A New Type of Adversarial Examples

Name: A New Type of Adversarial Examples
Uploaded: 2026-09-04T09:30:58.518934+00:00
Duration: 425 s
Description: The episode analyzes 'A New Type of Adversarial Examples,' a method that creates highly distorted inputs while maintaining correct classification. Unlike traditional attacks, this achieves stability despite massive visual divergence. The discussion covers iterative methods like NI-FGSM and findings demonstrating how these examples pose a significant, wide-ranging threat to AI system robustness.

The episode analyzes 'A New Type of Adversarial Examples,' a method that creates highly distorted inputs while maintaining correct classification. Unlike traditional attacks, this achieves stability despite massive visual divergence. The discussion covers iterative methods like NI-FGSM and findings demonstrating how these examples pose a significant, wide-ranging threat to AI system robustness.

Open the full episode