Search papers, labs, and topics across Lattice.
1
0
3
6
Multimodal models are surprisingly vulnerable to subtle text-based backdoor attacks using common words, making them far more practical and stealthy than previous visual trigger methods.
