Design And Implementation Of Safe Machine Learning Platform

With the continuous development of deep learning,deep learning has been applied in various aspects.Deep learning is also used in many environments that are quite strict about safety in today's society.Deep learning has been widely used in personal property and privacy.However,the current research found that adversarial examples can easily fool the deep learning network.Making the deep learning network produce the wrong classification results and produce the classification mode desired by the attacker.At the same time it is very difficult to distinguish against samples.The vulnerability of samples is one of the main risks of applying deep neural networks in security critical situations.Based on this background,this paper designs a security testing platform for deep learning model,which fully solves the shortage of people's knowledge on the security problem of deep learning.This paper makes a careful analysis of the entire platform and summarizes the overall architecture of the system as well as summarizes the popular deep learning model attack algorithm.This system is developed based on these algorithms.The security testing platform for deep learning model mainly includes user login and registration and deep learning attack module and defense module and evaluation module.Attack and defense modules use popular algorithms.In which,the deep learning model makes model produce correct results.The attack algorithm is used to generate adversarial examples,which is provided for users to download.The defense module uses the defense algorithm to generate the defense enhancement model so that the model can be more robust to defend against the example.The evaluation module designs several criteria to measuring the defense of the modelThe SafeML mainly adopts B/S architecture.The system mainly uses python as the main development language and the lightweight framework flask.The development of the whole system uses the MVC mode.The system's page and the logical relationship are separated.The storage of data using MySQL to store user information and return to the assessment results are stored in the database.This paper mainly completes the development and deployment of the system and tests the attack module,defense module and evaluation module after the completion of the development.Through the data obtained summarized the deficiencies of the system and carried on the outlook.