Data as of Jul 25, 2026 · Based on 2,718,867 AI responses across 9,511 prompts · See how Parse measures this
SOFTsoft-mia.github.io · Low evidenceSOFT is a defense technique that protects large language models from membership inference attacks during fine-tuning by selectively obfuscating influential training samples. It replaces vulnerable data with paraphrased alternatives to balance privacy protection and model utility, as demonstrated in a USENIX Security'25 paper from Purdue University and Cisco Research.
Parse Score