A Defense Method for Adversarial Attacks of Deep Reinforcement Learning Models

Abstract

The invention discloses a defense method and application for deep reinforcement learning model anti-attacks, including: using a visual prediction model to predict the input environment state at the previous moment, outputting and predicting the current environment state, and obtaining the predicted current environment state in depth reinforcement Predict the environment state value in the next frame under the learning strategy; obtain the actual current environment state output by the deep reinforcement learning model, and obtain the environment state value of the actual current environment state with the added disturbance under the deep reinforcement learning strategy; use the discriminant model to predict the environment state value and the value of the environment state with disturbance added, and obtain whether the deep reinforcement learning model is attacked according to the judgment result; when the depth reinforcement learning model is attacked, extract the actual current environment state, and use two defense models to defend the actual current environment state ; the deep reinforcement learning model uses the actual current environment state after the defense to learn to predict the output.

Description

technical field [0001] The invention belongs to the field of security defense, and in particular relates to a defense method for deep reinforcement learning models against attacks. Background technique [0002] With the rapid development of artificial intelligence technology, more and more fields have begun to use AI technology. Since the concept of "artificial intelligence" was first proposed in 1956, AI has attracted more and more attention. His research areas include knowledge representation, machine perception, machine thinking, machine learning, and machine behavior, and he has made some achievements in various fields. For example, AlphaGo, an artificial intelligence Go software developed by Google DeepMind in 2014, used deep learning and reinforcement learning, and defeated Lee Sedol, one of the world's top Go players, in 2016. Reinforcement learning is also a multi-disciplinary product, which itself is a decision science, so it can be found in many branches of disci...

AI Technical Summary

Problems solved by technology

However, neural networks are extremely vulnerable to adversarial attacks. Experts and scholars have also proposed many attack methods and defense methods. However, there are no patent proposals for defense methods for deep reinforcement learning.

Images