analiza aiqbench
The paper systematically tests how four attack types (membership inference, attribute inference, data extraction, and backdoors) vary with model and data factors, rather than studying each in isolation. For LLM developers, the practical signal is that leakage risk depends on specific choices: retrieval-augmented setups, repeated personal data in training sets, and unprotected fine-tuning pipelines. The authors' recommendations are concrete, though their generalizability across models and datasets should be checked against the full study.
On the Privacy of LLMs: An Ablation Study
arXiv:2605.02255v2 Announce Type: replace-cross Abstract: Large language models (LLMs) face diverse privacy threats, yet existing studies typically evaluate individual attacks in isolation, offering limited guidance on which system factors drive privacy vulnerabilities. This paper presents a unified threat model and a systematic ablation study of four representative attacks: Membership Inference (MIA), Attribute Inference (AIA), Data Extraction (DEA), and Backdoor Attacks (BA). Our findings reveal attack-specific sensitivities: membership inference depends strongly on dataset characteristics and attack strategy, attribute inference is influenced by model and training characteristics, data extraction increases with model scale and training-data repetition, and backdoor attacks achieve consistently high success rates. Based on these findings, we derive practical recommendations for privacy-aware LLM deployment, including prioritizing membership leakage assessments in retrieval-augmented systems, minimizing repeated personal information in training data, monitoring attribute inference risks, and safeguarding fine-tuning pipelines against backdoors.
Abstract:Large language models (LLMs) face diverse privacy threats, yet existing studies typically evaluate individual attacks in isolation, offering limited guidance on which system factors drive privacy vulnerabilities. This paper presents a unified threat model and a systematic ablation study of four representative attacks: Membership Inference (MIA), Attribute Inference (AIA), Data Extraction (DEA), and Backdoor Attacks (BA). Our findings reveal attack-specific sensitivities: membership inference depends strongly on dataset characteristics and attack strategy, attribute inference is influenced by model and training characteristics, data extraction increases with model scale and training-data repetition, and backdoor attacks achieve consistently high success rates. Based on these findings, we derive practical recommendations for privacy-aware LLM deployment, including prioritizing membership leakage assessments in retrieval-augmented systems, minimizing repeated personal information in training data, monitoring attribute inference risks, and safeguarding fine-tuning pipelines against backdoors.
Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI)
arXiv:2605.02255 [cs.CR]
(or arXiv:2605.02255v2 [cs.CR] for this version)
https://doi.org/10.48550/arXiv.2605.02255 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 230 cuvinte
Subiecte conexe
Research Note: Filtering Subversion-Relevant Information From Pretraining Data Is Feasible
LessWrong · 5 oct. 2026
The Interlocutor Effect: Why LLMs Leak More Personal Data to Agents Than Humans
arXiv · 10 iun. 2026
Natural Language Access to Domain-Specific Metadata: A Reusable Framework for LLM Query Generation
arXiv · 21 iul. 2026
Explaining and Tuning Transformer-based LLMs in Arithmetic Tasks with Human Strategies
arXiv · 21 iul. 2026