Research And Implementation Of Sentiment Analysis Method Based On Deep Learning In Adversarial Environment

Recently,deep learning has achieved brilliant outcomes in computer vision,speech processing,sentiment analysis and other domains due to its prominent property.A large number of deep learning applications appear in our daily lives,providing plenty of facility for enriching our lives.However,some recent research work have pointed out that deep learning models that perform well in various fields are very susceptible to attackers.When an attacker adds an extremely tiny disturbance to the input sample,it can cause the deep learning model to give wrong classification results.But so far,the research on the vulnerability of deep learning in adversarial environments has mainly focused on the field of images,and the vulnerability in the field of sentiment analysis is still unknown.In order to further study the vulnerability of deep learning models in the domain of sentiment analysis,this thesis conducted research from two aspects: adversarial attacks and adversarial defenses.In adversarial attacks,this thesis proposes an adversarial sample generation algorithm based on synonym substitution,which can generate usable and efficient adversarial samples under black box attacks.First,we calculate the word importance of the original sample by designing a new scoring function,and then use the principle of synonym substitution to modify the original sample to generate adversarial samples.We designed and implemented adversarial attacks on three deep learning models of LSTM,Bi-LSTM and CNN,and compared with the experimental results of the Random adversarial sample generation algorithm,the population-based optimization adversarial sample generation algorithm and the Deep Word Bug adversarial sample generation algorithm.The outcomes demonstrated that the adversarial sample generation approach based on synonym replacement is fabulous than the other three methods in attack success rate and word replacement rate.In addition,in adversarial defense,most of the current adversarial defense technologies are focused on image classification tasks.In the field of sentiment analysis,the defense technology against adversarial text is still at a disadvantage.Therefore,this thesis put forward an adversarial defense method based on adversarial training to improve the robustness of deep learning models.Through experimental evaluation,this method can successfully reduce the impact of adversarial attacks,and it is preliminarily verified that adversarial training is able to enhance the robustness of the deep learning model.