Method and system for encrypted authentication task offloading based on multi-objective optimization in edge collaborative power network

Abstract

The application discloses a kind of multi-objective optimization-based encryption authentication task unloading method and system in edge-end cooperative power network, comprising the following steps: collecting power terminal device information in the coverage range of edge gateway, obtaining power terminal device state vector;Obtain edge gateway northbound link, CPU and buffer information, and integrate the obtained power terminal device state vector to obtain state vector;Construct and train encryption authentication task unloading strategy network, input state vector into the trained strategy network, and obtain action vector;According to action vector, the signature of corresponding power terminal device is unloaded to gateway, and aggregated authentication is realized.The method of the present application overcomes the problems of millisecond-level real-time authentication difficulty, multi-objective competition optimization difficulty, slow training convergence and large update amplitude, and can be widely applied to various heterogeneous device access edge computing scenarios, with strong generalization ability.

Description

Technical Field

[0001] This invention belongs to the fields of cyberspace security, power Internet of Things and reinforcement learning, and specifically relates to a method and system for offloading encryption authentication tasks based on multi-objective optimization in edge-end collaborative power networks. Background Technology

[0002] With the deep integration of smart grids and IoT technologies, the number of power terminal devices (such as smart meters and distribution automation terminals) is exploding. These devices often require frequent identity authentication and encrypted data transmission to ensure the security of power big data. However, power terminal devices are usually limited by battery capacity and computing power, making it difficult for them to independently handle high-frequency, high-intensity encryption authentication computing tasks. The introduction of edge computing technology, by offloading computing tasks to edge gateways, provides an effective way to solve this contradiction. Therefore, how to rationally formulate task offloading and resource allocation strategies in a dynamically changing edge-coordinated power network environment to maximize authentication success rate while minimizing authentication latency and terminal energy consumption is of great practical significance.

[0003] To address the aforementioned computational offloading and resource allocation problems, existing solutions mainly fall into two categories: traditional optimization algorithms and deep reinforcement learning-based methods. The first category comprises traditional methods based on rules or convex optimization (such as Lyapunov optimization and game theory). These methods typically require pre-establishing a precise mathematical model of the system and iteratively solving for the optimal solution; for example, optimizing task scheduling energy consumption by constructing a queuing theory model. The second category is based on single-objective deep reinforcement learning (such as DQN and DDPG). These methods combine multiple indicators such as latency and energy consumption into a scalar reward function through weighted summation, utilizing neural networks to fit the mapping relationship between states and actions, achieving end-to-end decision control. The third category is conventional multi-objective policy gradient methods. These methods attempt to directly find Pareto optimal solutions in a multi-dimensional objective space, with each update calculating the gradient based on the currently collected samples to update the policy parameters.

[0004] However, the aforementioned existing methods all have some significant drawbacks when applied to edge-coordinated power networks with high dynamics and high security requirements:

[0005] First, traditional methods based on rules or convex optimization are highly dependent on the accuracy of the system model. However, the power Internet of Things environment is highly time-varying and random (such as power fluctuations and the sudden arrival of tasks), making it difficult to establish a precise model once and for all. Furthermore, traditional iterative algorithms have high online computational complexity, making it difficult to meet the millisecond-level real-time authentication requirements.

[0006] Second, for methods based on single-objective deep reinforcement learning, the linear weighting approach used is too simplistic and crude. In power scenarios, latency, energy consumption, and success rate often have a non-linear competitive relationship. The weighted summation with fixed weights will mask the conflicts between sub-objectives, causing the algorithm to fail to adaptively balance multiple objectives according to the actual power grid conditions. This often results in suboptimal decisions such as "sacrificing critical authentication timeliness for energy saving."

[0007] Third, conventional multi-objective policy gradient methods not only face the problem of conflicting gradients among multiple objectives, but more critically, they suffer from low sample efficiency and huge variance. Because their gradient estimation is usually based on small batches of samples and no effective variance reduction mechanism is introduced, the variance of the policy gradient accumulates with training iterations in a high-dimensional continuous action space (such as continuous CPU frequency allocation), making the model convergence speed extremely slow and making it difficult to train a usable policy within a limited number of power network interactions.

[0008] Fourth, the existing policy gradient-based methods mentioned above typically lack strict constraints on the magnitude of policy updates. When the power grid load fluctuates drastically, the magnitude of the gradient may suddenly increase, leading to excessively large parameter update steps. This causes the new policy to deviate from the trust region of the old policy. This not only disrupts the smoothness of policy iteration but may also cause performance collapse of the agent during training, seriously threatening the safe and stable operation of the power system. Summary of the Invention

[0009] To address the shortcomings of existing technologies, one of the objectives of this invention is to provide a method for offloading encryption authentication tasks based on multi-objective optimization in edge-coordinated power networks, so as to achieve millisecond-level real-time authentication, multi-objective cooperative optimization, fast training iteration convergence speed, and stable measurement update amplitude.

[0010] The second objective of this invention is to provide a system for implementing the encryption authentication task offloading method based on multi-objective optimization in the aforementioned edge-coordinated power network.

[0011] This invention provides a method for offloading encryption and authentication tasks based on multi-objective optimization in edge-to-edge collaborative power networks, comprising the following steps:

[0012] Collect information on power terminal equipment within the coverage area of ​​the edge gateway to obtain the power terminal equipment status vector;

[0013] The edge gateway northbound link, CPU, and buffer information are obtained and integrated with the obtained power terminal device state vector to obtain the state vector;

[0014] A network for offloading encryption authentication tasks is constructed based on a multilayer perceptron. The network includes a shared feature extraction module and an action output module. The shared feature extraction module takes the edge state vector as input and outputs a shared feature vector. The action output module takes the shared feature vector as input and outputs an action vector.

[0015] The historical state vectors of power terminal equipment are obtained to train the encryption authentication task offloading policy network. Edge state vectors are then input into the trained policy network to obtain action vectors; these action vectors include offloading decision vectors. Polymerization coefficient CPU allocation vector ;

[0016] Based on the action vector, the signature of the corresponding power terminal device is offloaded to the gateway.

[0017] The power terminal equipment state vector is obtained by the following method: obtaining the real-time task queue length, remaining battery power and computing power level of the power terminal equipment, and using the max-min normalization method to preprocess the data. Finally, the normalized feature values ​​are combined into a column vector and stacked vertically according to the device ID order to obtain the power terminal equipment state vector.

