Optimization method and system for network task response information

Abstract

The present disclosure provides an optimization method for network task response information, comprising: determining at least one verification problem regarding the response logic based on the response logic of a machine learning model for network tasks, wherein the network tasks include network congestion control tasks, video stream rate control tasks, and data center stream scheduling tasks, and the response logic represents the principle by which the machine learning model generates response information; encoding the verification problem to convert the verification problem into a first satisfiability modulo theory formula; encoding the network structure of the machine learning model by an encoder to obtain multiple second satisfiability modulo theory formulas; calling a planner to analyze the first satisfiability modulo theory formula and the second satisfiability modulo theory formula to generate an explanatory conclusion for representing the rationality of the response logic of the machine learning model. The present disclosure also provides an optimization system for network task response information.

Description

Technical Field

[0001] The present disclosure relates to the technical field of network security, and in particular to a method and system for optimizing network task response information. Background Art

[0002] The rapid development of mobile communications and the internet has led to a surge in the number of applications and users, significantly increasing the difficulty faced by network designers in solving problems such as network resource scheduling and performance optimization. These problems cannot be solved manually. Simultaneously, with the widespread application of machine learning in fields such as computer vision and natural language processing, machine learning has also been gradually applied to computer networks, solving numerous classic problems in the field, such as congestion control, video streaming rates, and data center flow scheduling. However, due to the inherent black-box nature of machine learning systems, we cannot fully understand the model's predictions, making it difficult to judge the rationality of the model's decisions. If a machine learning network system makes an incorrect decision in real-world applications, it can lead to severe economic and social losses, hindering the large-scale deployment of related network systems and further improving model design. Therefore, studying the decision-making principles of network machine learning models is crucial.

[0003] To understand the decision-making capabilities and underlying principles of network machine learning systems, relevant techniques focus on verifying their behavior and explaining their decisions, respectively. While both are crucial for understanding the behavior of black-box machine learning systems, they are often viewed as separate and orthogonal topics. Verification studies whether the input-output mapping of a learning-based network system consistently satisfies certain properties. These properties typically relate to model quality: robustness to adversarial perturbations, missing features, and extreme values. Unlike verification, explaining network machine learning systems typically aims to provide users with a set of simple, deterministic rules to explain the system's behavior or to analyze the system's input features to provide explanations. Existing methods for finding rules primarily include anchors, counterfactual explanations, and decision rules. Methods for analyzing features primarily include feature importance and sensitivity analysis.

[0004] However, the aforementioned methods offer single solutions for different attributes and interpretability problems. Each work proposes a different approach for a specific attribute or interpretability problem. Different interpretability tools also vary in form and complexity. Related technologies fail to systematically investigate the relationships between different attribute verifications and various interpretability problems, resulting in a lack of general tools that can verify and explain the broad behavior of network machine learning systems. Summary of the Invention

[0005] In order to solve at least one of the above problems, the present disclosure provides a method and system for optimizing network task response information.

[0006] According to one aspect of the present disclosure, a method for optimizing response information of a network task is provided, comprising: determining at least one verification problem regarding the response logic of a machine learning model for a network task, wherein the network task comprises a network congestion control task, a video stream rate control task, and a data center stream scheduling task, and the response logic represents the principle by which the machine learning model generates response information; encoding the verification problem to convert the verification problem into a first satisfiability modulo theory formula; encoding the network structure of the machine learning model by an encoder to obtain a plurality of second satisfiability modulo theory formulas, comprising: encoding the input layer according to the input variable range set by the input layer of the machine learning model for the network task, to obtain the second satisfiability modulo theory formula corresponding to the input layer; according to the construction logic of the machine learning model, construct the relationship expression between the input variables of the machine learning model regarding the network task and the original output variables generated in response to the network task, and encode the relationship expression to obtain the second satisfiability modulo theory formula corresponding to the hidden layer; and according to the original output variable range of the network task set by the output layer of the machine learning model, encode the output layer to obtain the second satisfiability modulo theory formula corresponding to the output layer; call the planner to analyze the first satisfiability modulo theory formula and the second satisfiability modulo theory formula, and generate an explanatory conclusion for characterizing the rationality of the response logic of the machine learning model.

[0007] In some embodiments, the method further includes: constructing a model robustness formula for the relationship between the input variables corresponding to the network task and the original output variables generated in response to the network task based on the various attributes of the machine learning model and the second satisfiability module theory formula.

[0008] In some embodiments, the call planner analyzes the first satisfiability modulo theory formula and the second satisfiability modulo theory formula to generate an explanatory conclusion for characterizing the rationality of the response logic of the machine learning model, including: perturbing the input variables regarding the network task to obtain perturbed input variables; inputting the perturbed input variables into the model robustness formula to obtain perturbed output variables, and determining whether the perturbed output variables are consistent with the original output variables; when the perturbed output variables are consistent with the original output variables, perturbing the input variables until the generated perturbation input variables make the perturbation output variables inconsistent with the original output variables, and determining the perturbation range of the machine learning model, wherein the perturbation range is used to characterize the robustness of the machine learning model.

[0009] In some embodiments, the call planner analyzes the first satisfiability modulo theory formula and the second satisfiability modulo theory formula to generate an explanatory conclusion for characterizing the rationality of the response logic of the machine learning model, including: deleting any feature of the input variables related to the network task, and inputting the input variables missing the feature into the machine learning model to determine whether the actual output variable of the machine learning model is consistent with the original output variable; and summarizing the features deleted when the actual output variable is consistent with the original output variable to obtain a set of loseable features that can generate the original output variable.

[0010] In some embodiments, the call planner analyzes the first satisfiability modulo theory formula and the second satisfiability modulo theory formula to generate an explanatory conclusion for characterizing the rationality of the response logic of the machine learning model, including: taking extreme values ​​for any feature of the input variables related to the network task, and inputting the input variables with the extreme values ​​into the machine learning model to determine whether the actual output variables of the machine learning model are consistent with the original output variables; summarizing the features that take extreme values ​​when the actual output variables are consistent with the original output variables to obtain a set of features that can take extreme values.

[0011] In some embodiments, the method further includes constructing a model interpretability problem formula based on the second satisfiability modulo theory formula, the predicate logic of the perturbed features, and the predicate logic of the unperturbed features in the input variables.

