Explainability and sensitivity-aware federated deep learning method for 6g slice

Abstract

The application discloses a network communication technical field for 6G slice explainability and sensitivity perception federated deep learning method, through the explainable federated learning local model, the call drop rate of each radio access network (RAN) slice is predicted, meanwhile, the sensitivity perception and explainability index are taken as the constraint in the setting of such non-independent and identically distributed local data set, the reliability of the interpreter is quantitatively verified through the logit score, and the score can be taken as the constraint condition in the runtime federated learning optimization task, and the method provided by the application is obviously superior to the performance of the traditional unconstrained integral gradient post-self-organizing federated deep learning, and can effectively reduce the complexity of network resource management.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of network communication, and in particular to a federated deep learning method for explainability and sensitivity awareness of 6G slices. BACKGROUND

[0002] Currently, the popular technology in 6G network slicing is to use end-to-end (E2E) network resource autonomous management and orchestration in the network domain, because the isolation of slices can lead to high cost in efficiency. Therefore, the ETSI organization has begun to consider the standardization of zero-touch network and service management framework, zero-touch refers to the automation and management of resources without human intervention. In addition, developing a cognitive slice management solution in the 6G network is crucial for automating and managing network slices, especially across different technology domains (TDs) of network resources, and ensuring the QoE and QoS of end users.

[0003] In addition, the 6G network technology has given birth to an AI-native network slice management solution to support emerging AI services, artificial intelligence algorithms are driven by the distributed nature of data sets to gain the full potential of network slice automation to address the problem behavior of traditional cloud-centric machine learning solutions. Therefore, a decentralized learning method is needed to effectively handle distributed network slices.

[0004] 6G networks will be "machine-centric" technology, which means that all corresponding "smart things" in the 6G network will run intelligently, like a smart black box that is opaque in its actions or decision-making process, which can adversely affect the network operation of 6G technology. On this issue, XAI provides a human-interpretable method to fully explain artificial intelligence systems and their decisions in the loop to gain human trust. In view of this fact, zero-touch XAI-driven federated learning will be of particular interest due to its automation and unique advantages, which are crucial for end-user trust and security procedures.

[0005] The present application proposes a federated deep learning method for joint explainability and sensitivity awareness of 6G slices, which is significantly better than the performance of unconstrained integral gradient post-self-organizing federated deep learning, and can effectively reduce the complexity of network resource management. SUMMARY

[0006] This section is intended to summarize some aspects of the embodiments of the present application and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract and title of the specification to avoid obscuring the purpose of this section, the abstract and the title, and such simplifications or omissions cannot be used to limit the scope of the present application.

[0007] Therefore, the application aims to provide an explainability and sensitivity-aware federated deep learning method for 6G slices, predict the call drop rate of each radio access network (RAN) slice through an explainable federated learning local model, and take the sensitivity awareness and explainability index as a constraint in such a non-identically distributed local dataset setting, quantitatively verify the reliability of the interpreter through a logit score, which can be used as a constraint condition in a runtime federated learning optimization task, and the method provided by the application is obviously superior to the performance of a conventional unconstrained integral gradient post-self-organizing federated deep learning, and can effectively reduce the complexity of network resource management.

[0008] To solve the above technical problems, the application provides an explainability and sensitivity-aware federated deep learning method for 6G slices, which adopts the following technical solution: the method is implemented through an application framework, and the application framework comprises

[0009] a radio access network composed of K base stations, wherein each base station is deployed with N parallel slices, each base station runs a local control closed loop, the local control closed loop collects monitoring data and performs call drop rate prediction, and the local control closed loop trains a local federated learning local model independently;

[0010] the method is iteratively run in a closed loop manner through the local control closed loop, and is provided with a runtime interpreter, a model tester and a logit mapper,

[0011] for each local control closed loop, the local federated learning local model feeds a model graph to the model tester, the model tester tests features and corresponding predicted features to the interpreter, the interpreter first generates feature attributes using an integrated gradient XAI method; then the logit mapper uses these attributes to select top features, and then calculates a logit score , which is fed back to the local federated learning local model to be included in the local constraint condition optimization of the local federated learning local model;

[0012] the application framework further comprises a recall mapper, which calculates a recall score from the predicted value and the true value output by the model tester and the interpreter, and includes the recall score in the local constraint condition optimization of the local federated learning local model;

[0013] for each local control closed loop, the predicted call drop rate is classified , the long-term statistical constraint defined on the joint local dataset sample, the corresponding recall score and the logit score Under the constraints, the associated true value should be Minimize the main loss function.

