Search papers, labs, and topics across Lattice.
3
0
6
Adversarial decoys can effectively misdirect attention-based defenses, revealing that attention magnitude is not a reliable indicator of adversarial relevance.
Surprisingly, even a non-robust pre-trained model can be adapted at test-time to achieve significantly better adversarial robustness than existing methods, using a novel label-free semantic anchoring technique.
Bridging the sim-to-real gap in railway perception, a new augmented reality framework generates realistic training data by seamlessly integrating virtual objects into real-world railway scenes, outperforming traditional methods.