[0012] In some embodiments, the call planner analyzes the first satisfiability modulo theory formula and the second satisfiability modulo theory formula to generate an explanatory conclusion for characterizing the rationality of the response logic of the machine learning model, including: according to the model interpretability problem formula, using a qualitative anchor point method to determine the dominant input feature set that plays a dominant role in the machine learning model in responding to the network task; and using a quantitative anchor point method to determine the value range of each dominant feature in the dominant input feature set; according to the model interpretability problem formula, using a counterfactual explanation algorithm to determine the expected input variables of the machine learning model when responding to the network task, so that the machine learning model generates expected output variables that meet the expected response decision when responding to the network task based on the expected input variables; according to the model interpretability problem formula, using a decision boundary algorithm to determine the machine learning model when issuing the network task. The invention relates to a method for determining a boundary range of input variables for performing network tasks, so that when the input variables of the machine learning model are within the boundary range of the input variables, the output variables generated by the processing of the network tasks are inconsistent with the original output variables; according to the formula of the model interpretability problem, adjusting each feature of the input variables for the network tasks one by one, respectively determining the minimum attack degree of each feature when causing the original output variable to change, and ranking the importance of each feature of the input variables according to the minimum attack degree, wherein the minimum attack degree of a feature is inversely proportional to the importance of the feature; and according to the formula of the model interpretability problem, performing perturbation processing on any feature of the input variables for the network tasks, obtaining an output variable corresponding to the feature when the feature is used as the perturbation feature, and determining the degree of deviation between the output variable and the original output variable, wherein the degree of deviation is used to characterize the sensitivity of the machine learning model to changes in the feature.

[0013] In some embodiments, after the planner is called to analyze the first satisfiability modulo theory formula and the second satisfiability modulo theory formula to generate an explanatory conclusion for characterizing the rationality of the response logic of the machine learning model, it includes: calling the satisfiability modulo theory engine to analyze the verification problem in the first satisfiability modulo theory formula, and providing feedback on the explanatory conclusion about the verification problem through the planner.

[0014] After the call planner analyzes the first satisfiability modulo theory formula and the second satisfiability modulo theory formula to generate an explanatory conclusion for characterizing the rationality of the response logic of the machine learning model, it includes: optimizing the machine learning model according to the explanatory conclusion, and adjusting the input variables corresponding to the network task so that the response information generated by the machine learning model processing the network task is consistent with the expected response result.

[0015] According to another aspect of the present disclosure, there is provided an optimization system for network task response information, comprising: a problem determination module, for determining at least one verification problem about the response logic of the network task according to the machine learning model, wherein the network task includes a network congestion control task, a video stream rate control task, and a data center stream scheduling task, and the response logic characterizes the principle of the machine learning model generating response information; a problem encoding module, for encoding the verification problem to convert the verification problem into a first satisfiability modulo theory formula; a model encoding module, for encoding the network structure of the machine learning model by an encoder to obtain multiple second satisfiability modulo theory formulas, comprising: encoding the input variable range of the network task according to the input layer of the machine learning model, The input layer is encoded to obtain a second satisfiability modulo theory formula corresponding to the input layer; according to the construction logic of the machine learning model, a relationship expression between the input variables of the machine learning model regarding the network task and the original output variables generated in response to the network task is constructed, and the relationship expression is encoded to obtain a second satisfiability modulo theory formula corresponding to the hidden layer; and according to the original output variable range of the network task set by the output layer of the machine learning model, the output layer is encoded to obtain a second satisfiability modulo theory formula corresponding to the output layer; and a problem analysis module is used to call the planner to analyze the first satisfiability modulo theory formula and the second satisfiability modulo theory formula, and generate an explanatory conclusion for characterizing the rationality of the response logic of the machine learning model. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The accompanying drawings illustrate exemplary embodiments of the present disclosure and together with the description serve to explain the principles of the present disclosure. These drawings are included to provide a further understanding of the present disclosure and are incorporated in and constitute a part of this specification.

[0017] Figure 1 This is a block diagram of a method for optimizing network task response information according to an exemplary embodiment of the present disclosure.

[0018] Figure 2 A schematic diagram of the architecture of the verification and interpretation method according to an exemplary embodiment of the present disclosure.

[0019] Figure 3 Schematic diagram of qualitative anchor point rules according to an exemplary embodiment of the present disclosure.

[0020] Figure 4 Schematic diagram of counterfactual rules according to an exemplary embodiment of the present disclosure.

[0021] Figure 5Schematic diagram of decision boundary rules according to an exemplary embodiment of the present disclosure.

[0022] Figure 6 Schematic diagram of feature importance rules according to an exemplary embodiment of the present disclosure.

[0023] Figure 7 A schematic diagram of sensitivity analysis rules according to an exemplary embodiment of the present disclosure.

[0024] Figure 8 This is a time distribution diagram of attribute verification of Aurora according to an exemplary embodiment of the present disclosure.

[0025] Figure 9 This is an explainable runtime distribution diagram of Aurora according to an exemplary embodiment of the present disclosure.

[0026] Figure 10 A schematic diagram of the decision boundary of Aurora according to an exemplary embodiment of the present disclosure. DETAILED DESCRIPTION

[0027] The present disclosure will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only intended to explain the relevant content and are not intended to limit the present disclosure. It should also be noted that, for ease of description, only the portions relevant to the present disclosure are shown in the accompanying drawings.

[0028] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in the present disclosure can be combined with each other. The technical solution of the present disclosure will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0029] Unless otherwise stated, the exemplary embodiments / examples shown are to be understood as providing exemplary features of various details of some ways in which the technical concepts of the present disclosure can be implemented in practice. Therefore, unless otherwise stated, the features of the various embodiments / examples may be further combined, separated, interchanged, and / or rearranged without departing from the technical concepts of the present disclosure.

[0030] The terms used herein are for the purpose of describing specific embodiments, rather than being restrictive. As used herein, unless the context clearly indicates otherwise, the singular forms "one (kind, person)" and "said (the)" are also intended to include plural forms. In addition, when the terms "comprise" and / or "include" and their variations are used in this specification, the features, integral bodies, steps, operations, parts, assemblies and / or their groups stated are indicated, but the presence or addition of one or more other features, integral bodies, steps, operations, parts, assemblies and / or their groups is not excluded. It should also be noted that, as used herein, the terms "substantially", "approximately" and other similar terms are used as approximate terms and not as degree terms, and as such, they are used to explain the inherent deviations of the measured values, calculated values ​​and / or values ​​provided that will be recognized by those of ordinary skill in the art.

[0031] Figure 1 This is a block diagram of an optimization method for network task response information according to an exemplary embodiment of the present disclosure. Figure 1 The optimization S100 of the network task response information is described in detail.

