Self-Refining Language Model Anonymizers via Adversarial Distillation (2506.01420v1)

Abstract: LLMs are increasingly used in sensitive domains, where their ability to infer personal data from seemingly benign text poses emerging privacy risks. While recent LLM-based anonymization methods help mitigate such risks, they often rely on proprietary models (e.g., GPT-4), raising concerns about cost and the potential exposure of sensitive data to untrusted external systems. To address this, we introduce SElf-refining Anonymization with LLM (SEAL), a novel distillation framework for training small LLMs (SLMs) to perform effective anonymization without relying on external costly models at inference time. We leverage adversarial interactions between an LLM anonymizer and an inference model to collect trajectories of anonymized texts and inferred attributes, which are used to distill anonymization, adversarial inference, and utility evaluation capabilities into SLMs via supervised fine-tuning and preference learning. The resulting models learn to both anonymize text and critique their outputs, enabling iterative improvement of anonymization quality via self-refinement. Experiments on SynthPAI, a dataset of synthetic personal profiles and text comments, demonstrate that SLMs trained with SEAL achieve substantial improvements in anonymization capabilities. Notably, 8B models attain a privacy-utility trade-off comparable to that of the GPT-4 anonymizer and, with self-refinement, even surpass it in terms of privacy. These results show the effectiveness of our adversarial distillation framework in training SLMs as efficient anonymizers. To facilitate further research, we release the full dataset used in our experiments.