Resisting Adversarial Attacks in Deep Neural Networks using Diverse Decision Boundaries

Name: Resisting Adversarial Attacks in Deep Neural Networks using Diverse Decision Boundaries
Uploaded: 2026-10-10T05:13:08.791245+00:00
Duration: 303 s
Description: The paper proposes an ensemble defense against adversarial attacks by creating models with diverse decision boundaries. It uses two methods—Split-and-Shuffle and Contrast-Significant-Features—to train a detector model that forces different models to make different classifications, making it harder for attackers to fool all of them simultaneously.

The paper proposes an ensemble defense against adversarial attacks by creating models with diverse decision boundaries. It uses two methods—Split-and-Shuffle and Contrast-Significant-Features—to train a detector model that forces different models to make different classifications, making it harder for attackers to fool all of them simultaneously.

Open the full episode