Research And Implementation Of Semi-Supervised Adversarially Robust Model-Agnostic Meta-Learning

With the development of computer technology,machine learning has achieved very excellent performance in many fields.Although new machine learning algorithms appear endlessly and the performance is gradually improved,the success of machine learning depends more on a large number of high-quality labeled data.However,in many professional fields,data or labeled data are scarce,data labeling is not only boring and time-consuming,but also requires a large number of people with professional knowledge.How to carry out in few-shot deep learning is not an easy task.The emergence of meta learning seeks a new breakthrough for the solution of the above problems.The meta model developed by meta learning,is expected to have the ability of"learning to learn".As so,it can quickly adapt to new tasks based on the learned"meta knowledge",with a small number of gradient descent steps to fine tune the model using a few labeled training data from new tasks,which is very consistent with the viewpoint of artificial intelligence.Model-agnostic meta-learning（MAML）has become one of the most successful meta learning techniques in few-shot learning.However,due to the scarcity of training sampling data,when the meta learning algorithm overtraining the existing tasks during meta-training phase,the decision boundary trained by meta learner is not accurate enough.What’s more,the unreasonable decision boundary makes the meta model more vulnerable to the adversarial perturbation,which may lead to bad robustness performance of the meta model on new tasks.This paper proposes semi-supervised adversarially robust model-agnostic meta-learning（semi-ARMAML）,and dynamic task-based tocal loss for semi-supervised adversarially robust MAML(semi-ARMAML_DFL).A semi-supervised adversarial robust regularization term and a task-agnostic regularization term based on informationentropy are brought into the objective function of semi-ARMAML to optimize the decision boundary.In particular,the calculation of the adversarial robust regularization term allows the unlabeled dataset containing the classes unseen in the labeled dataset.semi-ARMAML_DFL rewrites focus loss,uses dynamic parameters to adjust the attention of the model between difficult classification tasks vulnerable to adversarial samples and simple tasks in different stages of meta training.Task-agnostic regularization term based on informationentropy and semi supervised dynamic focus loss regularization term are also brought into the objective function to improve the robustness of the model.Through the comparative experiment with the current mainstream countermeasure meta learning algorithm on the two benchmark data sets of Mini Image Net and CIFAR-FS,it is verified that the meta model trained by semi-ARMAML can better adapt to the real application scenario and has universality,The meta model also greatly reduces the sensitivity to input disturbance and has good adversarial robustness performance.Adversarial robustness performance has significantly improved by semi-ARMAML_DFL,which lead to the highest overall robustness of the meta model.Finally,comparative experiments on miniimagenet and cifar-fs benchmark data sets show that semi-ARMAML can better adapt to real application scenarios,greatly reduce the sensitivity to input disturbances.It also has strong adversarial robustness and universality.Compared with semi-ARMAML,semi-ARMAML_DFL sacrifices clean sample robustness.However,it’s adversarial robustness performance is significantly improved,and better overall robustness is obtained.