This paper outlines how, under certain circumstances, text embeddings can be used to reconstruct the original embedded text.<br /><br />OUTLINE:<br />0:00 - Intro<br />6:50 - Vec2Text: Iterative Embedding Inversion<br />12:20 - How to train this?<br />21:20 - Experimental results<br />26:10 - How can we prevent this?<br />31:20 - Some thoughts on sequence lengths<br /><br />Paper: <a href="https://arxiv.org/abs/2310.06816" target="_blank" rel="nofollow">https://arxiv.org/abs/2310.06816</a><br /><br />Abstract:<br />How much private information do text embeddings reveal about the original text? We investigate the problem of embedding \textit{inversion}, reconstructing the full text represented in dense text embeddings. We frame the problem as controlled generation: generating text that, when reembedded, is close to a fixed point in latent space. We find that although a naïve model conditioned on the embedding performs poorly, a multi-step method that iteratively corrects and re-embeds text is able to recover 92% of 32-token text inputs exactly. We train our model to decode text embeddings from two state-of-the-art embedding models, and also show that our model can recover important personal information (full names) from a dataset of clinical notes. Our code is available on Github<br /><br />Authors: John X. Morris, Volodymyr Kuleshov, Vitaly Shmatikov, Alexander M. Rush<br /><br />Links:<br />Homepage: <a href="https://ykilcher.com" target="_blank" rel="nofollow">https://ykilcher.com</a><br />Merch: <a href="https://ykilcher.com/merch" target="_blank" rel="nofollow">https://ykilcher.com/merch</a><br />YouTube: <a href="https://www.youtube.com/c/yannickilcher" target="_blank" rel="nofollow">https://www.youtube.com/c/yannickilcher</a><br />Twitter: <a href="https://twitter.com/ykilcher" target="_blank" rel="nofollow">https://twitter.com/ykilcher</a><br />Discord: <a href="https://ykilcher.com/discord" target="_blank" rel="nofollow">https://ykilcher.com/discord</a><br />LinkedIn: <a href="https://www.linkedin.com/in/ykilcher" target="_blank" rel="nofollow">https://www.linkedin.com/in/ykilcher</a><br /><br />If you want to support me, the best thing to do is to share out the content :)<br /><br />If you want to support me financially (completely optional and voluntary, but a lot of people have asked for this):<br />SubscribeStar: <a href="https://www.subscribestar.com/yannickilcher" target="_blank" rel="nofollow">https://www.subscribestar.com/yannickilcher</a><br />Patreon: <a href="https://www.patreon.com/yannickilcher" target="_blank" rel="nofollow">https://www.patreon.com/yannickilcher</a><br />Bitcoin (BTC): bc1q49lsw3q325tr58ygf8sudx2dqfguclvngvy2cq<br />Ethereum (ETH): 0x7ad3513E3B8f66799f507Aa7874b1B0eBC7F85e2<br />Litecoin (LTC): LQW2TRyKYetVC8WjFkhpPhtpbDM4Vw7r9m<br />Monero (XMR): 4ACL8AGrEo5hAir8A9CeVrW8pEauWvnp1WnSDZxW7tziCDLhZAGsgzhRQABDnFy8yuM9fWJDviJPHKRjV4FWt19CJZN9D4n<br />...<br /><a href="https://www.youtube.com/watch?v=FY5j3P9tCeA" target="_blank" rel="nofollow">https://www.youtube.com/watch?v=FY5j3P9tCeA</a>