[0014] Optionally, the monitoring data collected by the local control closed loop is used to construct a local dataset for slice n; n=1, ..., N; the local dataset includes: average physical resource block, average data transmission delay, and SNR value characterizing channel quality, expressed by the formula:

[0015]

[0016] in, Indicates the first in the local dataset Each input feature, and Indicates the corresponding first There are 1 output feature; k=1, ..., K; the output feature includes: call drop rate.

[0017] Optionally, the application framework also includes a federated learning layer, which is used to collect local federated learning local model information trained independently by each local control closed loop, aggregate the local federated learning local model information trained independently by each local control closed loop through an algorithm, and then feed the obtained global model information back to the aggregation cloud server for resource management optimization of wireless access network.

[0018] Optionally, in the t-th round of iterative operation in the closed-loop mode, the locally optimized weights in the federated learning framework are... The data is sent to the aggregation cloud server, which then generates a global federated learning model for slice n. This can be expressed as a formula:

[0019]

[0020] in Let represent the weight factor of the nth slice in the kth base station in the tth round, which is used in subsequent constraint optimization to minimize the main loss function, thereby constructing the implementation framework of federated learning. , It is the total data sample of all datasets related to slice n. Then, the aggregation cloud server broadcasts the global federated learning model to all local K local control loops that use it to start the next round of local iterative optimization.

[0021] Optionally, the local federated learning local model information includes local federated learning local model parameters or local federated learning local model update parameters. After receiving the updated model graph, the model tester reconstructs the test prediction from a subset of data extracted from the local dataset. All model diagrams and test features and corresponding predictive features All of these will be provided to the interpreter.

[0022] Optionally, the formula for calculating the logarithmic odds score is:

[0023]

[0024] in, It is a predictive feature. It is the first in the local dataset Each input feature Represents the first part of the modified local dataset Features that are zero-padding.

[0025] Optionally, the recall score The calculation formula is:

[0026]

[0027] in, definition The proportion of positive values ​​in the classification It satisfies the expression of A subset of.

[0028] Optionally, a lower limit was imposed on the recall score. An upper limit was set for the logarithmic probability score. The constraint optimization is transformed into optimization within rounds specified by the iteration cycle and in the federated learning process. Solving constrained local classification problems, i.e.

[0029] ,

[0030] ,

[0031] .

[0032] Optionally, two Lagrange equations can be constructed to solve the above-mentioned constrained local classification problem, as shown in the following formula:

[0033] ,

[0034] ,

[0035] in, and Indicates the original constraint. and Indicates smooth substitution, and Represents the Lagrange multiplier.

[0036] wherein the smooth surrogate formula is as follows:

[0037] ,

[0038] .

[0039] Optionally, the constraint optimization is a non-zero-sum two-player game strategy, and the local constraint optimization is calculated by the following formula:

[0040] ,

[0041] .

[0042] In summary, the present application includes at least one of the following beneficial effects: predicting the call drop rate of each radio access network (RAN) slice through an interpretable federated learning local model, while taking sensitivity awareness and interpretability indicators as constraints in such non-identically distributed local dataset settings, quantitatively verifying the credibility of the interpreter through a logit score, which can be used as a constraint condition in the runtime federated learning optimization task, the method provided by the present application is significantly better than the performance of the traditional unconstrained integral gradient post-self-organizing federated deep learning, and can effectively reduce the complexity of network resource management. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0044] Figure 1 Flow chart of the federated deep learning method for 6G slice interpretability and sensitivity awareness of the present application;

[0045] Figure 2 The federated deep learning method for 6G slice interpretability and sensitivity awareness of the present application is iteratively run in a closed-loop manner through local control. DETAILED DESCRIPTION

[0046] The technical solutions of the embodiments of the present application will be described below in conjunction with the drawings of the embodiments of the present application; it is obvious that the described embodiments are only some embodiments of the present application, not all embodiments; based on the embodiments in the present application; all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0047] In the description of the present application, it should be noted that the terms "upper", "lower", "inner", "outer", "top / bottom end" and the like indicate the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first", "second" are only for the purpose of description, and cannot be understood as indicating or implying relative importance.

