Computer vision task confrontation sample generation method, terminal and medium

The invention discloses an adversarial sample generation method for a computer vision task, a terminal and a medium, and the method achieves a specific quantifiable constraint target for the disturbance invisibility of an adversarial sample in a fast gradient symbol method in the form of a structural similarity loss function, thereby generating the adversarial sample with higher disturbance invisibility. Meanwhile, the method can be combined with any prior art which globally or locally adopts a fast gradient symbol method, and has high expansibility and adaptability. 本发明公开了一种计算机视觉任务的对抗样本生成方法、终端及介质，本方法用结构相似性损失函数的形式在快速梯度符号法中为对抗样本的扰动不可见性实现了具体的可量化的约束目标，从而生成扰动不可见性更高的对抗样本，同时本方法可与任意全局或局部采用了快速梯度符号法的现有技术进行组合，具有高扩展性与适应性。