Reinforcement learning-oriented data exception detection defense method

Abstract

The invention discloses an abnormal data detection defense method for reinforcement learning, and the method comprises the steps: (1) building a car automatic driving environment, carrying out the reinforcement learning through employing a deep deterministic strategy gradient algorithm based on the state data provided according to the car automatic driving environment, and generating driving statedata as a training sample; (2) training a generative adversarial network composed of a generator and a discriminator by using the training sample; (3) according to the collected driving state data, generating predicted driving state data at the next moment according to the driving state data at the current moment by utilizing a trained generator; and (4) judging whether the real driving state data and the predicted driving state data at the next moment are normal or not by utilizing the trained discriminator, and when the real driving state data at the next moment are abnormal, judging that the predicted driving state data are normal, and replacing the real driving state data by utilizing the predicted driving state data.

Description

technical field [0001] The invention belongs to the reinforcement learning-oriented defense field, and in particular relates to a reinforcement learning-oriented abnormal data detection and defense method. Background technique [0002] Reinforcement learning is one of the directions of artificial intelligence that has attracted much attention in recent years. Its basic idea is to learn the optimal strategy to achieve the purpose of learning by maximizing the cumulative rewards that the agent can obtain from the environment. However, the training process of reinforcement learning is vulnerable to attack, which makes its training set data abnormal, which affects the decision-making judgment or action selection of the agent in the learning process, and finally makes the agent learn actions in the direction of failure, which is very important for reinforcement learning. The domain of decision-making security applications is a major challenge. [0003] At present, according to t...

AI Technical Summary

Problems solved by technology

However, this method generally faces two problems: one is the need to define a clear decision boundary to define normal points and abnormal points; the other is the high-frequency computing performance bottleneck between the disaster of dimensionality and cross-index calculations

Existing studies have shown that the decision poisoning attack can change the decision by changing the reward in the training set, so that the policy in the training set is poisoned, the action selection is wrong, and finally the agent cannot achieve the learning goal.

Images