Search papers, labs, and topics across Lattice.
1
0
4
Poisoned training data leaves a unique fingerprint in the spectral entropy of LLM gradients, enabling backdoor detection even at extreme poison ratios where clustering-based defenses fail.