[0018] For edge gateways and their subordinates Each power terminal device and its edge gateway are configured with a fixed status sampling period. (e.g., 100ms), at each sampling time, the parsing from the first... Terminal devices (of which) From the uplink status messages of the task queue, the remaining battery power and computing power level can be extracted.

[0019] The data is preprocessed using the max-min normalization method, expressed by the following formula: ; ; ;in, For terminal devices Maximum queue length, For terminal devices The total battery capacity; The normalized feature value for the task queue length; The normalized characteristic value of the remaining battery capacity; To calculate the normalized characteristic value of the capability level;

[0020] Will The three normalized feature values ​​of each terminal device are calculated according to... The order is used to form a column vector, and... The row vectors are stacked vertically in order of terminal device ID, thus constructing a dimension of Terminal device state vector .

[0021] Task queue length This represents the amount of data tasks currently backed up in the local buffer of the terminal device, awaiting authentication or computation, in bits.

[0022] Remaining battery power This indicates the remaining percentage or specific charge level of the terminal device's battery relative to its fully charged state, expressed in milliampere-hours (mAh).

[0023] Computational ability level This indicates the CPU processing frequency or floating-point operations per second of the terminal device, reflecting its local task execution speed.

[0024] The edge state vector is obtained by: acquiring the uplink bandwidth utilization, downlink bandwidth utilization, CPU utilization, and buffer utilization of the edge gateway's northbound link, and appending them to the terminal device's state vector. Finally, we obtain the dimension as Edge state vector ;in, Edge state vector Express it using the following formula: Among them, the uplink bandwidth utilization rate of the northbound link and uplink bandwidth utilization Indicates the communication capacity between the edge gateway and the upstream network; CPU utilization. Indicates the busy level of the edge server's computing unit, with a numerical range of [value missing]. Buffer occupancy rate The storage space occupancy ratio of the task input queue on the edge side of the edge gateway reflects the current task backlog and queuing congestion status, with a value range of [value missing]. .

[0025] The policy network to be trained includes a shared feature extraction module and an action output module; the edge state vector obtained as input to the shared feature extraction module is... The output is a dimension Shared feature vectors The action output module inputs a shared feature vector. The final output is the action vector; where, .

[0026] The shared feature extraction module is a multilayer perceptron, consisting of three layers, all using the ReLU activation function; the number of neurons in layers 1, 2, and 3 are respectively... , and .

[0027] The action output module includes three parallel sub-modules: an unloading decision module, an aggregation decision module, and a resource allocation module. The outputs of the three sub-modules are combined to obtain the output of the final policy network.

[0028] The input to the unloading decision module is a shared feature vector. The output dimension is The unloading decision vector; the unloading decision module consists of an MLP with four fully connected layers and a sampling module; the number of neurons in the 1st, 2nd, 3rd, and 4th layers of the MLP in the unloading decision module are respectively , , and The first three layers use the ReLU activation function, and the last layer uses the Sigmoid activation function; the input to the unloading decision module... The dimension is obtained after MLP processing in the unloading decision module. Unloading decision probability vector This includes the unloading probability of each terminal device; then the sampling module targets the vector... The first in element Constructing Bernoulli distribution The unloading decision variables were obtained by sampling using this distribution. Finally, construct the unloading decision vector. And output;

[0029] The input to the aggregation decision module is the shared feature vector. Output aggregation coefficient The aggregation decision module consists of a three-layer MLP, with the number of neurons in layers 1, 2, and 3 of the MLP being respectively... , and The first two layers use the ReLU activation function, while the third layer has no activation function; the shared feature vector is input to the aggregation decision module. The original polymerization coefficient was obtained after MLP processing. The aggregation coefficient is obtained after Gaussian noise reduction and activation. ; in, Gaussian noise was introduced during the training phase, and Follows a mean of 0 and a variance of normal distribution ;

[0030] The input to the resource allocation module is a shared feature vector. The output dimension is CPU allocation vector The resource allocation module consists of a three-layer MLP, a Softmax layer, and a sampling module connected in series. The number of neurons in the first, second, and third layers of the MLP are respectively... , and The first two layers use the ReLU activation function, while the third layer does not; they share feature vectors. The initial CPU allocation vector is obtained after processing by the MLP and Softmax layers. The sampling module is based on Calculate the distribution parameters .in, , Given a vector of length 3, where each element is a vector of length 3. Then, the sampling module constructs a Dirichlet distribution. And sample to obtain CPU allocation vector And output; where, The first concentration parameter of the Dirichlet distribution corresponds to the CPU initial allocation vector. The first component is used to determine the CPU allocation vector. The sampling distribution; The second concentration parameter represents the Dirichlet distribution and corresponds to the CPU initial allocation vector. The second component is used to determine the CPU allocation vector. The sampling distribution; The third concentration parameter represents the Dirichlet distribution and corresponds to the CPU initial allocation vector. The third component is used to determine the CPU allocation vector. The sampling distribution.

[0031] Using probability distribution functions This indicates the policy network. Indicates input Time output The probability, parameter Let the periodic index be a vector consisting of all the parameters to be learned in the policy network. Iterative index The training process of the policy network includes:

[0032] A. Obtain historical information of power terminal equipment within the coverage area of ​​the edge gateway to obtain the historical state vector of the power terminal equipment;

[0033] B. Obtain historical edge gateway northbound link, CPU, and buffer information, and integrate them with the corresponding historical state vector of the power terminal equipment to obtain the historical state vector;

[0034] C. The historical state vector is used as input to the policy network, and the terminal device authentication task is executed based on the obtained action vector. Next, obtain 1 line of length. trajectory ;in, It is a vector, represented using the following formula: ; For the first The edge state vector obtained this time; For the first The action vector obtained this time; It is a three-dimensional vector, specifically in the form of , For the first The average authentication latency obtained this time , For the first The average energy consumption obtained this time , For the first The success rate of the first authentication ;

[0035] D. Obtain the initial values ​​of the policy parameters through Gaussian initialization. Simultaneously, initialize relevant metadata parameters; among which, set the batch size. and trajectory length Learning rate Total return dimension Number of iterations per cycle Total number of cycles Gradient cutoff radius Reward Discount Rate ;

[0036] E. Based on historical data, repeatedly analyze the policy network. Get from The length of the strip is trajectory and collect them to obtain a set ;

