Mirage: a Clean-Label Backdoor against LiDAR 3D Object Detection
The episode discusses the paper "Mirage: a Clean-Label Backdoor against LiDAR 3D Object Detection." The hosts explain how this method uses specific poison samples during training to make an AI associate a trigger pattern with a target class. They conclude that defenses must be tailored to detect physically plausible, malicious point-cloud structures rather than just label errors.