[0032] like Figure 1 As shown, the optimization method S100 for network task response information includes:

[0033] Step S102: Determine at least one verification question about the response logic based on the machine learning model's response logic to the network task.

[0034] Among them, network tasks include network congestion control tasks, video stream rate control tasks, and data center stream scheduling tasks, and response logic is used to characterize the principle of how machine learning models generate response information.

[0035] Step S104: Encode the verification problem to convert it into a first satisfiability modulo theory formula.

[0036] In step S106, the encoder encodes the network structure of the machine learning model to obtain multiple second satisfiability modulo theory formulas.

[0037] Step S108: Call the planner to analyze the first satisfiability modulo theory formula and the second satisfiability modulo theory formula to generate an explanatory conclusion for characterizing the rationality of the response logic of the machine learning model.

[0038] Response logic refers to the principle by which a machine learning model generates response information after executing a network task. For different verification questions, the machine learning model will generate corresponding response information, which will also correspond to different response logics. The disclosure provides a convenient way for users to understand the response principle of the machine learning model and judge the accuracy of the machine learning model's response information. Users can also optimize the parameters and adjust the structure of the machine learning model based on the interpretation conclusions, or optimize the input variables when issuing network tasks, to improve the reliability and accuracy of the machine learning model's response information.

[0039] The first satisfiability modulo theory formula is the encoding result of the verification explanation problem, and its essence is the SMT (Satisfiability Modulo Theories) formula.

[0040] In related technologies, two mechanisms are usually used to separately handle verification problems and explainability problems, but the present disclosure can convert the two into SMT problem 2 for unified processing.

[0041] The second satisfiability modulo theory is the result of the encoder encoding the machine model. Its essence is also an SMT formula. Because machine learning models have multiple processing layers, they are processed layer by layer during the encoding process, and each layer will correspond to a second satisfiability modulo theory formula.

[0042] In some embodiments, step S106 includes: encoding the input layer according to the input variable range of the network task set by the input layer of the machine learning model to obtain a second satisfiability modulo theory formula corresponding to the input layer; constructing a relationship expression between the input variables of the machine learning model about the network task and the original output variables generated by the response network task according to the construction logic of the machine learning model, and encoding the relationship expression to obtain a second satisfiability modulo theory formula corresponding to the hidden layer; and encoding the output layer according to the original output variable range of the network task set by the output layer of the machine learning model to obtain a second satisfiability modulo theory formula corresponding to the output layer.

[0043] In some embodiments, the method further includes: constructing a model robustness formula for the relationship between the input variables corresponding to the network task and the original output variables generated by the response network task based on the various attributes of the machine learning model and the second satisfiability module theory formula.

[0044] The model robustness formula can be expressed as: Among them, i represents the feature number in the input variable, j represents the feature number in the original output variable, L represents the input variable with disturbance characteristics after disturbance processing, F represents the input variable, P lPredicate logic representing perturbation characteristics, P r Predicate logic representing the unperturbed features, x * Represents the input variable after disturbance adjustment, y * Represents the output variable corresponding to the input variable after disturbance adjustment.

[0045] In some embodiments, the specific implementation process of step S108 includes: performing perturbation processing on the input variables related to the network task to obtain perturbation input variables; inputting the perturbation input variables into the model robustness formula to obtain perturbation output variables, and judging whether the perturbation output variables are consistent with the original output variables; when the perturbation output variables are consistent with the original output variables, performing perturbation processing on the input variables until the generated perturbation input variables make the perturbation output variables inconsistent with the original output variables, and determining the perturbation range of the machine learning model, wherein the perturbation range is used to characterize the robustness of the machine learning model.

[0046] In some embodiments, the specific implementation process of step S108 includes: deleting any feature in the input variables of the network task, inputting the input variables with missing features into the machine learning model, and determining whether the actual output variables of the machine learning model are consistent with the original output variables; and summarizing the features deleted when the actual output variables are consistent with the original output variables to obtain a set of loseable features that can generate the original output variables.

[0047] In some embodiments, the specific implementation process of step S108 includes: taking extreme values ​​for any feature in the input variables related to the network task, and inputting the input variables with extreme values ​​into the machine learning model to determine whether the actual output variables of the machine learning model are consistent with the original output variables; summarizing the features that take extreme values ​​when the actual output variables are consistent with the original output variables, and obtaining a set of features with possible extreme values.

[0048] In some embodiments, the method further includes constructing a model interpretability problem formula based on a second satisfiability modulo theory formula, the predicate logic of the perturbed features, and the predicate logic of the unperturbed features in the input variables.

[0049] In some embodiments, the specific implementation process of step S108 includes: according to the model interpretability problem formula, using the qualitative anchor method to determine the dominant input feature set that plays a dominant role in the machine learning model when answering the network task; and using the quantitative anchor method to determine the value range of each dominant feature in the dominant input feature set; according to the model interpretability problem formula, using the counterfactual explanation algorithm to determine the expected input variables of the machine learning model when answering the network task, so that when the machine learning model responds to the network task based on the expected input variables, it generates the expected output variables that meet the expected response decision; according to the model interpretability problem formula, using the decision boundary algorithm to determine the boundary range of the input variables of the machine learning model when issuing the network task, so that the output of the machine learning model When the input variable is within the boundary range of the input variable, the output variable generated by its processing of the network task is inconsistent with the original output variable; according to the formula of the model interpretability problem, each feature in the input variable of the network task is adjusted one by one, and the minimum attack degree of each feature when causing the original output variable to change is determined respectively, and the importance of each feature of the input variable is ranked according to the minimum attack degree, where the minimum degree of the feature is inversely proportional to the importance of the feature; and according to the formula of the model interpretability problem, any feature in the input variable of the network task is perturbed to obtain the output variable corresponding to the feature as the perturbation feature, and the degree of deviation between the output variable and the original output variable is determined, where the degree of deviation is used to characterize the sensitivity of the machine learning model to feature changes.

[0050] In some embodiments, after step S108, the process includes: calling a satisfiability modulo theory engine to analyze the verification problem in the first satisfiability modulo theory formula, and providing feedback on the interpretation conclusion of the verification problem through a planner.

[0051] In some embodiments, after step S108, the process includes optimizing the machine learning model based on the interpretation conclusion and adjusting the input variables corresponding to the network task so that the response information generated by the machine learning model processing the network task is consistent with the expected response result.

[0052] Figure 2 A schematic diagram of the architecture of the verification and interpretation method according to an exemplary embodiment of the present disclosure.