[0048] In the description of the present application, it should be noted that unless otherwise explicitly specified and limited, the terms "mounting", "provided with", "sleeved / connected", "connected" and the like should be broadly understood, for example, "connected" can be fixedly connected, or can be detachably connected, or integrally connected; can be mechanically connected, or can be electrically connected; can be directly connected, or can be indirectly connected through an intermediate medium, or can be internal communication of two elements. For those of ordinary skill in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0049] Embodiment one

[0050] The application discloses a federated deep learning method for explainability and sensitivity awareness of 6G slices, which is mainly realized by applying a framework, as shown in the figure. Figure 1 The application framework includes

[0051] A wireless access network, a radio access network (RAN) is composed of K base stations (BS), each of which deploys a set of N parallel slices, each base station runs a local control closed loop that collects monitoring data and performs call drop rate prediction, and the monitoring data collected by the local control closed loop is used to construct a local data set for slice n (n = 1,..., N), that is,

[0052]

[0053] Wherein, represents the first input feature in the local data set, and represents the corresponding first output feature, base station k (k = 1,..., K), the local data set includes: average physical resource block, average data transmission delay, SNR value representing channel quality, etc., and the output features include: call drop rate, etc.

[0054] In detail, in the embodiment, the application framework further comprises a federated learning layer, which is configured to collect local federated learning local model information trained by each local control closed loop, aggregate the local federated learning local model information trained by each local control closed loop through an algorithm, and then feed back the obtained global model information to the aggregation cloud server for resource management optimization of the wireless access network.

[0055] In the field of communication anomaly detection, to train an accurate classifier, ideally, a very comprehensive data set covering all possible situations is needed. However, due to different traffic distributions caused by heterogeneous user distribution and channel conditions, the actual cumulative data set is a non-identically distributed data set, and it is difficult to include all possible abnormal situations in one data set. In the method provided by the present application, the model is trained using data sources (local control closed loop) from multiple slices without directly sharing data. Since each local control closed loop may have some unique data, which contains abnormal situations or features that other data sets do not cover, the end-to-end slice federated learning layer is a key part in the federated learning application framework. It manages and coordinates the federated learning process in a slice manner from the end-to-end perspective (from the source of data generation to the final model aggregation application), collects model parameters or model updates trained by each local closed loop on its own data, and then aggregates the model information from different local closed loops through a specific algorithm (such as weighted average), and then feeds back the obtained global model information to the aggregation cloud server for resource management optimization of the wireless access network (RAN).

[0056] The federated deep learning method for 6G slicing provided by the present application is interpretable and sensitive, in which the constrained call drop detection classifier and the interpreter exchange the feature attributes and prediction values in a closed loop manner to realize transparent zero-touch service management of 6G network slicing of the wireless access network (RAN) under the condition of non-identically distributed data sets. The integrated gradient explainable artificial intelligence (XAI) method is used to show the feature attributes of the wireless access network (RAN) slicing, the generated attributes are used to quantitatively verify the credibility of the explanation through the logit index, and the index is used as a constraint condition in the optimization task of federated learning. Further, the present application formulates a corresponding federated learning optimization problem with joint recall and logit constraints under the agent Lagrange framework, and solves the optimization through a non-zero-sum double-player game strategy. The specific processing flow is as shown in Figure 2 .

[0057] The local control closed loop trains the local federated learning local model independently; the method iteratively runs in a closed loop manner through the local control closed loop, with a runtime interpreter, a model tester and a logit mapper, for each local control closed loop, the local federated learning local model feeds the model graph to the model tester, the model tester tests the features and the corresponding predicted features to the interpreter, which first uses the integrated gradient XAI method to generate feature attributes; then the logit mapper uses these attributes to select the top features, and then calculates the logit score, which is fed back to the local federated learning local model for inclusion in the local constraint optimization of the local federated learning local model.

[0058] The application framework also includes a recall mapper that calculates a recall score from the predicted values and true values output by the model tester and interpreter, which is included in the local constraint optimization of the local federated learning local model; for each local control closed loop, the predicted call drop rate is classified , the joint local data set samples defined by the long-term statistical constraints and the corresponding recall score and logit score constraints should minimize the main loss function associated with the true value .

