Anchor-frame-free model class gradient global confrontation sample generation method, system and equipment

The invention belongs to the technical field of deep learning adversarial attack, and discloses an anchor-frame-free model class gradient global adversarial sample generation method, system and equipment, and the method comprises the steps: collecting class gradients by taking a class as a unit through employing a target class in an input image, and carrying out the collection of class gradients; gradient collection is carried out according to the mode, the similarity on intra-class target gradients can be utilized, the influence of direct superposition disturbance on disturbance performance is avoided, and then picture scale disturbance is generated. On the basis, global disturbance aiming at a human body posture estimation task is formed based on a certain amount of image scale disturbance, the global disturbance overcomes the defect that an existing confrontation disturbance generation method is based on a picture or a post selection box, effective disturbance can be formed aiming at a large amount of data