1 Summary
At present, the most effective way for deep models to resist adversarial attacks is adversarial training, where neural networks are trained by introducing adversarial samples to make the model robust to a certain extent. The current research direction of adversarial training mainly focuses on the training method in the multi-classification task, this paper tries to train the multi-classifier with the help of multi-labeled classifiers for adversarial training, in which the difference between multi-classification task and multi-labeled task can be learned from the article "TheMulti-label classifier and pytorch source code for python deep learningThe "Knowing.
As an example, a face picture will show a lot of labeling information, such as there will be name, gender, age, emotion and other labeling information, according to the idea of this paper is to first train a multi-label classifier about gender, age, and use this multi-label classifier to train against the multi-classifier about name. With this idea to start the theoretical analysis and experimental verification of this article.
2 Introduction to the methodology
2.1 Adversarial Samples for Multi-Categorization Tasks
2.2 Multi-label task adversarial samples
2.3 Dual Classifier Adversarial Training
According to the above schematic of the principle, the dual classifiers against training can be categorized into two ways as follows
Face Expression Confrontation Training
In this paper, the AU unit detection model (multi-label classifier) is utilized to assist the face expression recognition model (multi-classifier) for adversarial training.
The relevant description of each AU unit in the face is shown in the following figure
The basic facial expressions are Angry, Disgust, Fear, Happy, Sad, Surprise and Comtempt.
The relationship between the face base expression and the AU unit is shown schematically and the corresponding calculation formula is as follows
Above is the details of python deep learning adversarial training with the help of multi-label classifiers, for more information about multi-label classifiers adversarial training, please pay attention to my other related articles!