| Artificial intelligence（AI）has achieved a great development with the improvements of machine learning and big data technology.Specially,the deep learning technology based on neural network has gradually become a representative technology promoting the development of artificial intelligence due to its high intelligence and strong generalization ability.However,in the artificial intelligence environments,the data processing mechanisms of deep learning become more and more complex,which brings serious challenges to the research for guaranteeing the data security of deep learning.Specifically,under the data inference attacks,the existing schemes of privacy protection still have problems of low efficiency,function limitation and so on.In the heterogeneous data environments,the existing research results still have problems of low efficiency and low model accuracy when dealing with low-quality data,and lack of consideration of the diversity of computing architectures,so they cannot be applied to fog-computing architectures.Under the data poisoning attacks,the existing research results do not fully consider the issue of protecting data integrity,and cannot be applied to the fog-computing architectures.For addressing the above challenges,this dissertation puts research on the data security technologies of deep learning in the artificial intelligence environments from three aspects,in terms of data inference attacks,heterogeneous data environments and data poisoning attacks.1.Research on data security technologies of deep learning under data inference attacksAn efficient and secure privacy-preserving distributed training scheme is proposed.The scheme is based on threshold encryption algorithm and threshold signature technology,which can protect users’ data privacy in the training process when multiple users collude with the server.Besides,this scheme can guarantee the authenticity of each user’s identity and data under adaptive chosen message attack,preventing malicious external adversaries from sabotaging model training by forging identities or data.In addition,the scheme can ensure the robustness of the system to unexpected user disconnection,allowing users to drop out at any stage of the training process without consuming additional resource overhead.2.Research on data security technologies of deep learning in heterogeneous data environments（1）An efficient and accurate privacy-preserving distributed training scheme under cloud-computing is proposed.This dissertation designs a new gradient aggregation method to ensure that the neural network model is trained based on high-quality data while comprehensively considering the influence of gradient component symbols and values on the correlation of different gradients.In addition,a threshold encryption technology is utilized to design a semantically secure aggregation protocol to protect users’ data privacy during the model training process.In addition,this scheme can resist the collusion between users and servers,and support some users to accidentally drop out during the training process.（2）An efficient and accurate privacy-preserving distributed training scheme under fog-computing is proposed.This scheme can mitigate the negative impact of low-quality data on model accuracy.In addition,this dissertation utilizes Shamir secret sharing,arithmetic circuits and Lagrange interpolation theorem to design a secure computing protocol that can satisfy users’ requirements of privacy protection during the distributed training process under the fog-computing with multiple serving nodes.At the same time,this scheme can resist the collusion between multiple fog nodes,and support some fog nodes to drop unexpectedly.Compared with existing schemes,this scheme is superior in terms of scalability,security and robustness.3.Research on data security technologies of deep learning under data poisoning attacksThis dissertation proposes a privacy-preserving distributed training scheme that can resist data poisoning attacks under fog-computing.To mitigate the impact of poisoned data on model accuracy,a new gradient aggregation method is proposed with the combination of the methods of Euclidean distance and hierarchical aggregation.In addition,this dissertation utilizes verifiable Shamir secret sharing,multiplication triples and some other technologies to design a privacy protection protocol under the fogcomputing structure that can resist active adversaries.While protecting users’ data privacy during the training process,this scheme can prevent fog nodes from destroying the integrity of user’s training data.In addition,this scheme can also resist the collusion between multiple fog nodes,and ensure the robustness of the system to accidental disconnection of some fog nodes.This dissertation comprehensively analyzes the securities of the above schemes and prove them under the given threat models.In addition,this dissertation demonstrates the effect and performance of the above schemes through extensive experimental results,and illustrates the performance superiority through comparison with existing schemes. |