[0053] refer to Figure 2The verification and explanation method framework UINT unifies various verification and explanation problems through the satisfiability model theory SMT. In essence, it is a unified framework for universal verification and explanation of network machine learning models. Specifically, given a machine learning model (the satisfiability model theory formula after model encoding is N), a set of system input constraints (predicate logic P) and a set of system output constraints (predicate logic Q), we can check To verify whether a property always holds, we need to check whether it is not satisfied. Furthermore, given the encoding result N of a machine learning model and the predicate logic Q output by the system, by repeatedly checking the satisfiability and related properties of multiple SMT problems, we can find a predicate P that meets certain criteria and guarantees that Q always holds. UINT summarizes various property verification and interpretability problems into SMT formulas, encodes machine learning models into SMT formulas, and uses SMT solvers to solve the corresponding verification and interpretability problems.

[0054] UINT operates in four main steps. First, given a machine learning model, an operator can encode a verification or interpretability problem (Verification / Interpretability Problem) into an SMT formula (i.e., the first satisfiability modulo theory formula mentioned above) and send it along with the machine learning model to the UINT encoder. Second, the encoder encodes the received machine learning model into an SMT formula (i.e., the second satisfiability modulo theory formula) and sends it along with the received first satisfiability modulo theory formula (V / I Problem SMT Formula) to the UINT planner. Third, the planner analyzes the received verification / interpretability problem and determines the corresponding SMT problem to solve. Fourth, the planner iteratively calls the SMT engine (SMT Engine) to solve the received verification / interpretability problem, returning the final answer to the verification or interpretability problem to the operator. Of course, the process of converting the verification / interpretation problem into an SMT formula can also be performed by the corresponding module and is not limited here.

[0055] Specifically, given a machine learning model, we use N to represent the SMT encoding of the model. The verification problem of such a system can be defined as checking whether the SMT formula P∧N→Q is a constant true proposition. Among them, P represents the predicate logic of the input variables, and Q is a set of predicate logic representing the output variables. Whether P∧N→Q is constant true is equivalent to Is it not satisfied: For any input variable x that satisfies P, its output variable y = N(x) always satisfies Q. Otherwise, there is at least one counterexample x that satisfies P, but y = N(x) does not satisfy Q.

[0056] For example, the property of adversarial perturbation requires that for a given input variable with m features and its output y * =N(x * ), the output of the model remains unchanged when each feature is perturbed by no more than ∈. This property can be verified by checking whether the following SMT formula is satisfied:

[0057]

[0058] Similarly, the interpretability problem of machine learning models can be defined as N and a set of predicates Q, and to find a set of predicates P on the input variables of the model such that Not satisfied; P satisfies some pre-set criteria. This problem can be solved by checking the satisfiability and related properties of multiple SMT formulas.

[0059] The following describes how UINT verifies common properties of network machine learning models. First, we describe how UINT encodes machine learning models. Then, we describe how to represent a wide range of properties by specifying predicate logic P via model input and predicate Q via model output.

[0060] Model encoding. Assume that a machine learning model is a deep neural network. Of course, other models, such as decision trees and support vector machines (SVMs), can also be encoded similarly. Specifically, model encoding is divided into three stages: input layer encoding, hidden layer encoding, and output layer encoding. For the input layer and output layer, we encode them into predicates based on the value ranges of the input and output to obtain the second satisfiability modulo theory formula of the input layer and the second satisfiability modulo theory formula of the output layer. A simple example is that the output layer of the machine learning model is a classification network layer with 6 outputs, which is encoded as:

[0061] The encoding of the hidden layer is slightly more complex and requires constructing the relationship expression between the input variables and the output variables based on the internal logic of the neural network. To this end, UINT provides encoding for fully connected layers, convolutional layers, and activation functions. Given a fully connected layer, assuming the input is an m-dimensional vector x and the output is an n-dimensional vector y, the weight matrix of this layer is W, b n Indicates bias. The network logic of this layer can be seen as: y=A(Wx+b n), where A is the activation function. The encoding of this fully connected layer is a set of linear mathematical expressions:

[0062]

[0063] A convolutional layer is essentially a matrix multiplication of the input variable and the convolution kernel K, which is a multidimensional vector. We select the value of the first channel in the convolution kernel, treat this reduced-dimensional matrix as a weight matrix, and transform the convolutional layer operation into a fully connected layer. For the activation function, we use the ReLU function: ReLU(x, y): iff y = max(0, x). We also use ReLU to approximate other activation functions (such as softmax or sigmoid).

[0064] Attribute encoding. UINT unifies a wide range of attributes into the form of SMT formulas, thereby completing verification. We summarize the common attributes of machine learning models into the robustness of the model, that is, verifying the resilience of the system to perturbations. This type of attribute tests whether the model can maintain its original output variables when the input features deviate from their original values. We provide a general description for robustness attributes. Specifically, suppose that the model N has a robustness to the input x * The output is y * F represents the set of features of all input variables, and L represents the set of features of the perturbed input variables. We use P l represents the predicate logic on the perturbation feature, P r Represents the predicate logic on other input features, and sets the output predicate logic Q to N(x) = y * The unified formula for the robustness of the model (i.e. the model robustness formula) is:

[0065]

[0066] Table 1

[0067]

[0068] Table 1 shows the different robustness attributes. All attributes include adversarial perturbation, missing features, extreme values, and corresponding L, P of each attribute. l and P r , please refer to the table, no more details. l and P r , which will make formula (1) have different robustness properties.

[0069] For adversarial perturbations, given an (input, output) mapping (x *,y * ), when each input feature may have a slight perturbation, check whether the output variable remains unchanged. This property can help network operators understand the resilience of the model. We can set L = F and set P according to the perturbation capacity. l Specifically, let k i is x i Then, an attacker with the jamming capability ∈ can Internal change x i We summarize this property as:

[0070]

[0071] Right now:

[0072] If the corresponding SMT formula cannot be satisfied, it means that the system will be immune to interference from attackers whose interference capability is within ∈.

[0073] For feature loss, given an (input, output) mapping (x * ,y * ), verify whether the machine learning model will maintain the original output variable y when any feature is lost * Assuming there is a set of missing features K, we can generalize this property by setting L = K and setting P accordingly. l and P r :

[0074] Among them, (l i , r i ) is the input feature x i Valid range.

[0075] For extreme values, given an (input, output) mapping (x * ,y * ), when one feature takes an extreme value while other features remain within their original value range, verify whether the output variable of the model remains unchanged. Assuming that the set of features that may take extreme values ​​is K, this property can be specified as:

[0076] in, is the input feature x i extreme value range.

