Enhancing adversarial transferability via curvature-aware penalization

Key Points

• Improved adversarial transferability shows a 20% increase when curvature-awareness is applied.
• Key evidence indicates that penalizing curvature leads to a significant boost in model performance.
• This analysis uses curvature-aware penalization strategies across various attack types on neural networks.
• Highlights the need for adjusting model training methods to enhance defenses against adversarial attacks.