Kato Mivule

The paper introduces LLM-CEG, an extended framework that uses membership inference attack success rates and model perplexity to systematically audit and optimize the privacy-utility trade-off when fine-tuning Large Language Models.