Datasets:

---

pszemraj

/

LocalLLaMA-posts

like 0

ERROR: type should be string, got "https://preview.redd.it/5mf4dmhom01g1.jpg?width=1654&format=pjpg&auto=webp&s=b2b0e8a6cd49bde36263a1361a09d0734ce80470\n\nWe showcase that simple membership inference–style attacks can achieve over 60% success in predicting the presence of personally identifiable information (PII) in data input to LLMs  just by observing the privatized output, even when it doesn’t explicitly leak private information!\n\nTherefore, it’s imperative to use Differential Privacy (DP) with LLMs to protect private data passed to them. However, existing DP methods for LLMs often severely damage utility, even when offering only weak theoretical privacy guarantees.\n\nWe present DP-Fusion the first method that enables differentially private inference (at the token level) with LLMs, offering robust theoretical privacy guarantees without significantly hurting utility.\n\nOur approach bounds the LLM’s output probabilities to stay close to a public distribution, rather than injecting noise as in traditional methods. This yields over 6× higher utility (perplexity) compared to existing DP methods.\n\n📄 The arXiv paper is now live here: [https://arxiv.org/abs/2507.04531](https://arxiv.org/abs/2507.04531) \n💻 Code and data: [https://github.com/MBZUAI-Trustworthy-ML/DP-Fusion-DPI](https://github.com/MBZUAI-Trustworthy-ML/DP-Fusion-DPI)\n\n⚙️ Stay tuned for a PIP package for easy integration!"

ERROR: type should be string, got "https://preview.redd.it/5aqyfpfhk01g1.jpg?width=1654&format=pjpg&auto=webp&s=fb625930eb8fe36e6f8f52d6c7262ee0bfd4144f\n\nWe showcase that simple membership inference–style attacks can achieve over 60% success in predicting the presence of personally identifiable information (PII) in data input to LLMs  just by observing the privatized output, even when it doesn’t explicitly leak private information! \n \nTherefore, it’s imperative to use Differential Privacy (DP) with LLMs to protect private data passed to them. However, existing DP methods for LLMs often severely damage utility, even when offering only weak theoretical privacy guarantees. \n \nWe present DP-Fusion the first method that enables differentially private inference (at the token level) with LLMs, offering robust theoretical privacy guarantees without significantly hurting utility. \n \nOur approach bounds the LLM’s output probabilities to stay close to a public distribution, rather than injecting noise as in traditional methods. This yields over 6× higher utility (perplexity) compared to existing DP methods. \n \n📄 The arXiv paper is now live here: [https://arxiv.org/abs/2507.04531](https://arxiv.org/abs/2507.04531) \n 💻 Code and data: [https://github.com/MBZUAI-Trustworthy-ML/DP-Fusion-DPI](https://github.com/MBZUAI-Trustworthy-ML/DP-Fusion-DPI) \n \n ⚙️ Stay tuned for a PIP package for easy integration! \n\n\nlinkedin post - [https://www.linkedin.com/posts/rushil-thareja-a80400179\\_do-not-use-llms-to-privatize-your-data-activity-7393274346275807232-luBB?utm\\_source=share&utm\\_medium=member\\_desktop&rcm=ACoAACo9hgABughMeMAXoGUs0enD590i3TroMXM](https://www.linkedin.com/posts/rushil-thareja-a80400179_do-not-use-llms-to-privatize-your-data-activity-7393274346275807232-luBB?utm_source=share&utm_medium=member_desktop&rcm=ACoAACo9hgABughMeMAXoGUs0enD590i3TroMXM)"