Privacy protection computing resource dynamic scheduling method for weak current system

Abstract

The invention discloses a privacy protection computing resource dynamic scheduling method for a weak current system. The method comprises the following steps: acquiring physical states and network topology states of a plurality of computing nodes in the weak current system in real time; the method comprises the following steps: for a to-be-scheduled privacy calculation task, calculating a privacy entropy value for dynamically quantifying a data leakage risk of the to-be-scheduled privacy calculation task based on confidence and sensitivity weight of personal identifiable information in data content of the to-be-scheduled privacy calculation task and context information of the to-be-scheduled privacy calculation task; using a pre-trained deep reinforcement learning model to take the obtained state of the weak current system and the privacy entropy value as input, and outputting an optimal scheduling decision; executing the task on the target computing node according to the decision; according to the method, the privacy entropy is introduced and deep reinforcement learning is adopted, so that prospective joint scheduling of the privacy protection level and the computing resources is realized, and the resource utilization rate, especially the energy efficiency, of the weak current system is remarkably improved while the data security is guaranteed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of resource scheduling, and in particular to a privacy protection computing resource dynamic scheduling method for a weak current system. BACKGROUND

[0002] With the deep integration of Internet of Things (IoT) and edge computing technology, the weak current system taking devices such as cameras, sensors and intelligent access control as nodes is evolving from a traditional signal transmission network into an edge intelligent platform carrying complex computing tasks. Under this background, a large amount of unstructured data containing personally identifiable information (PII), such as high-definition video streams and audio signals, urgently needs to be processed in real time on heterogeneous and resource-constrained edge nodes close to the data source. Existing technical solutions usually use privacy computing technologies such as homomorphic encryption, trusted execution environment (TEE) or differential privacy to protect data security, and combine algorithms based on rules or greedy strategies for computing resource scheduling. The development of these technologies provides a foundation for deploying intelligent applications in weak current systems.

[0003] However, the existing technology has significant limitations in dealing with dynamic and complex application scenarios in weak current systems. First, existing privacy protection strategies are static and coarse-grained, and cannot be dynamically adjusted according to the actual privacy risks contained in the data content, which leads to "overprotection" of low-risk data, wasting valuable computing and energy resources, and "insufficient protection" of high-risk data, creating security risks. Second, traditional scheduling methods are essentially "reactive", and their decisions are based only on the current system state snapshot, lacking the ability to predict future load bottlenecks, network congestion or key node energy depletion, making the system less robust. More critically, there is an inherent conflict between task processing delay, privacy protection strength and system resource consumption, and existing technologies cannot find a dynamic Pareto optimal solution among these mutually restrictive goals, failing to maximize the overall utility of the system. SUMMARY

[0004] This section is intended to summarize some aspects of the embodiments of the present application and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract and title of the specification to avoid obscuring the purpose of this section, abstract and title, and such simplifications or omissions cannot be used to limit the scope of the present application.

[0005] In view of the above existing problems, the present application is proposed. Therefore, the present application provides a privacy protection computing resource dynamic scheduling method for a weak current system to solve the problems proposed in the background art.

[0006] To solve the above technical problems, the present application provides the following technical solutions: a privacy protection computing resource dynamic scheduling method for a weak electric system, comprising: Real-time acquisition of the physical state of a plurality of computing nodes in the weak electric system, and the task attribute of at least one to-be-scheduled privacy computing task; For the to-be-scheduled privacy computing task, based on its data content and context information, a privacy entropy value representing the data leakage risk of the task is calculated; Based on the physical state, the task attribute and the privacy entropy value, a pre-trained deep reinforcement learning model is used to output a scheduling decision, which specifies the selected privacy computing task, a target computing node, and a privacy protection scheme matched with the privacy entropy value; According to the scheduling decision, the selected privacy computing task is executed on the target computing node using the privacy protection scheme.

[0007] As a preferred scheme of the privacy protection computing resource dynamic scheduling method for a weak electric system of the present application, wherein: the physical state of the computing node at least includes: The central processor load, available memory, remaining power level and / or charging and discharging rate of the node.

[0008] As a preferred scheme of the privacy protection computing resource dynamic scheduling method for a weak electric system of the present application, wherein: in addition to the real-time acquisition of the physical state of a plurality of computing nodes in the weak electric system, and the task attribute of at least one to-be-scheduled privacy computing task, it also includes: Acquire the network topology state between nodes in the weak electric system, including the network delay and available bandwidth between nodes.

