0
$\begingroup$

It is fairly agreed in literature that from a given face-embedding (that is a vector of features values) it is possible, with a good amount of effort, to reconstruct the original face, (See here for instance) assuming to have access to the embedding extractor.

Now, for concreteness, let's say that I've a 2000 long embedding of floats. I randomly rearrange the 2000 entries of the vector in one the $2000!$ permutations. The random rearrangement is specific to the owner of the face.

Is it still possible to recover the original face?

(A brute force attack is clearly excluded, as $2000! \approx 10^{5735}$)

$\endgroup$
0
$\begingroup$

if you will permute input you would have to also permute coeficients in first layer of network and if these permutations "will agree" then yes, otherwise you will just generate face on random input vector

$\endgroup$
  • $\begingroup$ I would first generate the embeddings and then randomise them. Knowing the randomisation pattern, it's always possible to recover the original embedding. $\endgroup$ – Rexcirus Oct 12 at 13:25

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.