Artificial Intelligence:Chaotic Analysis And Research

In today’s rapid development of artificial intelligence,face recognition,intelligent transportation and smart home applications are inseparable from the support of artificial intelligence algorithms,so the safety of artificial intelligence algorithms is particularly important.To address this problem,chaos theory provides a new perspective to study the "butterfly effect" of artificial intelligence algorithms from the perspective of chaos theory for small perturbations in artificial intelligence algorithms.With this as a starting point,this paper investigates the chaotic nature analysis of artificial intelligence algorithms and the application of chaos theory to adversarial samples,and the main work and innovations of this paper are as follows:Firstly,the research of chaos detection algorithm based on 0-1 test method and its improvement.Due to the similarity between chaos and noise,the existing chaos detection methods cannot guarantee the accuracy of chaos detection when the signal-to-noise ratio is low.To address this problem,this paper proposes a chaos detection tree method based on fully adaptive noise ensemble empirical modal decomposition and local projection wavelet noise reduction,and verifies the effectiveness of the chaos detection algorithm by multi-system simulation.Experiments show that the chaos detection accuracy of this method is improved by 8.16% when the signal-to-noise ratio is 5-25 d B compared with the existing methods.Secondly,we analyze the chaos of artificial intelligence algorithms based on the chaos detection tree method."In order to study whether the AI algorithm itself has chaotic characteristics,this paper proposes to classify the results of AI algorithms based on the theory of complex adaptive systems and analyze the possible chaotic AI algorithms by chaos detection tree method.The experiments show that the algorithms of TCNN family,LSTM family and GRU family have chaos,and the algorithm chaos is closely related to the model parameters,data properties,training process and other factors.Finally,the experiments summarize the general model for the appearance of chaos in the algorithms.Thirdly,a generic adversarial sample generation algorithm based on Sinusoidal chaotic mapping.Few of the existing adversarial sample generation algorithms have been constructed using chaotic properties.To address this problem,this paper innovatively proposes adding Sinusoidal mapping to the adversarial sample generation algorithm,generating a learning rate and two-dimensional input chaotic perturbation based on Sinusoidal mapping,and finally proposing the algorithm combined with the principle of generic adversarial sample generation.Experiments show that the method in this paper has 81.2% average attack success rate with target and 86.95% average attack success rate without target for a given task.Compared with existing methods,the proposed method has higher attack success rate and better migration,and successfully attacks the commercially available Yu Pan face recognition access control system under the black box model.