MEC stable computing migration system based on security and energy consumption perception

Abstract

The invention discloses an MEC stable computing migration system based on safety and energy consumption perception. The MEC stable computing migration system comprises the steps that S1, a multi-time-slice constraint and cost optimization problem is decomposed into single time slices through Lyapunov optimization; s2, taking the data arrival scale and the channel gain in each time slot as the input of an act-critic model, and generating a plurality of candidate unloading actions; and S3, the risk probability constraint of the current time slot is satisfied, and the security service with the lowest time and energy consumption cost is selected. And S4, converting the transmission time distribution of the equipment which is determined to be unloaded into a knapsack problem for solving. And S5, selecting an unloading action with an optimal target value, and matching the unloading action with environment input to serve as a training sample. And S6, along with the continuous generation of the model, selecting, updating and outputting the optimal unloading strategy of each device in multiple time slots. The invention provides a security and energy consumption perception stable computation unloading (LySESO) scheme based on a Lyapunov framework in an MEC and an act-critic model in a DRL, and is used for solving the computation unloading problem with a dynamic MEC environment and security threats, so that the total energy consumption of a system is reduced while the stability of a task backlog queue is ensured. According to the method, openness and dynamism of the edge cloud environment are considered, long-term task queue stability and safety requirements are guaranteed, the total energy consumption of a task unloading system is reduced, and the method has excellent commercial value.

Description

Technical Field

[0001] The present invention relates to edge computing in the field of cloud computing in computer technology, and more specifically to a MEC stable computing migration system based on security and energy consumption perception. Background Art

[0002] With the rapid development of the Internet of Things and the continuous upgrading of mobile communication technology, mobile Internet has completely entered people's lives. The number of smart mobile devices (such as smartphones, tablets, smart watches, etc.) has grown exponentially, and computing-intensive applications (such as autonomous driving, virtual reality, augmented reality, real-time online games, telemedicine, etc.) have emerged rapidly. Although processor technology continues to advance, the battery capacity and computing power of wireless mobile devices are still bottlenecks, and the quality of service (QoS) of computing-intensive and delay-sensitive tasks cannot be guaranteed.

[0003] In order to solve the problem of long round-trip communication links between mobile devices and cloud servers in traditional mobile cloud computing (MCC), mobile edge computing (MEC) is a new computing paradigm that adds edge servers in addition to cloud servers, allocating remote cloud resources to edge servers near mobile users, enabling mobile users to offload part or all of the computing tasks of mobile applications to edge servers for collaborative execution, greatly alleviating the conflict between resource supply and demand, and effectively reducing application completion time and energy consumption of mobile devices.

[0004] In mobile edge computing (MEC), computing task offloading also faces some challenges: First, the dynamic nature of the edge cloud environment, such as the channel quality of the edge server and the workload of the edge server are constantly changing, which will affect the task offloading decision on the mobile device. Choosing the right offloading decision for the mobile device can not only reduce the task computing time and ensure the stability of the task queue, but also reduce the system energy consumption. Second, the data security issue during the task offloading process. Due to the openness of the edge cloud environment, the task offloading process from the mobile device to the edge server will be subject to malicious attacks such as data leakage and tampering, which poses a serious security threat to the successful execution of these offloaded tasks. Therefore, it is necessary to select appropriate security services to effectively defend against hostile attacks and protect the data of these offloaded tasks. Summary of the invention

[0005] The present invention aims to solve the problem of minimizing the energy consumption cost of mobile edge computing network computing migration under the constraints of task queue stability and offload data security.

[0006] The present invention adopts the following technical solutions:

[0007] According to the mobile edge computing (MEC) network information, the goal of minimizing energy consumption cost under the constraints of queue stability and security is constructed and solved. The overall process of the algorithm is as follows:Figure 1 As shown in the application scenario diagram Figure 2 As shown, the method architecture diagram is as follows Figure 3 shown.

[0008] The security threats and task processing between each mobile device and edge server in the MEC network are modeled as follows: Figure 2 The following are the main steps of the algorithm:

[0009] S1. For the problem of minimizing multi-time-slot energy consumption and stable computational migration of task queues under safety constraints, Lyapunov optimization is used to decompose multi-time-slot optimization into a single-time-slot problem.

[0010] S2. Initialize the DNN network parameters in the actor-critic model, take the data arrival scale and channel gain in each time slot as the input of the actor-critic model, and the DNN network generates multiple candidate offloading actions.

[0011] S3. For the optional confidentiality and integrity algorithms, calculate the risk probability and cost respectively. Under the risk probability constraint of the current time slot, select the security service with the lowest time and energy consumption cost.