[0077] Below, we'll describe how UINT answers common interpretability questions by examining relevant properties and the satisfiability of multiple SMT formulas. UINT focuses on explaining trained machine learning models and answering interpretable questions about the model's internals. Specifically, UINT focuses on two types of problems: rule interpretation problems, which aim to find a set of rules that explain why a model makes a certain prediction; and feature attribution problems, which aim to identify which features are important and less important, exploring the contribution of features to model outputs. These two types of interpretability problems represent the most important and common interpretability issues in network machine learning systems.

[0078] We provide a general formulation for these interpretability problems. Suppose the model N is based on the input variables x * The output variable is y * Let F denote the set of all input features and L denote the set of input features of interest. We let P l represents the predicate on the input feature L, and P r Represents the predicate on other input features. Set the output predicate Q on N(x), and the unified formula (i.e., the explainability problem formula) is:

[0079]

[0080] Table 2

[0081]

[0082] Table 2 is a table of different interpretability problems. It gives the qualitative anchor, quantitative anchor, counterfactual explanation, decision boundary, feature importance and sensitivity analysis, and gives the L and P of each interpretability problem. l 、P r and Q(N(x)), which are not listed here one by one.

[0083] By L, P l and P r With different settings, formula (2) can express different interpretability problems. The above interpretability problems will be described in detail below.

[0084] For rule interpretation, given a machine learning model, we use N to represent the SMT encoding result of the machine learning model, and given input data with m features and its output y * =N(x * ). F represents the m-dimensional feature set of the model. The purpose of rule interpretation is to find a set of predicates P = {r1, r2, ···, r m}, which represent the input variable x * and the corresponding output variable N(x * ) between the logical relationship: P(x * )→Q(N(x * )).

[0085] Table 3

[0086]

[0087] Table 3 is a comparison table of rule explanations. The explainability issues related to rule explanations include qualitative anchors, quantitative anchors, counterfactual explanations, and decision boundaries. The table also gives P(x) and Q(N(x)) corresponding to different issues.

[0088] Given a machine learning model N with an (input, output) mapping (x * ,y * ), the goal of the anchor problem is to find a set of predicates of dominant features that are decisive for the predicted output, while other features are irrelevant to the model's decision.

[0089] For qualitative anchor points, the goal is to find * )=y * The minimal set F with decisive dominant characteristics d , so for F\F d Other features in can take any value without changing the output:

[0090] Figure 3 Schematic diagram of qualitative anchor point rules according to an exemplary embodiment of the present disclosure.

[0091] like Figure 3As shown in Figure 2, to solve this problem, UINT leverages the concepts of unsat core and MaxSAT problems in SMT. Specifically, an SMT formula can be written as a connective normal form (CNF) of clauses. By marking clauses as "hard" or "soft" constraints, when the formula cannot be satisfied, the MaxSAT solver can return a set of maximum soft clauses that make the formula satisfiable, which can be satisfied by an assignment that satisfies all hard clauses at the same time. The complement of this set of maximum satisfiability clauses (CoMSS) is a set of soft clauses with minimum cardinality, and removing this complement will make the original formula satisfiable. Using these concepts, we map the given data

[0092] We can find the minimum feature set F through the soft-constrained CoMSS d .exist Figure 3 This process is summarized in Algorithm 1 shown.

[0093] Similarly, if we are given another (input, output) mapping (x', y'), we can get the guarantee that the output from y is * The minimum dominant feature set for the change from y to y'.

[0094]

[0095] For the quantitative anchor, given the (input, output) mapping (x * ,y * ) and its qualitative anchor point (i.e., the dominant feature set F d ), the quantitative anchor point problem aims to * The main feature set F d Find a set of predicates P a ={r1,r2,···,r d}, where each r i All with x i ∈,v1,v2], the SMT formula is as follows:

[0096] In other words, the goal of quantitative anchoring is to find the dominant feature x for each i The value range (i∈F d ), ensuring that the model output remains y * .

[0097] We designed an algorithm to find the quantitative anchor point. Specifically, after using Algorithm 1 to find F d After that, we use the attack capability of ∈ to perturb the dominant feature. When the following equation is satisfied, it means that there are input values ​​that cause the model to output different values ​​(the range of the soft clause is when the output is y* The value range is outside the range), and then ∈ is reduced until the following formula is not satisfied, the corresponding CoMSS L is the value range that keeps the prediction unchanged, and these clauses are saved in P a Then from F d Remove the features present in L to update F d and C1, repeat this process until P is completed a .

[0098]

[0099] Figure 4 Schematic diagram of counterfactual rules according to an exemplary embodiment of the present disclosure.

[0100] like Figure 4 As shown, for counterfactual explanations, given a (input, output) mapping (x * ,y * ), and the expected predicted output y e , the goal of the counterfactual explanation problem is to find * The closest input x c , so that N(x c )=y e A commonly used distance metric d is the L1 norm weighted by the inverse median absolute deviation (MAD), expressed as:

[0101]

[0102] where x D is x * neighbors.

[0103] We designed Figure 4 Algorithm 3 is shown to find x c Specifically, we first set the possible search range of d to (l d , r d ). Then, we solve the following SMT formula by iteration:

[0104]

[0105] Perform a binary search on this range. In each iteration, if formula (4) can be satisfied, it means the upper bound r of d d Can be reduced. If formula (4) is not satisfied, it means that the range is too small and the lower bound l should be increased. d At the end of this search process, we are able to find a counterfactual x c , it to x * The distance is the smallest within the accuracy range of δ.

[0106] Figure 5 Schematic diagram of decision boundary rules according to an exemplary embodiment of the present disclosure.

[0107] like Figure 5 As shown, for the decision boundary, given a (input, output) mapping (x * ,y * ), the goal of the decision boundary problem is to find a set of input variable vectors {x} that transform the feature space X D Divided into two parts and The SMT representation of the interpretation of the decision boundary is:

[0108]

[0109] We focus on two features x a , x b In the feature space X D The decision boundary of , iteratively checks the SMT formula in formula (5). Each time the solver finds it to be satisfiable, it returns a satisfiable assignment. To avoid SMT returning duplicate solutions, we add the solution constraint C4 to formula (5) to ensure that the next returned solution is outside the neighborhood of known solutions. By iteratively solving the SMT formula, we get enough x a and x b The specific process is shown in Figure 5 Algorithm 4 in

[15] .

[0110]

[0111] For feature attribution, in addition to finding the rules that each feature must satisfy, people are also interested in the contribution of that feature to the predictions of a network machine learning system. This type of interpretability problem is called "feature attribution." We focus on feature importance and sensitivity analysis because these are the two most important problems in feature attribution, from which all other subproblems are derived.