[0059] In detail, in the tthround of iterative running in a closed loop manner, the local closed loop optimized weights in the federated learning framework are sent to the aggregation cloud server, which generates a global federated learning model for slice n ,

[0060]

[0061] where is the weight factor of the nthslice in the kthbase station in the tthround, used in subsequent constraint optimization to minimize the main loss function, thereby constructing the implementation framework of federated learning , is the total data sample of all data sets related to slice n, then the aggregation cloud server broadcasts the global federated learning model to all local K local control closed loops that use it to start the next round of local iterative optimization.

[0062] In detail, in the present embodiment, the local federated learning local model information comprises local federated learning local model parameters or local federated learning local model update parameters, and the model tester reconstructs the test prediction from the data subset extracted from the local data set after receiving the updated model graph , all model graphs, test features and corresponding predicted features will be provided to the interpreter.

[0063] In detail, in the present embodiment, the interpreter exhibits feature properties of the radio access network (RAN) slice by utilizing a low-complexity integrated gradients explainable artificial intelligence (XAI) method, which utilizes the properties generated by integrated gradients to explain the prediction results of the model. If the model predicts that a sample belongs to a certain class, the interpreter can find out the features that contribute most to the prediction result by analyzing the values, thereby providing intuitive explanations to the user about the features on which the model makes decisions. The interpreter can explain the performance differences of the model on different slices by analyzing the distribution of integrated gradients on different slice data. If the performance of the model is poor on a certain slice, the interpreter can check the feature importance distribution of the slice data to see whether there are some key features whose integrated gradients are abnormal or some features have large differences in importance on different slices, so as to find out the reasons that may cause the performance difference and take corresponding measures to improve.

[0064] In detail, in the present embodiment, in order to represent the reliability of the local model, the logit score is preferably calculated, which measures the influence of the top attribute features on the model prediction. Specifically, the logit score is defined as the average difference in the negative log probability of the predicted class before and after masking the features with zeros. The logit mapper first selects the top features according to the attributes collected from the interpreter, and replaces them with zero padding. That is,

[0065]

[0066] wherein, is the predicted feature, is the i-th input feature in the local data set, indicates that the first features in the modified local data set are zero padding features. Finally, the logit mapper calculates the logit score, which is used as one of the constraints for the FL optimization task.

[0067] ​​In detail, in the present embodiment, the sensitivity-aware score (recall mapping) is preferably adopted to call the recall rate as the score of the sensitivity of the local classifier of the federated learning, denoted as i.e.

[0068]

[0069] wherein, defined the proportion of the classification as positive values, is a subset satisfying the expression

[0070] In detail, in the present embodiment, in order to make the abnormal detection / classification of the dropped call more credible, a protocol of federated learning of artificial intelligence service quality is established between the slice tenant and the infrastructure provider, wherein a lower limit is imposed on the recall score , and an upper limit is set for the logit score The constraint condition is optimized to solve the constrained local classification problem in the iteration cycle specified by the iteration cycle and in the round of federated learning , that is

[0071] ,

[0072] ,

[0073] .

[0074] In detail, in the present embodiment, the above-mentioned constrained local classification problem can be solved by the Lagrange framework, and two Lagrange equations need to be further constructed, as follows:

[0075] ,

[0076] ,

[0077] wherein, and represent the original constraint, and represent the smooth replacement, and represent the Lagrange multiplier,

[0078] wherein, the smooth replacement formula is as follows:

[0079] ,

[0080] Because the negative logarithm is already a convex function, this also proves that the solution of the optimization problem is equivalent to the solution obtained when only the original constraint is used.​​

[0081] In detail, in the present embodiment, the constraint optimization is a non-zero-sum two-player game strategy, and the local constraint optimization is calculated by the following formula:

[0082] ,

[0083] .

[0084] The above are preferred embodiments of the present application, and are not intended to limit the protection scope of the present application, and thus: any equivalent changes made according to the structure, shape, principle of the present application should be covered within the protection scope of the present application.

