## **PoC**
Reconstruct text sequences from embeddings as well as run pre-trained models.
The library can be used to embed text and then invert it, or invert directly from embeddings.
**Most of the code in this repository facilitates training inversion models**
https://github.com/jxmorris12/vec2text
## **Details**
It has been previously assumed that embeddings could not be reconstructed to reveal the sensitive data within, that should never be the assumption.
[video](https://www.youtube.com/watch?v=lguThumFUl4)
[Code demo](https://colab.research.google.com/drive/14RQFRF2It2Kb8gG3_YDhP_6qE0780L8h?usp=sharing )
[Paper](https://arxiv.org/abs/2310.06816)
A paper for a multilingual implementation of this attack exists [paper](https://arxiv.org/pdf/2401.12192.pdf)
paper exists for editing multimodal large language models [paper](https://arxiv.org/abs/2310.08475))
ID: AML.T0057