[0037] F. Based on trajectory sets The trajectory is estimated using the following function. Total return: Then estimate the current total return of the trajectory set. ;

[0038] G. Project the current total return using the function. Head-guided hyperrectangular feasible region In this process, a feasible total return is obtained. ; where, projection function Express it using the following formula: ; Hyperrectangular feasible region Express it using the following formula: Total feasible return Express it using the following formula: ;in, The feasible range for the corresponding certification delay target reward component. The feasible range for the corresponding energy consumption target return component. The feasible range for the corresponding certification success rate target return component;

[0039] H. Based on historical data, repeat the process from the policy network. Get from The length of the strip is trajectory and collect them to obtain a set ;

[0040] I. Based on trajectory sets Estimate the gradient of the current policy It can be expressed using the following formula: ; where, function Used for trajectory estimation The policy gradient for total return is expressed using the following formula: ; where, convex function Used to merge total returns The three components are defined as follows: ;in, Total return vector The reward component on the certification latency target; Total return vector The return on energy consumption targets; Total return vector The reward component in the certification success rate target;

[0041] J. Based on the current policy gradient Update the strategy parameters to obtain It can be expressed using the following formula: ; where, function Indicates the use of policy gradient Update strategy parameters The result is defined using the following formula: ;

[0042] K. Based on historical data, repeat the process from the policy network. Get from The length of the strip is trajectory and collect them to obtain a set ;

[0043] L. Based on trajectory sets Estimate current total return Then project it onto the hyperrectangular feasible region. In this process, a feasible total return is obtained. ; of which, total return The estimation formula is: ;function Indicates based on sample The estimated policy network Policy gradient, sample From policy network The sampled data is represented by the following formula: ;in, The ratio of importance sampling is expressed using the following formula: ;

[0044] M. Based on historical data, repeat the process from the policy network. Get from The length of the strip is trajectory and collect them to obtain a set ;

[0045] N. Based on trajectory sets Estimate the gradient of the current policy It can be expressed using the following formula: ;function Indicates the use of trajectory Policy Network Policy gradient, where trajectory From policy network The sampled data is represented by the following formula: ;

[0046] O. Based on the current gradient Update the strategy parameters to obtain ;

[0047] P. Iterative Index Increase by 1, and then determine If the condition is true, return to step K and continue the iteration; otherwise, obtain the updated condition. , and ;

[0048] Q. Set the initial values ​​for the strategy parameters in the next cycle. , will the cycle Increase by 1, and then determine Check if the condition is met; if it is met, return to step E; if it is not met, end the training algorithm and obtain the final policy parameters. .

[0049] The trained policy network includes a shared feature extraction module and an action output module; the shared feature extraction module takes as input the edge state vector. The output is a dimension Shared feature vectors The action output module inputs a shared feature vector. The final output is the action vector; where, .

[0050] The shared feature extraction module is a multilayer perceptron, comprising three fully connected layers, all using the ReLU activation function; the number of neurons in layers 1, 2, and 3 are respectively... , and .

[0051] The action output module includes three parallel sub-modules: an unloading decision module, an aggregation decision module, and a resource allocation module. The outputs of the three sub-modules are combined to obtain the output of the final policy network.

[0052] The input to the unloading decision module is a shared feature vector. The output dimension is The unloading decision vector; the unloading decision module consists of an MLP with three fully connected layers and a decision module; the number of neurons in the first, second, and third layers of the MLP in the unloading decision module are respectively , and The first two layers use the ReLU activation function, and the last layer uses the Sigmoid activation function; the input to the unloading decision module... The dimension is obtained after three-layer MLP processing in the unloading decision module. Unloading decision probability vector This includes the offloading probability of each terminal device; then the decision module targets the vector... The first in element Calculate unloading decision variables , It is a step function; finally, construct the unloading decision vector. And output;

[0053] The input to the aggregation decision module is the shared feature vector. Output aggregation coefficient The aggregation decision module consists of a three-layer MLP, with the number of neurons in layers 1, 2, and 3 of the MLP being respectively... , and The first two layers use the ReLU activation function, and the third layer uses the Sigmoid activation function.

[0054] The input to the resource allocation module is a shared feature vector. The output dimension is CPU allocation vector The resource allocation module consists of a three-layer MLP and a Softmax layer connected in series. The number of neurons in the first, second, and third layers of the MLP are respectively... , and The first two layers use the ReLU activation function, while the third layer does not.

[0055] The edge gateway obtains the offloading decision vector based on the action vector. Polymerization coefficient CPU allocation vector Then, based on the unloading decision vector... The power terminal equipment within its coverage area is divided into a local computing group and an offloading group, and authentication instructions and authentication type codes are broadcast to all devices in the group respectively.

[0056] For all terminal devices within the computing group, upon receiving the instruction, they will invoke the onboard microcontroller to execute the elliptic curve signature algorithm and then upload the generated short signature to the edge gateway.

[0057] For all terminal devices in the unloading group, upon receiving the instruction, they will directly encapsulate the original data to be authenticated and upload it to the edge gateway.

[0058] While sending authentication commands to the terminal devices, the edge server calls the operating system interface to perform real-time slicing of the total system frequency and allocates CPU vectors accordingly. The CPU quotas for its signature container and aggregation container, i.e., the percentages of the total system frequency, are readjusted. and The signature container is specifically designed to process the raw data uploaded by terminal devices within the unloading group and generate corresponding signatures for it.

[0059] Signatures from both the terminal device and the signature container are placed into the verification queue of the aggregation container; when the number of signatures in the queue exceeds... At that time, the aggregation container will start a batch verification algorithm to process this... Each signature is aggregated and verified;

[0060] Finally, the edge gateway calculates and records the average authentication latency. Average normalized energy consumption and certification success rate Use the following formula to represent: ; ; ;in, This refers to the total time elapsed from when the edge gateway broadcasts and sends the authentication command until the aggregation container completes the authentication of all signatures; Is this it? The maximum average time taken for a terminal device to complete all signatures is preset based on historical observations and calculations. It refers to the first The amount of electricity consumed by a terminal device from receiving an instruction to sending the signature or raw data; This refers to the number of terminal devices that have been successfully authenticated.

