Adversarial sample generation method, detector training method and related equipment

Abstract

The invention provides an adversarial sample generation method, a detector training method and related equipment. The method for generating the adversarial sample comprises the following steps: acquiring a clean training sample set; for each training sample in the training sample set, performing adversarial attack on the training sample to obtain a plurality of adversarial instances; selecting a target label according to the confrontation attack information of each target confrontation instance indicating successful confrontation in the plurality of confrontation instances; generating an adversarial sample corresponding to the training sample based on the target label; and storing the confrontation samples. The detector training method comprises the following steps: generating an adversarial sample set based on a training sample set by utilizing an adversarial sample generation method; performing first training of a binary classification task on the detector by using the training sample set; and performing second training on the detector subjected to the first training by using the confrontation sample set. The detector obtained through the training method is used for detecting whether the sample data input into the deep neural network contains the deep neural network Trojan horse or not.

Description

technical field [0001] This application relates to neural network model training, and in particular to a method for generating adversarial samples, a method for training a detector, and related equipment. Background technique [0002] In the past few decades, machine learning has developed rapidly in the application fields of artificial intelligence, such as computer vision, natural language processing and so on. In machine learning, the deep neural network (DNN) model is one of the most important models to promote the development of the field of artificial intelligence, but applications based on deep neural networks are difficult to deploy in key systems with extremely high safety requirements. In recent years, security issues related to deep neural networks have been continuously raised, e.g., an attacker can trick a DNN model into misclassifying a sample input by applying an intentionally chosen perturbation to a given sample. [0003] Researching the Trojan horse detect...

AI Technical Summary

Problems solved by technology

In machine learning, the deep neural network (DNN) model is one of the most important models to promote the development of artificial intelligence, but applications based on deep neural networks are difficult to deploy in critical systems with extremely high safety requirements.

Images