Search papers, labs, and topics across Lattice.
Email:
1
0
2
Decoy activations fail to protect split-LLM training because zero-valued backward gradients unmask private frames with 100% accuracy, rendering forward-only privacy audits dangerously deceptive.