[0061] The present invention also provides a system for implementing a multi-objective optimization-based encryption authentication task offloading method in the aforementioned edge-end collaborative power network, comprising a terminal device information acquisition module, an edge-end state vector fusion module, a policy network construction module, a policy network training module, and an encryption authentication task offloading module;

[0062] The terminal device information collection module collects information on power terminal devices within the coverage area of ​​the edge gateway, obtains the power terminal device status vector, and uploads the data to the edge status vector fusion module.

[0063] The edge state vector fusion module obtains the northbound link, CPU and buffer information of the edge gateway, integrates it with the received power terminal device state vector to obtain the edge state vector, and uploads the data to the encryption authentication task offloading module.

[0064] The policy network construction module builds an encryption authentication task offloading policy network based on a multilayer perceptron and uploads the data to the policy network training module.

[0065] The policy network training module obtains the historical state vector of the power terminal equipment to train the encryption authentication task offloading policy network, obtains the trained policy network, and uploads the data to the encryption authentication task offloading module.

[0066] The encryption authentication task offloading module inputs the edge state vector into the trained policy network to obtain the action vector; based on the action vector, it offloads the signature of the corresponding power terminal device to the gateway.

[0067] This invention discloses a method and system for offloading encryption authentication tasks based on multi-objective optimization in edge-coordinated power networks, aiming to achieve millisecond-level real-time authentication, multi-objective collaborative optimization, fast training iteration convergence speed, and stable measurement update amplitude. The method of this invention enables the three sub-tasks—offloading decision, aggregation coefficient, and CPU allocation—to share the underlying environmental representation by setting a shared feature extraction module in the first layer of the policy network. This reduces the number of network parameters, accelerates training and inference speed, and promotes information complementarity between different tasks. Furthermore, this method not only adapts to the needs of refined resource scheduling through continuous action space modeling (such as aggregation coefficient and CPU frequency allocation), but also adapts to binary offloading selection through discrete offloading decision modeling. It can be widely applied to edge computing scenarios with various heterogeneous device access and has strong generalization ability. Attached Figure Description

[0068] Figure 1 This is a schematic flowchart of the method of the present invention;

[0069] Figure 2 This is a schematic diagram of the strategy network structure of the present invention. Detailed Implementation

[0070] This invention provides a method for offloading encryption and authentication tasks based on multi-objective optimization in edge-to-edge collaborative power networks, the flowchart of which is shown below. Figure 1 As shown, it includes the following steps:

[0071] Collect information on power terminal equipment within the coverage area of ​​the edge gateway to obtain the power terminal equipment status vector;

[0072] The power terminal equipment state vector is obtained by the following method: obtaining the real-time task queue length, remaining battery power and computing power level of the power terminal equipment, and using the max-min normalization method to preprocess the data. Finally, the normalized feature values ​​are combined into a column vector and stacked vertically according to the device ID order to obtain the power terminal equipment state vector.

[0073] For edge gateways and their subordinates Each power terminal device and its edge gateway are configured with a fixed status sampling period. (e.g., 100ms), at each sampling time, the parsing from the first... Terminal devices (of which) From the uplink status messages of the task queue, the remaining battery power and computing power level can be extracted.

[0074] The data is preprocessed using the max-min normalization method, expressed by the following formula: ; ; ;in, For terminal devices Maximum queue length, For terminal devices The total battery capacity; The normalized feature value for the task queue length; The normalized characteristic value of the remaining battery capacity; To calculate the normalized characteristic value of the capability level;

[0075] Will The three normalized feature values ​​of each terminal device are calculated according to... The order is used to form a column vector, and... The row vectors are stacked vertically in order of terminal device ID, thus constructing a dimension of Terminal device state vector .

[0076] Task queue length This represents the amount of data tasks currently backed up in the local buffer of the terminal device, awaiting authentication or computation, in bits.

[0077] Remaining battery power This indicates the remaining percentage or specific charge level of the terminal device's battery relative to its fully charged state, expressed in milliampere-hours (mAh).

[0078] Computational ability level This indicates the CPU processing frequency or floating-point operations per second of the terminal device, reflecting its local task execution speed.

[0079] The edge gateway northbound link, CPU, and buffer information are obtained and integrated with the obtained power terminal device state vector to obtain the state vector;

[0080] The edge state vector is obtained using the following method: obtain the uplink bandwidth utilization, downlink bandwidth utilization, CPU utilization, and buffer utilization of the northbound link of the edge gateway, and append them to the terminal device state vector. Finally, we obtain the dimension as Edge state vector ;in, Edge state vector Express it using the following formula: Among them, the uplink bandwidth utilization rate of the northbound link and uplink bandwidth utilization Indicates the communication capacity between the edge gateway and the upstream network; CPU utilization. Indicates the busy level of the edge server's computing unit, with a numerical range of [value missing]. Buffer occupancy rate The storage space occupancy ratio of the task input queue on the edge side of the edge gateway reflects the current task backlog and queuing congestion status, with a value range of [value missing]. .

[0081] Construct and train the encryption authentication task offloading policy network, and input the state vector into the trained policy network to obtain the action vector;

[0082] The policy network to be trained includes a shared feature extraction module and an action output module; the edge state vector obtained as input to the shared feature extraction module is... The output is a dimension Shared feature vectors The action output module inputs a shared feature vector. The final output is the action vector; where, .

[0083] The shared feature extraction module is a multilayer perceptron, consisting of three layers, all using the ReLU activation function; the number of neurons in layers 1, 2, and 3 are respectively... , and .

[0084] The action output module includes three parallel sub-modules: an unloading decision module, an aggregation decision module, and a resource allocation module. The outputs of the three sub-modules are combined to obtain the output of the final policy network.