[0012] S4. After determining the security service, the time cost of the security service is removed from each time slice, and the time allocation for the devices that have been determined to be unloaded in TDMA mode is converted into a knapsack problem for solution.

[0013] S5. Calculate the optimal target value after the unloading strategy is determined in the candidate unloading action group. Select the unloading action with the best target value and pair it with the environmental input as a training sample for the DNN network.

[0014] S6. As the model is continuously generated, selected, and updated, the optimal offloading strategy for each device in multiple time slots is output.

[0015] Preferably, Lyapunov optimization in S1 is often used in the control of dynamic systems with queuing networks, and is used to solve the random optimization problem of dynamic changes in system variables in each time slice, thereby optimizing system performance while ensuring the stability of the network queue.

[0016] Preferably, the actor-critic model in S2 is an algorithm that combines the ideas of policy base and value base. Actor is implemented using the policy gradient algorithm, and Critic is implemented using temporal difference. Actor is the policy function π θ (s), usually implemented as a neural network, the input is the current state, and the output is an action. The training goal of the network is to maximize the expected cumulative return.

[0017] Preferably, in S4, TDMA is a communication technology for realizing shared transmission medium (generally in the radio field) or network. It allows multiple users to use the same frequency in different time slices (time slots). Under the conditions of timing and synchronization, the base station can receive the signals of each mobile terminal in each time slot without interference.

[0018] Compared with the prior art, the present invention has the following advantages:

[0019] The present invention is a MEC stable computing migration system based on security and energy consumption perception. It takes into account the openness and dynamics of the edge cloud environment, adds the security issue of offloading task data into the computing migration strategy problem, and makes a time-varying optimization strategy for the random and dynamic changes of environmental variables such as channel status and data arrival in the MEC network.

[0020] Compared with the prior art, the present invention considers the long-term data queue stability constraints and system performance, and models the optimization problem as a time-varying MINLP problem in a multi-time-slice environment. Assuming that the random channel conditions and data arrival in the future are unknown, the Lyapunov framework is used to decouple it into a single-time-slice optimization problem. Through the actor-critic model in DRL, the optimal unloading and safety decisions in each time slice are trained. It has the advantages of low algorithm complexity, long-term performance guarantee of task unloading, and time-varying task unloading strategies for each time slot. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is the overall flow chart of the invention

[0022] Figure 2 This is the application scenario diagram of the invention

[0023] Figure 3 This is the architecture diagram of the invented method DETAILED DESCRIPTION

[0024] The present invention is fully described and illustrated below in conjunction with the accompanying drawings in the embodiments of the present invention, and this embodiment does not limit the present invention.

[0025] Consider a MEC system consisting of an edge server ES and N wireless mobile devices WDs within the communication coverage of the server in a continuous time T. The time period T is divided into T time slices of equal size, and the length of each time slice is represented by ΔT = 1. The model adopts a binary offloading rule. In the time slice t, it is determined that the tasks on the WDs are either all offloaded to the ES for execution or all calculated locally. Specifically, the offloaded WDs share a common bandwidth W and transmit the task data to the ES in a TDMA manner to avoid mutual interference between the communications of different WDs and ES.

[0026] In the tth time slice, is the task data arriving at the WD i data queue. Assume that the arriving data Follow the independent and identically distributed (iid) with bounded second-order moments, Assume that η is estimated from past observations i The value of is known. is the channel gain between WD i and ES. Under the block fading assumption, Remains constant within a time slice, but varies independently between time slices.

[0027] The dynamic task queue of WDs is defined as:

[0028] In time slice t, the transmission rate between WD i and ES is calculated as follows:

[0029]

[0030] The risk probability of data being leaked or tampered after using algorithm j in security service type is:

[0031]

[0032] When the task i Assigned to instance V l The reliability expression formula during execution is as follows:

[0033] Different security protections will result in different security costs. During the WD i task offloading process, the time and energy consumption cost calculation formula is as follows:

[0034]

[0035]

[0036] When WD i decides to process data locally within time slice t Express the local CPU frequency as In the local computing mode, the raw data (bit), computing rate, and energy consumption processed by WD i in time slice t are:

[0037]

[0038]

[0039]

[0040] When WD i decides to offload task data to ES computing within time slice t Security service and offload transfer time limits are:

[0041]

[0042] When WD i decides to unload a task within time slice t The processing task volume and energy consumption are:

[0043]

[0044]

[0045] According to the established model, the optimization objectives and constraints are constructed as follows:

[0046] Optimization goal:

[0047] Constraints:

[0048] ① The amount of local computing and offloading tasks should be less than the task queue length requirement:

[0049] ② The sum of the security service and unloading transmission time should be less than the time slot length requirement:

[0050] ③The task queue should maintain a long-term stable state:

[0051] ④ The task transmission risk of the mobile device to be offloaded should be less than the risk threshold requirement of the current time slot:

[0052] A MEC stable computing migration system based on security and energy consumption awareness realizes the task unloading with optimal energy consumption cost under the constraints of task queue stability and security. Figure 1 As shown, the following steps are included:

[0053] S1. For the problem of minimizing multi-time-slot energy consumption and stable computational migration of task queues under safety constraints, Lyapunov optimization is used to decompose multi-time-slot optimization into a single-time-slot problem.

