Method and system for secure application of machine learning model

Abstract

The invention discloses a safe application method and system of a machine learning model. The method includes steps such as front-end service model processing, prior information extraction, verification strategy analysis, back-end verification module processing, and fusion decision-making. The back-end verification and the front-end service machine learning service model of the present invention are relatively independent, and the anti-attack ability can be enhanced through the separation of module functions, and the application security of the machine learning model can be improved; the back-end verification supports the multi-dimensional back-end of inherent attributes and external information Verification, the verification method is more comprehensive and stricter, and is suitable for the definition of verification strategies for various types of machine learning tasks. For the processing results of the front-end service machine learning model, the influence coefficient of the main feature can be defined, and the verification strategy can be flexibly formulated according to the influence coefficient. It is suitable for application scenarios of various machine learning models, and has good flexibility and versatility.

Description

technical field [0001] The invention relates to the field of computer technology, in particular to a method and system for safely applying a machine learning model. Background technique [0002] Machine learning, especially deep learning, is becoming a powerful tool for building intelligent systems, which can greatly improve the efficiency and quality of production and life. Machine learning is mainly based on the statistical analysis of data, which can automatically extract the hidden features of data from massive input data and form the decision boundary of data processing. [0003] However, the data in many application scenarios has the characteristics of high dimensionality and large random fluctuations. For example, for images with a resolution of 1080P, it is difficult to cover all sample spaces by manually collecting and labeling training data sets. At the same time, the statistics of the data itself The distribution will also be affected by the statistical distribut...

AI Technical Summary

Problems solved by technology

[0005] Furthermore, when training the model with data, machine learning often expresses the target as a whole with a unified parameter model. Due to the incompleteness of the training data, the model cannot express all possible data. At the same time, this method does not have local Or global reasoning capabilities, resulting in performance loss of machine learning models and increasing the risk of attacks

Images