[0085] The input to the unloading decision module is a shared feature vector. The output dimension is The unloading decision vector; the unloading decision module consists of an MLP with four fully connected layers and a sampling module; the number of neurons in the 1st, 2nd, 3rd, and 4th layers of the MLP in the unloading decision module are respectively , , and The first three layers use the ReLU activation function, and the last layer uses the Sigmoid activation function; the input to the unloading decision module... The dimension is obtained after MLP processing in the unloading decision module. Unloading decision probability vector This includes the unloading probability of each terminal device; then the sampling module targets the vector... The first in element Constructing Bernoulli distribution The unloading decision variables were obtained by sampling using this distribution. Finally, construct the unloading decision vector. And output;

[0086] The input to the aggregation decision module is the shared feature vector. Output aggregation coefficient The aggregation decision module consists of a three-layer MLP, with the number of neurons in layers 1, 2, and 3 of the MLP being respectively... , and The first two layers use the ReLU activation function, while the third layer has no activation function; the shared feature vector is input to the aggregation decision module. The original polymerization coefficient was obtained after MLP processing. The aggregation coefficient is obtained after Gaussian noise reduction and activation. ;in, Gaussian noise was introduced during the training phase, and Follows a mean of 0 and a variance of normal distribution ;

[0087] The input to the resource allocation module is a shared feature vector. The output dimension is CPU allocation vector The resource allocation module consists of a three-layer MLP, a Softmax layer, and a sampling module connected in series. The number of neurons in the first, second, and third layers of the MLP are respectively... , and The first two layers use the ReLU activation function, while the third layer does not; they share feature vectors. The initial CPU allocation vector is obtained after processing by the MLP and Softmax layers. The sampling module is based on Calculate the distribution parameters .in, , Given a vector of length 3, where each element is a vector of length 3. Then, the sampling module constructs a Dirichlet distribution. And sample to obtain CPU allocation vector And output; where, The first concentration parameter of the Dirichlet distribution corresponds to the CPU initial allocation vector. The first component is used to determine the CPU allocation vector. The sampling distribution; The second concentration parameter represents the Dirichlet distribution and corresponds to the CPU initial allocation vector. The second component is used to determine the CPU allocation vector. The sampling distribution; The third concentration parameter represents the Dirichlet distribution and corresponds to the CPU initial allocation vector. The third component is used to determine the CPU allocation vector. The sampling distribution;

[0088] Using probability distribution functions This indicates the policy network. Indicates input Time output The probability, parameter Let the periodic index be a vector consisting of all the parameters to be learned in the policy network. Iterative index The training process of the policy network includes:

[0089] A. Obtain historical information of power terminal equipment within the coverage area of ​​the edge gateway to obtain the historical state vector of the power terminal equipment;

[0090] B. Obtain historical edge gateway northbound link, CPU, and buffer information, and integrate them with the corresponding historical state vector of the power terminal equipment to obtain the historical state vector;

[0091] C. The historical state vector is used as input to the policy network, and the terminal device authentication task is executed based on the obtained action vector. Next, obtain 1 line of length. trajectory ;in, It is a vector, represented using the following formula: ; For the first The edge state vector obtained this time; For the first The action vector obtained this time; It is a three-dimensional vector, specifically in the form of , For the first The average authentication latency obtained this time , For the first The average energy consumption obtained this time , For the first The success rate of the first authentication ;

[0092] D. Obtain the initial values ​​of the policy parameters through Gaussian initialization. Simultaneously, initialize relevant metadata parameters; among which, set the batch size. and trajectory length Learning rate Total return dimension Number of iterations per cycle Total number of cycles Gradient cutoff radius Reward Discount Rate ;

[0093] E. Based on historical data, repeatedly analyze the policy network. Get from The length of the strip is trajectory and collect them to obtain a set ;

[0094] F. Based on trajectory sets The trajectory is estimated using the following function. Total return: Then estimate the current total return of the trajectory set. ;

[0095] G. Project the current total return using the function. Head-guided hyperrectangular feasible region In this process, a feasible total return is obtained. ; where, projection function Express it using the following formula: ; Hyperrectangular feasible region Express it using the following formula: Total feasible return Express it using the following formula: ;in, The feasible range for the corresponding certification delay target reward component. The feasible range for the corresponding energy consumption target return component. The feasible range for the corresponding certification success rate target return component;

[0096] H. Based on historical data, repeat the process from the policy network. Get from The length of the strip is trajectory and collect them to obtain a set ;

[0097] I. Based on trajectory sets Estimate the gradient of the current policy It can be expressed using the following formula: ; where, function Used for trajectory estimation The policy gradient for total return is expressed using the following formula: ; where, convex function Used to merge total returns The three components are defined as follows: ;in, Total return vector The reward component on the certification latency target; Total return vector The return on energy consumption targets; Total return vector The reward component in the certification success rate target;

[0098] J. Based on the current policy gradient Update the strategy parameters to obtain It can be expressed using the following formula: ; where, function Indicates the use of policy gradient Update strategy parameters The result is defined using the following formula: ;

[0099] K. Based on historical data, repeat the process from the policy network. Get from The length of the strip is trajectory and collect them to obtain a set ;

[0100] L. Based on trajectory sets Estimate current total return Then project it onto the hyperrectangular feasible region. In this process, a feasible total return is obtained. ; of which, total return The estimation formula is: ;function Indicates based on sample The estimated policy network Policy gradient, sample From policy network The sampled data is represented by the following formula: ;in, The ratio of importance sampling is expressed using the following formula: ;

[0101] M. Based on historical data, repeat the process from the policy network. Get from The length of the strip is trajectory and collect them to obtain a set ;

[0102] N. Based on trajectory sets Estimate the gradient of the current policy It can be expressed using the following formula: ;function Indicates the use of trajectory Policy network Policy gradient, where trajectory From policy network The sampled data is represented by the following formula: ;

[0103] O. Based on the current gradient Update the strategy parameters to obtain ;

[0104] P. Iterative Index Increase by 1, and then determine If the condition is true, return to step K and continue the iteration; otherwise, obtain the updated condition. , and ;

[0105] Q. Set the initial values ​​for the strategy parameters in the next cycle. , will the cycle Increase by 1, and then determine Check if the condition is met; if it is met, return to step E; if it is not met, end the training algorithm and obtain the final policy parameters. .

[0106] The trained policy network includes a shared feature extraction module and an action output module; the shared feature extraction module takes as input the edge state vector. The output is a dimension Shared feature vectors The action output module inputs a shared feature vector. The final output is the action vector; where, .

[0107] The shared feature extraction module is a multilayer perceptron, consisting of three layers, all using the ReLU activation function; the number of neurons in layers 1, 2, and 3 are respectively... , and .

[0108] The action output module includes three parallel sub-modules: an unloading decision module, an aggregation decision module, and a resource allocation module. The outputs of the three sub-modules are combined to obtain the output of the final policy network.