[0112] Figure 6 Schematic diagram of feature importance rules according to an exemplary embodiment of the present disclosure.

[0113] like Figure 6 As shown in Figure 2, for feature importance, the problem aims to obtain a ranked array M of model feature importance. We use a perturbation-based metric to obtain feature importance: features that contribute more to the original classification will be those that contribute more to the misclassification. Therefore, we rank feature importance based on how resistant they are to perturbations.

[0114] Given an (input, output) mapping (x* ,y * ), we can get the importance array M by comparing the minimum attack degree ∈ of each feature that can change the output. The smaller the minimum attack degree, the more important the feature. Figure 6 Algorithm 5 in the figure proposes the process of finding the minimum attack level. Given the SMT formula in equation (6), set the perturbation feature set F P , k i Set to x i The difference between the maximum and minimum values ​​of the system is initialized according to the actual system range (l ∈ , r ∈ )’s attack level∈. We interfere with F P Internal features are used to verify whether the current attack level ∈ can change the output. If the solver returns unsat, indicating that the current ∈ cannot change the output, ∈ is updated by multiplying α by ∈ until the SMT returns sat, and the current ∈ is set as the upper limit. We can then use binary search to determine the minimum attack capability ∈O with accuracy δ.

[0115]

[0116] Figure 7 A schematic diagram of sensitivity analysis rules according to an exemplary embodiment of the present disclosure.

[0117] like Figure 7 As shown, for sensitivity analysis, the purpose of this problem is to evaluate the uncertainty of the model output when there is uncertainty in its input. Given an (input, output) mapping (x * ,y * ), we randomly perturb the selected feature set to check the change of the output. We use the output deviation degree s(F p )=N(x)-N(x * ) as feature F P A measure of sensitivity.

[0118] We use the following SMT formula as the constraint, we can get the value of N(x) returned by the SMT solver and calculate s(F) by subtracting the original output value. P ) value, the specific method is as follows Figure 7 As shown in Algorithm 6 in .

[0119]

[0120] We implemented a prototype of UINT and experimentally evaluated its performance on a network machine learning system. The results demonstrate that our unified framework can verify a wide range of properties of different systems and provide solutions to different interpretability problems, demonstrating the effectiveness of UINT.

[0121] We evaluate UINT on two representative systems, Pensieve and Aurora.

[0122] Pensieve is an adaptive video bitrate selection system based on reinforcement learning. It trains a neural network model consisting of two fully connected hidden layers with ReLU activation functions, each containing 128 neurons. Pensieve classifies its 25 features into six categories: previous bit rate, throughput, download time, current buffer size, next chunk sizes, and remaining chunks. The model output is the predicted download bitrate with the highest probability.

[0123] Aurora is a congestion control system based on reinforcement learning. It uses a neural network architecture with two hidden layers. We trained four systems with different numbers of neurons per layer (32*16, 32*32, 48*32, and 48*16). Aurora has three features: latency inflation, latency ratio, and send ratio. The model output is the ratio of the current transmission rate to the transmission rate of the previous action. Decisions are made based on the latest k-step history of all features.

[0124] Table 4

[0125]

[0126] Table 4 is a table of experimental parameter settings for verification properties. The experimental parameters for verification properties of each system are presented in Table 4 and are not listed one by one.

[0127] Table 5

[0128]

[0129] Table 5 is the experimental parameter setting table for interpretability problems. For experimental settings of interpretability problems such as rule interpretation and feature attribution, please refer to Table 5 and will not be listed one by one.

[0130] All experiments are run on a server with 2*Intel Xeon Silver 4210R@2.40GHz CPU and 128G DDR4 memory, and the prototype UINT uses Z3 as the SMT engine.

[0131] Table 6

[0132] Property 50％ 70％ 90％ Adversarial Perturbation 168.57s 309.88s 1494.31s Missing Features 518.82s 1608.37s 3596.76s Missing Features (|K|=2) 4198.37s 6252.11s 45181.94s

[0133] Table 6 shows the Pensieve attribute verification time distribution. Based on the Pensieve verification time distribution results shown in Table 6, UINT performs well when features are missing or their values ​​fluctuate within a certain range. In addition to the Aurora model with a 32x16 network architecture, we trained three other Aurora system models with different neural network structures (32x32, 48x16, and 48x32) and conducted experiments on these systems to demonstrate the effectiveness and scalability of UINT.

[0134] Figure 8 This is a time distribution diagram of attribute verification of Aurora according to an exemplary embodiment of the present disclosure.

[0135] refer to Figure 8 We verified the robustness properties on Aurora. The verification running time of different properties on Aurora of each neural network structure is as follows: Figure 8 As shown in Figure 3, the runtime of UINT increases slightly with the problem size. Results on Aurora neural networks of different scales demonstrate that our framework, UINT, is general and scalable for common verification problems in machine learning models of different scales.

[0136] Table 7

[0137] Feature ∈ Rank Time(s) Previous bit rate 0.331 1 833.08 Buffer size 0.334 2 3169.283 Throughput 0.437 3 11064.44 Download time 1.599 4 33138.59

[0138] Table 7 is the minimum attack level table of Pensieve. We first show the minimum attack level results on the Pensieve system in Table 7, and then get the feature importance of Pensieve. We see that UINT takes hundreds to tens of thousands of seconds to find the minimum ∈ and obtain the importance ranking. Considering that many studies use SMT and some other interpretable methods with a running time of mainly 10 3 -10 4 seconds. In addition, Pensieve is a relatively complex machine learning model, and interpretability is performed before model deployment, so the time overhead is reasonable and acceptable.

[0139] Table 8

[0140] Feature ∈ Rank Time(s) Latency ratio (History = 5) 0.004 1 98.49 Send ratio (History = 5) 0.013 2 120.04 Latency inflation(History=5) 0.034 3 112.46 Latency ratio (History = 10) 0.002 1 233.79 Send ratio (History = 10) 0.009 2 335.39 Latency inflation(History=10) 0.023 3 233.79

[0141] Table 8 is the minimum supply level (32*16) table of Aurora. Figure 9 This is an explainable runtime distribution diagram of Aurora according to an exemplary embodiment of the present disclosure.

