Statistical Adversaries: Natural Backdoor-like Features in Vision Datasets
The episode discusses a paper titled "Statistical Adversaries: Natural Backdoor-like Features in Vision Datasets." The authors show that ordinary datasets, like ImageNet, contain statistical structures that can be used to create targeted perturbations pushing models toward wrong answers without any malicious injection. The key finding is that these dataset statistics can be used to build adversarial directions, shifting the focus from model attacks to data vulnerabilities.