Generating Adversarial Text Based On Genetic Algorithm

In recent years,the rapid development of artificial intelligence technology has shown its powerful application value in many fields such as computer vision,natural language processing,and speech recognition.The security of artificial intelligence technology has also become more and more popular with these applications.The more attention is paid,the existence of adversarial samples is increasingly attracting the attention of relevant researchers.By studying the problems related to the adversarial sample,we can better think about artificial intelligence technology and understand the essence of various algorithm models in the third connection-based artificial intelligence wave represented by deep learning.By focusing on the research of sample generation and defense,to better control artificial intelligence technology,there are more possibilities to avoid some potential security problems when the future artificial intelligence technology is further deployed.Since the discovery of the adversarial sample in 2014,researchers at home and abroad have proposed a series of methods to combat sample generation and defense methods,making research on the adversarial sample problem more and more in-depth.This paper mainly studies the method of generating adversarial texts.By referring to the traditional algorithmic design algorithm to attack a well-performing convolutional neural network for sentiment analysis tasks,we can find the adversarial samples that are least likely to be perceived by humans in the shortest time.By referring to the idea of genetic algorithm,the whole iterative process of genetic algorithm,such as seed population generation,population selection,population crossover,and population variation,is targeted for the generation of adversarial samples.The design of the fitness function is based on the attack model score and the change scale score,and try to find a good adversarial sample as soon as possible.The designed attack method has been successfully implemented.The final experimental results show that the improved sample generation method based on the genetic algorithm has achieved very good results in attack success rate and disturbance ratio.The generated adversarial samples are difficult to be perceived by humans.At the same time,the designed attacks are designed.The method also achieved good results in combating the rate of sample production.