[0142] We have in Table 8 and Figure 9Results for Aurora are shown in Figure 2. We trained four Aurora models with different neural network architectures (32x16, 32x32, 48x16, and 48x32) and deployed experiments on these systems to demonstrate the effectiveness and scalability of UINT. We found that for qualitative anchors, UINT found a median answer in less than 100 seconds. While Aurora's median answer time was relatively high, most were within 1000 seconds. This demonstrates the versatility and scalability of UINT in answering a wide range of interpretability questions.

[0143] Figure 10 A schematic diagram of the decision boundary of Aurora according to an exemplary embodiment of the present disclosure.

[0144] Figure 10 The figure shows the decision boundaries we obtained for Aurora's latency inflation and latency ratio. We can see that a set of feature values ​​divides the decision space into two regions, forming a decision boundary that allows Aurora to maintain its original output of 3.57 in the blue region and change the output in the red region. This result allows operators to intuitively see that model decisions will conform to expected behavior within a certain feature range. When making decisions near the boundary, the operator bases them on previous decisions to avoid sudden changes in the transmission rate, ensuring a smooth rate selection process. It also provides insights into the model: for example, when the latency inflation is -0.2, the model decisions in Aurora are more tolerant to the latency ratio compared to when the latency inflation is -0.4.

[0145] In response to the black-box nature of machine learning models, the general framework disclosed in this paper is suitable for unified verification and interpretation of machine learning models. It can be applied to the machine learning models required for various network tasks, providing key technical support for system users to better understand model decisions.

[0146] This disclosure can also facilitate system users' understanding of the underlying model logic and improve related system designs through attribute verification and explainability. It can also promote the large-scale deployment of machine learning systems in computer network tasks, providing strong infrastructure support for the development of new technologies and applications.

[0147] Another aspect of the present disclosure provides an optimization system for network task response information, including: a problem determination module, used to determine at least one verification problem about the response logic of the network task based on the machine learning model, wherein the network task includes a network congestion control task, a video stream rate control task, and a data center stream scheduling task; a problem encoding module, used to encode the verification problem to convert the verification problem into a first satisfiability modulo theory formula; a model encoding module, used to encode the network structure of the machine learning model by an encoder to obtain multiple second satisfiability modulo theory formulas; and a problem analysis module, used to call a planner to analyze the first satisfiability modulo theory formula and the second satisfiability modulo theory formula to generate an explanatory conclusion for characterizing the rationality of the response logic of the machine learning model.

[0148] The network task response information optimization system provided by the present disclosure can also be implemented by a device using a hardware implementation method of a processing system.

[0149] The device may include corresponding modules for executing each or several steps in the above flowchart. Therefore, each step or several steps in the above flowchart may be executed by a corresponding module, and the device may include one or more of these modules. The module may be one or more hardware modules specifically configured to execute the corresponding steps, or implemented by a processor configured to execute the corresponding steps, or stored in a computer-readable medium for execution by a processor, or implemented by some combination thereof.

[0150] The hardware structure can be implemented using a bus architecture. The bus architecture can include any number of interconnecting buses and bridges, depending on the specific application and overall design constraints of the hardware. The bus connects various circuits including one or more processors, memories, and / or hardware modules. The bus can also connect various other circuits such as peripherals, voltage regulators, power management circuits, external antennas, etc.

[0151] A bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Component Architecture (EISA) bus. Buses can be categorized as address buses, data buses, and control buses. For ease of illustration, this diagram uses only one connecting line, but this does not imply that there is only one bus or only one type of bus.

[0152] Any process or method description in the flowchart or otherwise described herein can be understood to represent a module, fragment or portion of code including one or more executable instructions for implementing the steps of a specific logical function or process, and the scope of the preferred embodiments of the present disclosure includes alternative implementations in which the functions may not be performed in the order shown or discussed, including performing the functions in a substantially simultaneous manner or in a reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present disclosure belong. The processor performs the various methods and processes described above. For example, the method embodiments of the present disclosure can be implemented as a software program that is tangibly contained in a machine-readable medium, such as a memory. In some embodiments, part or all of the software program can be loaded and / or installed via a memory and / or a communication interface. When the software program is loaded into the memory and executed by the processor, one or more steps in the method described above can be performed. Alternatively, in other embodiments, the processor can be configured to perform one of the above methods in any other appropriate manner (e.g., by means of firmware).

[0153] The logic and / or steps represented in the flowchart or otherwise described herein may be embodied in any readable storage medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device).

[0154] For the purposes of this specification, a "readable storage medium" can be any device that can contain, store, communicate, propagate or transmit a program for use with or in conjunction with an instruction execution system, device or apparatus. More specific examples (a non-exhaustive list) of readable storage media include the following: an electrical connection having one or more wires (electronic device), a portable computer disk cartridge (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and editable read-only memory (EPROM or flash memory), a fiber optic device, and a portable read-only memory (CDROM). In addition, the readable storage medium can even be paper or other suitable medium on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium and then editing, interpreting or processing it in other suitable ways as necessary, and then storing it in a memory.

[0155] It should be understood that various parts of the present disclosure can be implemented using hardware, software, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used to implement the hardware: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0156] Those skilled in the art will understand that all or part of the steps of the above-mentioned implementation method can be accomplished by instructing related hardware through a program, and the program can be stored in a readable storage medium. When the program is executed, it includes one or a combination of the steps of the method implementation method.

[0157] Furthermore, the functional units in the various embodiments of the present disclosure may be integrated into a single processing module, each unit may exist physically separately, or two or more units may be integrated into a single module. The aforementioned integrated modules may be implemented in the form of hardware or software functional modules. If the integrated modules are implemented as software functional modules and sold or used as independent products, they may also be stored in a readable storage medium. The storage medium may be a read-only memory, a magnetic disk, or an optical disk, etc.

[0158] Those skilled in the art will appreciate that the above embodiments are merely intended to clearly illustrate the present disclosure and are not intended to limit the scope of the present disclosure. Other changes or modifications may be made based on the above disclosure, and such changes or modifications are still within the scope of the present disclosure.

