Human face forgery clue migration method based on knowledge distillation

Abstract

The invention discloses a face forgery clue migration method based on knowledge distillation. The method comprises the following steps: selecting a pre-training data set; selecting the first classification network model as a teacher model, and pre-training the teacher model by using the pre-training data set to obtain a pre-trained teacher model; selecting the second classification network model as a student model, and constructing a teacher-student joint network model by using the student model and the trained teacher model; selecting a training data set, inputting the training data set into the teacher-student joint network model, and training the student model by using the training data set to obtain a trained student model; and inputting the to-be-recognized face image into the trained student model to obtain a probability value that the to-be-recognized face image is a real image or a forged image, and further judging the authenticity of the face image. According to the invention, under the condition of avoiding loss of priori counterfeit clue knowledge, the accuracy and generalization of face counterfeit detection are improved.

Description

technical field [0001] The invention belongs to the technical field of visual identity forgery and detection, and in particular relates to a face forgery clue migration method based on knowledge distillation. Background technique [0002] In recent years, with the rapid development of artificial intelligence technology, it has been widely used in fields such as finance, medical care, urban services, industrial manufacturing, and life services. The world has entered the era of intelligence. As a popular research direction in the field of artificial intelligence, deep learning provides strong technical support for innovations in the fields of face recognition, driverless driving, natural language processing and speech recognition. However, while deep learning leads a new round of technological revolution, it also raises security issues such as visual media identity forgery, posing potential threats to personal privacy data, social stability, and national security. [0003] T...

AI Technical Summary

Problems solved by technology

Low reusability between existing forgery detection models

Although the same model can use the finetuning strategy to learn another model, different models cannot learn from each other, and can only re-use large data sets to retrain on the new network. On the one hand, it may cause data loss in the field of face forgery detection. The problem of privacy violation, on the other hand, will increase additional training costs

[0008] In addition, most of the existing technologies are based on forgery detection research based on visible light images, and rarely consider the forgery of face data in other modalities such as near-infrared mode. Considering the particularity of near-infrared face data, based on Methods such as color distortion and optical flow field difference have failed

Most of the existing technologies assume that the training and testing samples are the same forgery technology, which often causes a certain deviation and does not meet the requirements of the actual application environment.

Moreover, the method based on GAN (Generative Adversarial Network, Generative Adversarial Network) features will rely on the structure of GAN, making the feature classifier overfit on the existing generator behavior, and unable to deal with unknown generators, and the generalization ability is poor.

Images