Alignment Defends LLMs from Property Inference Attacks

Abstract

Large language models (LLMs) are increasingly fine-tuned on domain-specific datasets that may contain sensitive, dataset-level properties. Recent work has shown that such dataset-level information can be effectively extracted through property inference attacks, posing a confidentiality risk. Existing defenses against these attacks primarily operate by modifying the training data distribution and hence require access to the original data and retraining the model, limiting their applicability to s

Categories

@misc{huang2026alignment,
  title = {{Alignment Defends LLMs from Property Inference Attacks}},
  author = {Pengrun Huang and Chhavi Yadav and Ruihan Wu and Kamalika Chaudhuri},
  year = {2026},
  month = jun,
  eprint = {2606.10217},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2606.10217}
}