multi-Document Evidence Summarization for Fact-Checkingcited by 10
- 2023Assessing Privacy Risks in Language Models: A Case Study on Summarization Taskscited by 8
- 2023Automatically Summarizing Evidence from Clinical Trials: A Prototype Highlighting Current Challengescited by 7
- 2023Automated Refugee Case Analysis: An NLP Pipeline for Supporting Legal Practitionerscited by 6
- 2022What to Read in a Contract? Party-Specific Summarization of Legal Obligations, Entitlements, and Prohibitionscited by 5
- 2024Automate or Assist? The Role of Computational Models in Identifying Gendered Discourse in US Capital Trial Transcriptscited by 4
- 2023Private Meeting Summarization Without Performance Losscited by 3
- 2023STRONG -- Structure Controllable Legal Opinion Summary Generationcited by 3
Built on by
Citing papers per year: 2025: 2.
2 papers.