Harnessing the Universal Geometry of Embeddings (Category Security)

Literally the other day I performer Speaking at PHDays about reliability and security as a fundamental part of our systems, highlighting the new risks from AI. I thought the threat model was relatively clear to me, but yesterday I read Sergey Nikolenko’s Sinekur channel. (@sinecor)I came across a story about whitepaper.Harnessing the Universal Geometry of Embeddings", in which the authors literally developed the vec2vec method to transform vectors from a single representation space (embeddings) The other. Sergei is further in his post explaining

Nothing seems super-genius in the method itself is not, but the quality of the result is amazing: 3 It is shown how five vectors almost orthogonal to each other turned into vectors with scalar products from each other. 0.8 before 0.95.

In practice, it means a lot, and mostly not very good. It turns out that databases that contain vector representations need to be protected as carefully as the source code. (What no one is doing now, as far as I know.). In one experiment, they took Enron’s corporate e-mail embeddings and moved them into the space of a well-known model. (rice. 4) And they were able to extract sensitive information. (names, dates, amounts) from 80Percent of documents. In general, another interesting vector of attack, only not on the bases with original data, but on vector bases with embeddings. Maybe I'll get my hands on this whitepaper soon:)

P.S. I subscribed to the channel Sergey Nikolenko after I wrote recall The book is "Deep Learning." Immersion in the world of neural networks 2018 One year, co-authored by Sergey. The bottom line is that one of the channel's subscribers said that the author has a cool tg channel, I checked - the channel turned out to be really interesting and now I sometimes read it. Thank you for this recommendation!

#Security #AI #Software #Engineering #Database #Data