0
$\begingroup$

all, I am working on federated learning and here goes my question:

Suppose there are two participants to do the federated learning. For some of the models (e.g. Logistic Regression models, assume one party has some features $X_1$ and the label $y$, the other has some other features $X_2$. The coefficients are denoted as $W_1$ and $W_2$, respectively), the schemes use "full encryption/mask", which encrypt/secret share all the intermediate results of the training process.

However, there are some schemes which merely hides parts of the intermediate results, for the most "aggressive" one, only $W_2X_2$ is encrypted or secret shared. Then, it is sent to the other party, and the other party decrypts/recovers $W_2X_2$ and continues computation as in plaintext. These schemes argue that the exposure of $W_2X_2$ will not further reveal information of $X_2$ and is considered secure.

My question is how to evaluate the security of the later one, are there any possible attacks to recover the original data or to build a new model which is almost the same as the final federated model by only one of the participants?

$\endgroup$

1 Answer 1

1
$\begingroup$

Here's one attack : https://arxiv.org/pdf/2011.09290.pdf
Hope that helps.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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