[0054] S2. Initialize the DNN network parameters in the actor-critic model, take the data arrival scale and channel gain in each time slot as the input of the actor-critic model, and the DNN network generates multiple candidate offloading actions.

[0055] S3. For the optional confidentiality and integrity algorithms, calculate the risk probability and cost respectively. Under the risk probability constraint of the current time slot, select the security service with the lowest time and energy consumption cost.

[0056] S4. After determining the security service, the time cost of the security service is removed from each time slice, and the time allocation for the devices that have been determined to be unloaded in TDMA mode is converted into a knapsack problem for solution.

[0057] S5. Calculate the optimal target value after the unloading strategy is determined in the candidate unloading action group. Select the unloading action with the best target value and pair it with the environmental input as a training sample for the DNN network.

[0058] S6. As the model is continuously generated, selected, and updated, the optimal offloading strategy for each device in multiple time slots is output.

[0059] In S1, the Lyapunov optimization theory optimizes the target under long-term constraints by taking the weighted sum Δ(Θ t )+Vy t As new optimization goals, the system adjusts the importance of them through weight V.

[0060] In S2, the actor-critic model combines the algorithm of policy base and value base. The actor is implemented using the policy gradient algorithm, and the critic is implemented using temporal difference. The actor is the policy function π θ (s), usually implemented with a neural network, the input is the current state, and the output is an action. The training goal of the network is to maximize the expected cumulative return. The critic is the value function V π (s), the network can estimate the value function of the current strategy, that is, it can evaluate the quality of the Actor's strategy function.

[0061] In S3, encryption services mainly include IDEA, DES, Blowfish, AES and RC4 encryption algorithms, and integrity services mainly include TIGER, RipeMD160, SHA-1, RipeMD128 and MD5 hash functions. During task offloading, different hash algorithms with different security levels can be flexibly selected to protect data from tampering while minimizing the time and energy consumption cost of integrity services. When WDs perform computational offloading in the MEC network, they can flexibly select encryption and hash algorithms with different security levels to form an integrated security protection to prevent security threats.

[0062] The transmission time allocation of the unloading tasks of each device in each time slice in S4 is converted into the solution of the backpack problem. The "objects" are sorted in descending order according to their "value" k, and as many items with high "value" as possible are loaded into the backpack to make the total "value" in the backpack the highest, until the backpack "capacity" is 0.

[0063] In S5, the security service selection and transmission time allocation algorithms in the critic model are used to calculate the optimal target values ​​of the candidate offloading action groups respectively, and then the optimal offloading action group is selected.

[0064] In S6, (β t ,x t ) as the input-output samples for updating the DNN network parameters in the actor module, set a replay memory that can store up to m data samples, and start training the DNN when the number of samples collected in the replay memory exceeds m / 2. In addition, at every Δ t A batch of data samples will be randomly selected every 10 seconds to train the DNN regularly to avoid overfitting of the model. Using the data samples, the parameters of the DNN are updated by minimizing the average cross entropy loss function. The cross entropy loss function can be expressed as:

[0065]

[0066] The above is a preferred implementation of the present invention, but the protection scope of the present invention is not limited to this. The drawings and implementation expansions involved in this article directly or indirectly applied to related fields should be covered in the protection scope of the present invention, and the protection scope should be based on the protection scope of the claims.

Claims

1. A MEC stable computing migration system based on security and energy consumption awareness realizes task unloading with optimal energy consumption cost under task queue stability and security constraints, mainly including the following steps: S1. For the problem of minimizing multi-time-slot energy consumption and stable computational migration of task queues under safety constraints, Lyapunov optimization is used to decompose multi-time-slot optimization into a single-time-slot problem. S2. Initialize the DNN network parameters in the actor-critic model, take the data arrival scale and channel gain in each time slot as the input of the actor-critic model, and the DNN network generates multiple candidate offloading actions. S3. For the optional confidentiality and integrity algorithms, calculate the risk probability and cost respectively. Under the risk probability constraint of the current time slot, select the security service with the lowest time and energy consumption cost. S4. After determining the security service, the time cost of the security service is removed from each time slice, and the time allocation for the devices that have been determined to be unloaded in TDMA mode is converted into a knapsack problem for solution. S5. Calculate the optimal target value after the unloading strategy is determined in the candidate unloading action group. Select the unloading action with the best target value and pair it with the environmental input as a training sample for the DNN network. S6. As the model is continuously generated, selected, and updated, the optimal offloading strategy for each device in multiple time slots is output.

