A Trust Quantization Model For Access Control Of Medical Big Data

With the rapid development of data technology,data resources have become an important driving force for economic transformation and development.With the popularisation of smart mobile terminals and the digitisation of medical management in recent years,healthcare data has also grown explosively,and we call the current aggregation of all healthcare data medical big data.Seeing the strategic value of developing big data in healthcare,China is stepping up its efforts to call on all units to actively build big data in healthcare.In a massive medical data environment,it is an extremely powerful aid to both diagnosis and other medical research.But at the same time,big data contains a lot of personal information of patients and practitioners,which is vulnerable to external attacks or leaked by practitioners for profit,making hospitals reluctant to share data from their information systems.Access control is an important tool for data security and is uniquely suited to handling reasonable user access to appropriate shared resources.However,traditional access control methods are clearly unable to support such large volumes of data due to their static approach to permission allocation,so there is an urgent need for dynamic access control methods that are suitable for healthcare big data.One of the more promising approaches is trust-based access control,but the current trust-based access control methods are not well integrated with medical big data in terms of quantification of trust and insufficient granularity of permission assignment,so this paper focuses on exploring trust quantification methods suitable for trust-based access control in the current medical big data environment,and reflects the following two points.To construct a comprehensive trust quantification model based on doctors’ historical behaviours.The aim is to learn the low-dimensional representation of doctors’ historical behaviours through deep neural networks and obtain their comprehensive trust values by comparing the relationships of low-dimensional representations of doctors within the whole department.The model first n-hot codes the information contained in the case,then aggregates each feature according to the doctor’s history of visits by moving exponential averaging from old to new,and uses entropy weighting to obtain the weights of the features of interest to the doctor,and multiplies the two by element to obtain the doctor’s history of visits feature.The doctor’s visit process is then transformed into a doctor learning process like another doctor according to a certain strategy,thus constructing a doctor-doctor directed interaction topology map.Finally,a multilayer graph convolutional neural network(GCN)is used to affine transform the doctor features to predict the department to which the doctor belongs,and the last layer of hidden layer features is used as a low-dimensional representation of each doctor,and the cosine distance between the doctor and the department centre representation is used to represent the doctor’s combined trust value.Finally,the model is shown to be more effective in discriminating doctors with malicious access behaviour by simulating experiments against existing models.To construct a quantitative model for the trustworthiness of a doctor’s current visit,the aim is to obtain a low-dimensional representation of the case and the doctor’s current work goal through deep neural networks that learn the connection between case data and the connection between the visit history and the current visit behaviour,and to obtain the trustworthiness of the current visit through the similarity between the work goal and the case representation of the next visit.The model starts with a case relationship,and the case-case interaction matrix can be very large,so the aggregation of case features using random wandering is performed so that the aggregated case representation has both its own and neighbouring nodes’ features and topological relationships.The doctor’s visit record is then fed through the GRU network,with the output of the final hidden layer as a representation of the doctor’s current work status.A simple normalisation of the representation vector followed by multiplication gives the confidence level of the current visit behaviour.Due to the lack of similar related studies,this model focuses on exploring the effect of various hyperparameters on the model.