Search papers, labs, and topics across Lattice.
2
0
5
4
You can steal secrets from locally fine-tuned LLMs by backdooring their model code, even bypassing common defenses like differential privacy and code audits.
Achieve near-plaintext LLM inference speeds with strong privacy guarantees and minimal accuracy loss by jointly obfuscating data and model parameters – a first for models at the 671B scale.