2. According to the MEC stable computing migration system based on security and energy consumption awareness according to claim 1, the mobile edge computing (MEC) system is modeled as follows: Consider a MEC system consisting of an edge server ES and N wireless mobile devices WDs within the communication coverage of the server in continuous time T. The time period T is divided into T time slices of equal size, and the length of each time slice is represented by ΔT = 1. The model adopts a binary offloading rule. Within the time slice t, it is determined that the tasks on the WDs are either all offloaded to the ES for execution or all calculated locally. Specifically, the offloaded WDs share a common bandwidth W and transmit the task data to the ES in a TDMA manner to avoid mutual interference between the communications of different WDs and ES. Binary variables represents the uninstall decision, where and 0 respectively indicate that WD i performs task offloading and local computation in time slice t. In the tth time slice, is the task data arriving at the WD i data queue. Assume that the arriving data Follow the independent and identically distributed (iid) with bounded second-order moments, Assume that η is estimated from past observations i The value of is known. is the channel gain between WD i and ES. Under the block fading assumption, Remains constant within a time slice, but varies independently between time slices. In time slice t, WD i has corresponding data arrival amount and data processing volume set up represents the queue length of WD i at the beginning of the tth time slice. Then, the dynamic task queue of WDs can be modeled as 3. According to the MEC stable computing migration system based on security and energy consumption awareness according to claim 1, its communication model is as follows: In time slice t, the transmission rate between WDi and ES is calculated as follows: Where W represents the transmission bandwidth between WDs and ES, represents the transmission power of WD i, represents the wireless channel gain between WD i and ES, σ 2 represents the Gaussian white noise power.

4. According to the MEC stable computing migration system based on security and energy consumption awareness according to claim 1, its risk probability model and security cost model are as follows: In order to measure the risk level of task data during the process of offloading computation from WDs to ES, a risk probability model needs to be established to quantify the risk probability of these offloaded tasks. Assume that the malicious attack of the offloaded task follows the parameter λ ep and λ ig Poisson distribution. After using algorithm j in security service type, the risk probability of data being leaked or tampered is calculated as Therefore, after selecting a specific encryption algorithm and hash function for security protection within time slice t, the risk probability of WD i performing task offloading being leaked or tampered is Different security protections will result in different security costs. During the WD i task offloading process, the time cost calculation formula is as follows: in is the safety decision of WD i in time slice t, Indicates that WD i uses algorithm j to perform type security services. Indicates the amount of tasks offloaded by WD i, represents the speed of algorithm j in security service type. The energy cost consumed can be expressed as: Where α is the proportional coefficient between safety energy consumption cost and time cost.

5. According to the MEC stable computing migration system based on security and energy consumption awareness according to claim 1, its computing rate and energy consumption cost model is as follows: When WD i decides to process data locally within time slice t Express the local CPU frequency as The original data (bit) processed by WD i in time slice t in local computing mode is The calculation rate is The energy consumed is When WD i decides to offload task data to ES computing within time slice t Due to the constraints of TDMA, all WDs that decide to offload need to be executed serially with the data encryption action within one time slice, which is τ i ΔT is the offloading time of the WD i. Assuming that the computation speed of ES is much faster than that of WDs, it may be more than three orders of magnitude. In addition, the amount of data from WDs to ES is much smaller than the amount of data offloaded from WDs to ES. Therefore, the time spent by the task on ES computing and returning data to WDs can be ignored, so that each time slice is only occupied by data encryption and task offloading, that is, In time slice t, it is known that the transmission rate between WD i and ES is Then the amount of tasks processed by WD i that performs computation offloading is The energy consumed during task offloading transmission is Therefore, in the tth time slice, the task computation amount on each mobile device in MEC is The total energy consumption of the system is 6. According to the security and energy consumption-aware MEC stable computing migration system of claim 1, its optimization objectives and constraints are as follows: Optimization goal: Constraints: ① The amount of local computing and offloading tasks should be less than the task queue length requirement: ② The sum of the security service and unloading transmission time should be less than the time slot length requirement: ③The task queue should maintain a long-term stable state: ④ The task transmission risk of the mobile device to be offloaded should be less than the risk threshold requirement of the current time slot:

Images