Search papers, labs, and topics across Lattice.
1
0
2
1
Even with differential privacy guarantees ($\epsilon=200$), federated learning models still leak membership information, as demonstrated by a novel stacking-based attack that outperforms existing methods.