[0109] The input to the unloading decision module is a shared feature vector. The output dimension is The unloading decision vector; the unloading decision module consists of an MLP with three fully connected layers and a decision module; the number of neurons in the first, second, and third layers of the MLP in the unloading decision module are respectively , and The first two layers use the ReLU activation function, and the last layer uses the Sigmoid activation function; the input to the unloading decision module... The dimension is obtained after three-layer MLP processing in the unloading decision module. Unloading decision probability vector This includes the offloading probability of each terminal device; then the decision module targets the vector... The first in element Calculate unloading decision variables , It is a step function; finally, construct the unloading decision vector. And output;

[0110] The input to the aggregation decision module is the shared feature vector. Output aggregation coefficient The aggregation decision module consists of a three-layer MLP, with the number of neurons in layers 1, 2, and 3 of the MLP being respectively... , and The first two layers use the ReLU activation function, and the third layer uses the Sigmoid activation function.

[0111] The input to the resource allocation module is a shared feature vector. The output dimension is CPU allocation vector The resource allocation module consists of a three-layer MLP and a Softmax layer connected in series. The number of neurons in the first, second, and third layers of the MLP are respectively... , and The first two layers use the ReLU activation function, while the third layer does not.

[0112] The structural diagram of the policy network is shown below. Figure 2 As shown, the dashed line represents the data flow and structure of the policy network during training only.

[0113] Based on the action vector, the signature of the corresponding power terminal device is offloaded to the gateway to achieve aggregate authentication.

[0114] The edge gateway obtains the offloading decision vector based on the action vector. Polymerization coefficient CPU allocation vector Then, based on the unloading decision vector... The power terminal equipment within its coverage area is divided into a local computing group and an offloading group, and authentication instructions and authentication type codes are broadcast to all devices in the group respectively.

[0115] For all terminal devices within the computing group, upon receiving the instruction, they will invoke the onboard microcontroller to execute the elliptic curve signature algorithm and then upload the generated short signature to the edge gateway.

[0116] For all terminal devices in the unloading group, upon receiving the instruction, they will directly encapsulate the original data to be authenticated and upload it to the edge gateway.

[0117] While sending authentication commands to the terminal devices, the edge server calls the operating system interface to perform real-time slicing of the total system frequency and allocates CPU vectors accordingly. The CPU quotas for its signature container and aggregation container, i.e., the percentages of the total system frequency, are readjusted. and The signature container is specifically designed to process the raw data uploaded by terminal devices within the unloading group and generate corresponding signatures for it.

[0118] Signatures from both the terminal device and the signature container are placed into the verification queue of the aggregation container; when the number of signatures in the queue exceeds... At that time, the aggregation container will start a batch verification algorithm to process this... Each signature is aggregated and verified;

[0119] Finally, the edge gateway calculates and records the average authentication latency. Average normalized energy consumption and certification success rate Use the following formula to represent: ; ; ;in, This refers to the total time elapsed from when the edge gateway broadcasts and sends the authentication command until the aggregation container completes the authentication of all signatures; Is this it? The maximum average time taken for a terminal device to complete all signatures is preset based on historical observations and calculations. It refers to the first The amount of electricity consumed by a terminal device from receiving an instruction to sending the signature or raw data; This refers to the number of terminal devices that have been successfully authenticated.

[0120] The present invention also provides a system for implementing a multi-objective optimization-based encryption authentication task offloading method in the aforementioned edge-end collaborative power network, comprising a terminal device information acquisition module, an edge-end state vector fusion module, a policy network construction module, a policy network training module, and an encryption authentication task offloading module;

[0121] The terminal device information collection module collects information on power terminal devices within the coverage area of ​​the edge gateway, obtains the power terminal device status vector, and uploads the data to the edge status vector fusion module.

[0122] The edge state vector fusion module obtains the northbound link, CPU and buffer information of the edge gateway, integrates it with the received power terminal device state vector to obtain the edge state vector, and uploads the data to the encryption authentication task offloading module.

[0123] The policy network construction module builds an encryption authentication task offloading policy network based on a multilayer perceptron and uploads the data to the policy network training module.

[0124] The policy network training module obtains the historical state vector of the power terminal equipment to train the encryption authentication task offloading policy network, obtains the trained policy network, and uploads the data to the encryption authentication task offloading module.

[0125] The encryption authentication task offloading module inputs the edge state vector into the trained policy network to obtain the action vector; based on the action vector, it offloads the signature of the corresponding power terminal device to the gateway.

Claims

1. A method for offloading encryption and authentication tasks based on multi-objective optimization in an edge-to-edge collaborative power network, characterized in that, Includes the following steps: Collect information on power terminal equipment within the coverage area of ​​the edge gateway to obtain the power terminal equipment status vector; Obtain the northbound link, CPU, and buffer information of the edge gateway, and integrate them with the obtained power terminal device state vector to obtain the edge state vector; A network for offloading encryption authentication tasks is constructed based on a multilayer perceptron. The network includes a shared feature extraction module and an action output module. The shared feature extraction module takes the edge state vector as input and outputs a shared feature vector. The action output module takes a shared feature vector as input and outputs an action vector. The historical state vectors of power terminal equipment are obtained to train the encryption authentication task offloading policy network. Edge state vectors are then input into the trained policy network to obtain action vectors; these action vectors include offloading decision vectors. Polymerization coefficient CPU allocation vector ; Based on the action vector, the signature of the corresponding power terminal device is offloaded to the gateway.

2. The encryption authentication task offloading method based on multi-objective optimization in an edge-to-edge collaborative power network according to claim 1, characterized in that, The state vector of the power terminal equipment is obtained by the following method: obtaining the real-time task queue length, remaining battery power and computing power level of the power terminal equipment, and using the max-min normalization method to preprocess the data. Finally, the normalized feature values ​​are combined into a column vector and stacked vertically according to the device ID order to obtain the state vector of the power terminal equipment.

3. The encryption authentication task offloading method based on multi-objective optimization in edge-to-edge collaborative power networks according to claim 2, characterized in that, For edge gateways and their subordinates Each power terminal device and its edge gateway are configured with a fixed status sampling period. At each sampling time, the parsing is performed on the first... From the uplink status messages of each terminal device, extract the real-time task queue length, remaining battery power, and computing power level; task queue length This indicates the amount of data tasks currently backed up in the local buffer of the terminal device, awaiting authentication or computation, in bits; remaining battery power. This indicates the remaining percentage or specific charge level of the terminal device's battery relative to its full charge state, expressed in milliampere-hours (mAh); computing power level. This indicates the CPU processing frequency or floating-point operations per second of the terminal device, reflecting its local task execution rate; The data is preprocessed using the max-min normalization method, expressed by the following formula: ; ; ;in, For terminal devices Maximum queue length, For terminal devices The total battery capacity; The normalized feature value for the task queue length; The normalized characteristic value of the remaining battery capacity; To calculate the normalized characteristic value of the capability level; Will The three normalized feature values ​​of each terminal device are calculated according to... The order is used to form a column vector, and... The row vectors are stacked vertically in order of terminal device ID, thus constructing a dimension of Terminal device state vector .

