核心概念
Gradient Shaping (GRASP) enhances backdoor attack detection by reducing trigger effective radius.
摘要
The content discusses the effectiveness of gradient-based trigger inversion in detecting backdoor attacks and introduces a new method, Gradient Shaping (GRASP), to enhance backdoor attack detection by reducing the trigger effective radius. The study analyzes the impact of GRASP on various environmental factors, learning optimizers, noise levels, and datasets. It evaluates the performance of GRASP against different backdoor detection methods and backdoor attacks across multiple datasets.
統計資料
Gradient Shaping (GRASP) reduces trigger effective radius.
GRASP enhances backdoor stealthiness through data poisoning.
GRASP is effective in evading trigger inversion techniques.
引述
"Our study shows that existing attacks tend to inject the backdoor characterized by a low change rate around trigger-carrying inputs."
"GRASP can be combined with existing stealthy backdoor methods to enhance their capability to evade trigger inversion-based defenses."