TD-CAG: Enhancing adversarial transferability via curvature awareness and spatial dislocation

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Transfer-based adversarial attacks evaluate deep neural network (DNN) robustness in black-box settings.
• Improving cross-model transferability of these attacks is crucial but challenging due to local linearity and saliency issues.

Purpose of the Study:

• To propose a novel adversarial attack framework, Translation-Dislocation and Curvature-Aware Gradient (TD-CAG), to enhance transferability.
• To address limitations of existing methods by modeling local nonlinearity and saliency misalignment.

Main Methods:

• Developed a curvature-aware gradient (CAG) module to approximate second-order directional curvature, mitigating local linearity.
• Introduced a translation-dislocation (TD) module with spatial shifts to simulate saliency misalignments across architectures.
• Designed TD-CAG as a plug-and-play framework compatible with existing attack pipelines.

Main Results:

• TD-CAG demonstrated consistently superior transferability compared to state-of-the-art adversarial attack methods.
• The framework maintained high compatibility with existing methods.
• TD-CAG achieved this improvement with a low computational cost.

Conclusions:

• TD-CAG effectively enhances the transferability of adversarial attacks on DNNs.
• The proposed modules successfully address key limitations in gradient approximation and saliency guidance.
• TD-CAG offers a practical and efficient solution for robust DNN evaluation.