Multi-view defense method based on noise reduction self-coding

Abstract

The invention discloses a multi-view defense method based on noise reduction self-coding, and the method comprises the steps: firstly training a multi-view adversarial self-coding model through the original samples and adversarial samples of all views before a defense capability is provided; and after the training is finished, acquiring the confrontation and defense capability. The innovation point of the invention is to construct a multi-view adversarial self-encoding model capable of effectively improving the adversarial robustness of a multi-view depth model. According to the multi-view adversarial self-encoding model, noise reduction self-encoding serves as a prototype, the weight of each view is adjusted in a self-adaptive mode according to the reconstruction error, and pre-noise reduction processing on multi-view samples is achieved. Robustness of the multi-view depth model to the anti-attack is improved, the original performance of the model on a clean sample is maintained, and meanwhile, good generalization is shown for other unseen anti-disturbance.

Description

technical field [0001] The present invention relates to the field of computer technology, to confrontation defense technology, and in particular to a multi-view defense method based on noise reduction autoencoding. Background technique [0002] The background technology involves four major contents, namely multi-view learning, noise reduction autoencoder model, adversarial attack (multi-view adversarial attack), and adversarial defense. [0003] 1) Multiview Learning [0004] Since single-view learning only utilizes data from a single view, the information obtained is limited. Using the consistency and complementarity between different views can often further improve the performance of the model, so multi-view learning is introduced. When only one natural view is available, multi-view learning can be done by artificially generating other views. Even so, the performance of the model can be improved. [0005] According to different strategies for using multi-view data, exi...

AI Technical Summary

Problems solved by technology

Adversarial training can indeed alleviate the vulnerability of the model to adversarial attacks, but it has certain limitations: adversarial training can only defend against specific types of adversarial perturbations, and is still vulnerable to other unseen adversarial perturbations

Images