Claims

1. A federated deep learning method for explainability and sensitivity awareness of 6G slices, characterized in that: The method is realized by applying a framework, the framework comprises A wireless access network composed of K base stations, wherein each base station deploys N parallel slices, each base station runs a local control closed loop, the local control closed loop collects monitoring data and performs call drop rate prediction, wherein the local control closed loop trains a local federated learning local model independently; The method is iteratively run in a closed loop manner by the local control closed loop, with a runtime interpreter, a model tester and a logit mapper, For each local control loop, the local federated learning local model feeds the model graph to a model tester, which tests features and corresponding predicted features to an explainer, which first generates feature properties using the integrated gradients XAI method; then uses these properties to select top features for a logistic map which is fed back to the local federated learning local model to be included in the local constraint optimization of the local federated learning local model; The application framework further includes a recall mapper that calculates a recall score from the predicted values and true values output by the model tester and the explainer into the local constraint optimization of the local federated learning local model; For each local control loop, the predicted call drop rate is classified in conjunction with the local dataset defined long-term statistical constraints on the sample and corresponding recall scores and logit scores The real value associated with it should minimize the main loss function under the constraint conditions.

2. The federated deep learning method for 6G slice explainability and sensitivity awareness according to claim 1, characterized in that: The monitoring data collected by the local control closed loop is used to construct a local data set for slice n; n = 1,..., N; The local data set comprises: average physical resource block, average data transmission delay, SNR value representing channel quality, which is expressed by the formula: ， in, Indicates the first in the local dataset Each input feature, and Indicates the corresponding first There are 1 output feature; k=1, ..., K; the output feature includes: call drop rate.

3. The interpretable and sensitivity-aware federated deep learning method for 6G slices of claim 1, wherein: The framework further comprises A federated learning layer for collecting local federated learning local model information trained by each local control closed loop independently, aggregating the local federated learning local model information trained by each local control closed loop independently through an algorithm, and then feeding back the obtained global model information to the aggregation cloud server for resource management optimization of the wireless access network.

4. The interpretable and sensitivity-aware federated deep learning method for 6G slices of claim 3, wherein: In the t-th round of iterative running in a closed loop mode, the local closed loop optimized weight in the federated learning framework is sent to the aggregation cloud server, and the aggregation cloud server generates a global federated learning model for slice n , which is expressed by the formula: ， wherein represents the weight factor of the nth slice in the kth base station in the tth round, which is used in the subsequent constraint optimization to minimize the main loss function, thereby constructing the implementation framework of federated learning, , is the total data sample of all data sets related to slice n, and then the aggregation cloud server broadcasts the global federated learning model to all local K local control closed loops using it to start the next round of local iterative optimization.

5. The interpretable and sensitivity-aware federated deep learning method for 6G slices of claim 3, wherein: The local federated learning local model information includes local federated learning local model parameters or local federated learning local model update parameters, and the model tester, after receiving the updated model graph, reconstructs the test prediction from the data subset extracted from the local data set , all model graphs, test features and corresponding predicted features will be provided to the interpreter.

6. The interpretable and sensitivity-aware federated deep learning method for 6G slices of claim 1, wherein: The calculation formula of the logit score is: ， in, It is a predictive feature. It is the first in the local dataset Each input feature Represents the first part of the modified local dataset Features that are zero-padding.

7. The interpretable and sensitivity-aware federated deep learning method for 6G slices of claim 1, wherein: The recall score The formula for calculating the recall score is: ， wherein Definitions the proportion classified as positive, is a subset of the set of values satisfying the expression .

8. The interpretable and sensitivity-aware federated deep learning method for 6G slices of claim 4, wherein: A lower limit was imposed on the recall score. An upper limit was set for the logarithmic probability score. The constraint optimization is transformed into optimization within rounds specified by the iteration cycle and in the federated learning process. Solving constrained local classification problems, i.e. ， ， 。 9. The interpretable and sensitivity-aware federated deep learning method for 6G slices according to claim 8, characterized in that: Two Lagrange equation frameworks are constructed to solve the above constrained local classification problem, and the formula is as follows: ， ， wherein, and denotes the original constraint, and denotes the smooth substitute, and denotes the Lagrange multiplier, The smoothing substitution formula is as follows: ， 。 10. The interpretable and sensitivity-aware federated deep learning method for 6G slices according to claim 9, characterized in that: The constraint condition optimization is a non-zero-sum two-player game strategy, and the local constraint condition optimization is calculated by the following formula: ， 。

Images