The art of defense: letting networks fool the attacker

3D perception of objects is critical for many real-world applications, such as autonomous cars and robots. Among them, most state-of-the-art (SOTA) systems are based on deep learning models. Recently, the research community found that object classifiers point cloud easily fooled by adversarial craft attackers. To overcome this, defenses considered effective ways to improve robustness models, focused input transformation. However, all previous defense methods decrease natural accuracy, nature itself has been overlooked. this end, in paper, we propose a novel makes full use classifiers. Due disorder cloud, have property permutation invariant cloud. Based nature, design transformations (IT-Defense). We show that, even after accounting obfuscated gradients, our IT-Defense resilient against SOTA attacks. Moreover, does not hurt clean accuracy compared defenses. Our code will be available at: https://github.com/cuge1995/IT-Defense.