4. The encryption and authentication task offloading method based on multi-objective optimization in an edge-to-edge collaborative power network according to claim 1, characterized in that, The policy network to be trained includes a shared feature extraction module and an action output module; The edge state vector obtained from the input of the shared feature extraction module The output is a dimension Shared feature vectors ; Action output module input shared feature vector The final output is the action vector; where, ; The shared feature extraction module is a multilayer perceptron, consisting of three layers, all using the ReLU activation function; the number of neurons in layers 1, 2, and 3 are respectively... , and ; The action output module includes three parallel sub-modules: an unloading decision module, an aggregation decision module, and a resource allocation module. The outputs of the three sub-modules are combined to obtain the output of the final policy network. The input to the unloading decision module is a shared feature vector. The output dimension is The unloading decision vector; the unloading decision module consists of an MLP with four fully connected layers and a sampling module; the number of neurons in the 1st, 2nd, 3rd, and 4th layers of the MLP in the unloading decision module are respectively , , and The first three layers use the ReLU activation function, and the last layer uses the Sigmoid activation function; the input to the unloading decision module... The dimension is obtained after MLP processing in the unloading decision module. Unloading decision probability vector This includes the unloading probability of each terminal device; then the sampling module targets the vector... The first in element Constructing Bernoulli distribution The unloading decision variables were obtained by sampling using this distribution. Finally, construct the unloading decision vector. And output; The input to the aggregation decision module is the shared feature vector. Output aggregation coefficient The aggregation decision module consists of a three-layer MLP, with the number of neurons in layers 1, 2, and 3 of the MLP being respectively... , and The first two layers use the ReLU activation function, while the third layer has no activation function; the shared feature vector is input to the aggregation decision module. The original polymerization coefficient was obtained after MLP processing. The aggregation coefficient is obtained after Gaussian noise reduction and activation. ; in, Gaussian noise was introduced during the training phase, and Follows a mean of 0 and a variance of normal distribution ; The input to the resource allocation module is a shared feature vector. The output dimension is CPU allocation vector The resource allocation module consists of a three-layer MLP, a Softmax layer, and a sampling module connected in series. The number of neurons in the first, second, and third layers of the MLP are respectively... , and The first two layers use the ReLU activation function, while the third layer does not; they share feature vectors. The initial CPU allocation vector is obtained after processing by the MLP and Softmax layers. The sampling module is based on Calculate the distribution parameters .in, , Given a vector of length 3, where each element is a vector of length 3. Then, the sampling module constructs a Dirichlet distribution. And sample to obtain CPU allocation vector And output; where, The first concentration parameter of the Dirichlet distribution corresponds to the CPU initial allocation vector. The first component is used to determine the CPU allocation vector. The sampling distribution; The second concentration parameter represents the Dirichlet distribution and corresponds to the CPU initial allocation vector. The second component is used to determine the CPU allocation vector. The sampling distribution; The third concentration parameter represents the Dirichlet distribution and corresponds to the CPU initial allocation vector. The third component is used to determine the CPU allocation vector. The sampling distribution.

5. The encryption and authentication task offloading method based on multi-objective optimization in an edge-to-edge collaborative power network according to claim 4, characterized in that, Using probability distribution functions This indicates the policy network. Indicates input Time output The probability, parameter Let the periodic index be a vector consisting of all the parameters to be learned in the policy network. Iterative index The training process of the policy network includes: A. Obtain historical information of power terminal equipment within the coverage area of ​​the edge gateway to obtain the historical state vector of the power terminal equipment; B. Obtain historical edge gateway northbound link, CPU, and buffer information, and integrate them with the corresponding historical state vector of the power terminal equipment to obtain the historical state vector; C. The historical state vector is used as input to the policy network, and the terminal device authentication task is executed based on the obtained action vector. Next, obtain 1 line of length. trajectory ;in, It is a vector, represented using the following formula: ; For the first The edge state vector obtained this time; For the first The action vector obtained this time; It is a three-dimensional vector, specifically in the form of , For the first The average authentication latency obtained this time , For the first The average energy consumption obtained this time , For the first The success rate of the first authentication ; D. Obtain the initial values ​​of the policy parameters through Gaussian initialization. Simultaneously, initialize relevant metadata parameters; among which, set the batch size. and trajectory length Learning rate Total return dimension Number of iterations per cycle Total number of cycles Gradient cutoff radius Reward Discount Rate ; E. Based on historical data, repeatedly analyze the policy network. Get from The length of the strip is trajectory and collect them to obtain a set ; F. Based on trajectory sets The trajectory is estimated using the following function. Total return: Then estimate the current total return of the trajectory set. ; G. Project the current total return using the function. Head-guided hyperrectangular feasible region In this process, a feasible total return is obtained. ; where, projection function Express it using the following formula: ; Hyperrectangular feasible region Express it using the following formula: Total feasible return Express it using the following formula: ;in, The feasible range for the corresponding certification delay target reward component. The feasible range for the corresponding energy consumption target return component. The feasible range for the corresponding certification success rate target return component; H. Based on historical data, repeat the process from the policy network. Get from The length of the strip is trajectory and collect them to obtain a set ; I. Based on trajectory sets Estimate the gradient of the current policy It can be expressed using the following formula: ; where, function Used for trajectory estimation The policy gradient for total return is expressed using the following formula: ; where, convex function Used to merge total returns The three components are defined as follows: ;in, Total return vector The reward component on the certification latency target; Total return vector The return on energy consumption targets; Total return vector The reward component in the certification success rate target; J. Based on the current policy gradient Update the strategy parameters to obtain It can be expressed using the following formula: ; where, function Indicates the use of policy gradient Update strategy parameters The result is defined using the following formula: ; K. Based on historical data, repeat the process from the policy network. Get from The length of the strip is trajectory and collect them to obtain a set ; L. Based on trajectory sets Estimate current total return Then project it onto the hyperrectangular feasible region. In this process, a feasible total return is obtained. ; of which, total return The estimation formula is: ;function Indicates based on sample The estimated policy network Policy gradient, sample From policy network The sampled data is represented by the following formula: ;in, The ratio of importance sampling is expressed using the following formula: ; M. Based on historical data, repeat the process from the policy network. Get from The length of the strip is trajectory and collect them to obtain a set ; N. Based on trajectory sets Estimate the gradient of the current policy It can be expressed using the following formula: ;function Indicates the use of trajectory Policy network Policy gradient, where trajectory From policy network The sampled data is represented by the following formula: ; O. Based on the current gradient Update the strategy parameters to obtain ; P. Iterative Index Increase by 1, and then determine If the condition is true, return to step K and continue the iteration; otherwise, obtain the updated condition. , and ; Q. Set the initial values ​​for the strategy parameters in the next cycle. , will the cycle Increase by 1, and then determine Check if the condition is met; if it is met, return to step E; if it is not met, end the training algorithm and obtain the final policy parameters. .

