Search papers, labs, and topics across Lattice.
1
0
3
Safety classifiers leak surprisingly sensitive information: a boundary-targeted attack recovers 19% of user distress conversations from the training data, far exceeding existing membership inference methods.