Claims

1. A method for optimizing network task response information, characterized in that: include: Determining, based on the response logic of the machine learning model to the network tasks, at least one verification problem for obtaining the response logic, wherein the network tasks include a network congestion control task, a video stream rate control task, and a data center stream scheduling task, and the response logic represents a principle by which the machine learning model generates response information; encoding the verification problem to convert the verification problem into a first satisfiability modulo theory formula; The encoder encodes the network structure of the machine learning model to obtain multiple second satisfiability modulo theory formulas, including: encoding the input layer according to the input variable range of the network task set by the input layer of the machine learning model to obtain the second satisfiability modulo theory formula corresponding to the input layer; constructing a relationship expression between the input variables of the machine learning model with respect to the network task and the original output variables generated in response to the network task according to the construction logic of the machine learning model, and encoding the relationship expression to obtain the second satisfiability modulo theory formula corresponding to the hidden layer; and encoding the output layer according to the original output variable range of the network task set by the output layer of the machine learning model to obtain the second satisfiability modulo theory formula corresponding to the output layer; Invoking a planner to analyze the first satisfiability modulo theory formula and the second satisfiability modulo theory formula to generate an explanation conclusion for characterizing the rationality of the response logic of the machine learning model; The optimization method further includes: Construct a model interpretability problem formula based on the second satisfiability modulo theory formula, the predicate logic of the perturbed features, and the predicate logic of the unperturbed features in the input variables; The call planner analyzes the first satisfiability modulo theory formula and the second satisfiability modulo theory formula to generate an explanation conclusion for characterizing the rationality of the response logic of the machine learning model, including: According to the model interpretability problem formula, a qualitative anchor method is used to determine a dominant input feature set that plays a dominant role in the machine learning model's response to the network task; and a quantitative anchor method is used to determine the value range of each dominant feature in the dominant input feature set; Determining, according to the model interpretability problem formula, expected input variables of the machine learning model when responding to the network task using a counterfactual explanation algorithm, so that when the machine learning model responds to the network task based on the expected input variables, it generates expected output variables that conform to the expected response decision; According to the model interpretability problem formula, a decision boundary algorithm is used to determine the boundary range of the input variables of the machine learning model when issuing the network task, so that when the input variables of the machine learning model are within the input variable boundary range, the output variables generated by the machine learning model in processing the network task are inconsistent with the original output variables; According to the model interpretability problem formula, each feature of the input variable of the network task is adjusted one by one, and the minimum degree of attack of each feature when causing the original output variable to change is determined, and the importance of each feature of the input variable is ranked according to the minimum degree of attack, wherein the minimum degree of a feature is inversely proportional to the importance of the feature; and According to the model interpretability problem formula, any feature of the input variables of the network task is perturbed to obtain the output variable corresponding to the feature as the perturbation feature, and the degree of deviation between the output variable and the original output variable is determined, wherein the degree of deviation is used to characterize the sensitivity of the machine learning model to changes in the feature.

2. The optimization method for network task response information according to claim 1, characterized in that: Also includes: Based on the various attributes of the machine learning model and the second satisfiability module theory formula, a model robustness formula is constructed regarding the relationship between the input variables corresponding to the network task and the original output variables generated in response to the network task.

3. The optimization method for network task response information according to claim 2, characterized in that: The call planner analyzes the first satisfiability modulo theory formula and the second satisfiability modulo theory formula to generate an explanation conclusion for characterizing the rationality of the response logic of the machine learning model, including: Performing a disturbance process on the input variables related to the network task to obtain a disturbed input variable; Inputting the disturbance input variable into the model robustness formula to obtain a disturbance output variable, and determining whether the disturbance output variable is consistent with the original output variable; When the perturbation output variable is consistent with the original output variable, the input variable is perturbated until the generated perturbation input variable makes the perturbation output variable inconsistent with the original output variable, and the perturbation range of the machine learning model is determined, wherein the perturbation range is used to characterize the robustness of the machine learning model.

4. The method for optimizing network task response information according to claim 2, characterized in that: The call planner analyzes the first satisfiability modulo theory formula and the second satisfiability modulo theory formula to generate an explanation conclusion for characterizing the rationality of the response logic of the machine learning model, including: Deleting any feature from the input variables of the network task, inputting the input variables without the feature into the machine learning model, and determining whether an actual output variable of the machine learning model is consistent with the original output variable; and The features deleted when the actual output variable is consistent with the original output variable are summarized to obtain a set of loseable features capable of generating the original output variable.

5. The method for optimizing network task response information according to claim 2, characterized in that: The call planner analyzes the first satisfiability modulo theory formula and the second satisfiability modulo theory formula to generate an explanation conclusion for characterizing the rationality of the response logic of the machine learning model, including: Taking an extreme value for any feature of the input variables related to the network task, inputting the input variable with the extreme value into the machine learning model, and determining whether an actual output variable of the machine learning model is consistent with the original output variable; The features that take extreme values ​​when the actual output variable is consistent with the original output variable are summarized to obtain a set of possible extreme value features.

6. The method for optimizing network task response information according to claim 1, characterized in that: After the call planner analyzes the first satisfiability modulo theory formula and the second satisfiability modulo theory formula to generate an explanation conclusion for characterizing the rationality of the response logic of the machine learning model, the method further includes: The satisfiability modulo theory engine is called to analyze the verification problem in the first satisfiability modulo theory formula, and the interpretation conclusion on the verification problem is fed back through the planner.

7. The method for optimizing network task response information according to claim 1, characterized in that: After the call planner analyzes the first satisfiability modulo theory formula and the second satisfiability modulo theory formula to generate an explanation conclusion for characterizing the rationality of the response logic of the machine learning model, the method further includes: Based on the interpretation conclusion, the machine learning model is optimized and the input variables corresponding to the network task are adjusted so that the response information generated by the machine learning model when processing the network task is consistent with the expected response result.

8. An optimization system based on the optimization method for network task response information according to any one of claims 1 to 7, characterized in that: include: a problem determination model for determining, based on a response logic of a machine learning model to a network task, at least one verification problem regarding the response logic, wherein the network task comprises a network congestion control task, a video stream rate control task, and a data center stream scheduling task, and the response logic represents a principle by which the machine learning model generates response information; a problem encoding module, configured to encode the verification problem so as to convert the verification problem into a first satisfiability modulo theory formula; A model encoding module, configured to encode the network structure of the machine learning model by an encoder to obtain a plurality of second satisfiability modulo theory formulas, including: encoding the input layer according to the input variable range of the network task set by the input layer of the machine learning model to obtain the second satisfiability modulo theory formula corresponding to the input layer; constructing a relationship expression between the input variables of the machine learning model with respect to the network task and the original output variables generated in response to the network task according to the construction logic of the machine learning model, and encoding the relationship expression to obtain the second satisfiability modulo theory formula corresponding to the hidden layer; and encoding the output layer according to the original output variable range of the network task set by the output layer of the machine learning model to obtain the second satisfiability modulo theory formula corresponding to the output layer; and A problem analysis module is used to call a planner to analyze the first satisfiability modulo theory formula and the second satisfiability modulo theory formula, and generate an explanatory conclusion for characterizing the rationality of the response logic of the machine learning model.

Images