6. The encryption and authentication task offloading method based on multi-objective optimization in an edge-to-edge collaborative power network according to claim 1, characterized in that, The trained policy network includes a shared feature extraction module and an action output module; The shared feature extraction module inputs the edge state vector. The output is a dimension Shared feature vectors The action output module inputs a shared feature vector. The final output is the action vector; where, ; The shared feature extraction module is a multilayer perceptron, consisting of three layers, all using the ReLU activation function; the number of neurons in layers 1, 2, and 3 are respectively... , and ; The action output module includes three parallel sub-modules: an unloading decision module, an aggregation decision module, and a resource allocation module. The outputs of the three sub-modules are combined to obtain the output of the final policy network. The input to the unloading decision module is a shared feature vector. The output dimension is The unloading decision vector; the unloading decision module consists of an MLP with three fully connected layers and a decision module; the number of neurons in the first, second, and third layers of the MLP in the unloading decision module are respectively , and The first two layers use the ReLU activation function, and the last layer uses the Sigmoid activation function; the input to the unloading decision module... The dimension is obtained after three-layer MLP processing in the unloading decision module. Unloading decision probability vector This includes the offloading probability of each terminal device; then the decision module targets the vector... The first in element Calculate unloading decision variables , It is a step function; finally, construct the unloading decision vector. And output; The input to the aggregation decision module is the shared feature vector. Output aggregation coefficient The aggregation decision module consists of a three-layer MLP, with the number of neurons in layers 1, 2, and 3 of the MLP being respectively... , and The first two layers use the ReLU activation function, and the third layer uses the Sigmoid activation function. The input to the resource allocation module is a shared feature vector. The output dimension is CPU allocation vector The resource allocation module consists of a three-layer MLP and a Softmax layer connected in series. The number of neurons in the first, second, and third layers of the MLP are respectively... , and The first two layers use the ReLU activation function, while the third layer does not.

7. The encryption authentication task offloading method based on multi-objective optimization in an edge-to-edge collaborative power network according to claim 1, characterized in that, The edge gateway obtains the offloading decision vector based on the action vector. Polymerization coefficient CPU allocation vector ; Then based on the unloading decision vector The power terminal equipment within its coverage area is divided into a local computing group and an offloading group, and authentication instructions and authentication type codes are broadcast to all devices in the group respectively. For all terminal devices within the computing group, upon receiving the instruction, they will invoke the onboard microcontroller to execute the elliptic curve signature algorithm and then upload the generated short signature to the edge gateway. For all terminal devices within the unloading group, upon receiving the instruction, they will directly encapsulate the original data to be authenticated and upload it to the edge gateway.

8. The encryption and authentication task offloading method based on multi-objective optimization in an edge-to-edge collaborative power network according to claim 7, characterized in that, While sending authentication commands to the terminal devices, the edge server calls the operating system interface to perform real-time slicing of the total system frequency and allocates CPU vectors accordingly. The CPU quotas for its signature container and aggregation container, i.e., the percentages of the total system frequency, are readjusted. and The signature container is specifically designed to process the raw data uploaded by terminal devices within the unloading group and generate corresponding signatures for it. Signatures from both the terminal device and the signature container are placed into the verification queue of the aggregation container; when the number of signatures in the queue exceeds... At that time, the aggregation container will start a batch verification algorithm to process this... Each signature is aggregated and verified; Finally, the edge gateway calculates and records the average authentication latency. Average normalized energy consumption and certification success rate .

9. The encryption authentication task offloading method based on multi-objective optimization in an edge-to-edge collaborative power network according to claim 8, characterized in that, Average authentication latency Express it using the following formula: Average normalized energy consumption Express it using the following formula: Authentication success rate Express it using the following formula: ;in, This refers to the total time elapsed from when the edge gateway broadcasts and sends the authentication command until the aggregation container completes the authentication of all signatures; Is this it? The maximum average time taken for a terminal device to complete all signatures is preset based on historical observations and calculations. It refers to the first The amount of electricity consumed by a terminal device from receiving an instruction to sending the signature or raw data; This refers to the number of terminal devices that have been successfully authenticated.

10. A system for implementing the multi-objective optimization-based encryption authentication task offloading method in an edge-to-edge cooperative power network as described in any one of claims 1 to 9, characterized in that, The system includes a terminal device information acquisition module, an edge-end state vector fusion module, a policy network construction module, a policy network training module, and an encryption authentication task unloading module. The terminal device information collection module collects information on power terminal devices within the coverage area of ​​the edge gateway, obtains the power terminal device status vector, and uploads the data to the edge status vector fusion module. The edge state vector fusion module obtains the northbound link, CPU and buffer information of the edge gateway, integrates it with the received power terminal device state vector to obtain the edge state vector, and uploads the data to the encryption authentication task offloading module. The policy network construction module builds an encryption authentication task offloading policy network based on a multilayer perceptron and uploads the data to the policy network training module. The policy network training module obtains the historical state vector of the power terminal equipment to train the encryption authentication task offloading policy network, obtains the trained policy network, and uploads the data to the encryption authentication task offloading module. The encryption authentication task offloading module inputs the edge state vector into the trained policy network to obtain the action vector; based on the action vector, it offloads the signature of the corresponding power terminal device to the gateway.

Images