[0009] As a preferred scheme of the privacy protection computing resource dynamic scheduling method for a weak electric system of the present application, wherein: for the to-be-scheduled privacy computing task, based on its data content and context information, a privacy entropy value representing the data leakage risk of the task is calculated, further comprising: Pre-detect the input data of the privacy computing task, and acquire the confidence of one or more pre-defined personally identifiable information contained in the input data; According to the confidence, combining the sensitivity weight of each personally identifiable information pre-set, a pre-set entropy calculation function is used to obtain a basic privacy risk value; According to the context information of the privacy computing task, a context risk correction factor is determined; The context risk correction factor is used to correct the basic privacy risk value to obtain the privacy entropy value.

[0010] As a preferred scheme of the privacy protection computing resource dynamic scheduling method for a weak electric system according to the application, wherein: the deep reinforcement learning model is trained by receiving a composite reward signal containing a quality of service reward, a resource cost penalty and a privacy matching degree reward.

[0011] As a preferred scheme of the privacy protection computing resource dynamic scheduling method for a weak electric system according to the application, wherein: the determination method of the privacy matching degree reward comprises: Comparing the preset protection strength of the privacy protection scheme specified in the scheduling decision with the privacy entropy value, and determining the reward value according to the comparison result: If the protection strength is lower than the privacy entropy value, it is determined that the reward value is a negative reward, and the penalty amplitude of the negative reward increases at an accelerated rate as the difference between them increases. If the protection strength is not lower than the privacy entropy value, it is determined that the reward value is a positive reward, and the size of the positive reward decreases as the difference between the protection strength and the privacy entropy value increases.

[0012] As a preferred scheme of the privacy protection computing resource dynamic scheduling method for a weak electric system according to the application, wherein: the deep reinforcement learning model comprises: a policy network and a value network; The policy network is used to output the probability distribution on all possible scheduling decisions according to the physical state, task attribute and privacy entropy value of the current weak electric system; The value network is used to evaluate the long-term value of the current state according to the physical state, task attribute and privacy entropy value of the current weak electric system; Wherein, when the policy network generates a scheduling decision, the probability distribution of the scheduling decision is affected by the long-term value evaluated by the value network.

[0013] As a preferred scheme of the privacy protection computing resource dynamic scheduling method for a weak electric system according to the application, wherein: the resource cost penalty in the composite reward signal is determined according to the weighted sum of the power consumption, computing resource and / or network bandwidth consumed in executing the scheduling decision, wherein the weight of the power consumption is set to the highest.

[0014] As a preferred scheme of the privacy protection computing resource dynamic scheduling method for a weak electric system according to the application, wherein: the privacy protection scheme is selected from a predefined scheme set, each scheme in the set has a preset protection strength and an associated computing cost model.

[0015] As a preferred scheme of the privacy protection computing resource dynamic scheduling method for a weak electric system according to the application, wherein the pre-defined scheme set at least includes the following two schemes: The scheme without privacy protection, the scheme based on differential privacy, the scheme executed in a trusted execution environment and the scheme based on homomorphic encryption.

[0016] Compared with the prior art, the beneficial effects of the scheme of the application are: 1. By introducing the concept of privacy entropy value, the privacy risk of each task is dynamically and quantitatively evaluated by real-time analysis of task data content and context environment, which overcomes the drawbacks of the static protection strategy of the prior art, avoids resource waste caused by over-protection of low-risk data, and prevents security risks left by insufficient protection of high-risk data. 2. The pre-trained deep reinforcement learning model is adopted, the current and future possible dynamic changes (such as node power trend, network congestion, etc.) of the weak electric system are comprehensively considered, and forward-looking and global scheduling planning is performed, so that the model learns to find a dynamic optimal balance among the three conflicting goals of task performance, resource cost and privacy security, which significantly improves the overall utility and robustness of the system in complex environment. 3. In addition, in the training process of the deep reinforcement learning model, a high penalty weight is given to power consumption in the reward function to guide the scheduling decision to give priority to energy efficiency, and through the privacy matching degree reward mechanism, unnecessary energy consumption caused by over-protection is avoided, thereby effectively prolonging the endurance time of each node in the system and ensuring the long-term stable operation of the entire weak electric system. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical scheme of the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor. Among them: Figure 1 The overall flowchart of the privacy protection computing resource dynamic scheduling method for a weak electric system according to an embodiment of the application. DETAILED DESCRIPTION

[0018] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0019] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0020] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0021] This invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of this invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not adhering to the usual scale. Furthermore, the schematic diagrams are merely examples and should not be construed as limiting the scope of protection of this invention. In actual fabrication, the three-dimensional spatial dimensions of length, width, and depth should be included.

[0022] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0023] Unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" in this invention should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; similarly, they can refer to mechanical connections, electrical connections, or direct connections, or indirect connections through an intermediate medium, or internal connections between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0024] Example 1 Reference Figure 1This is the first embodiment of the present invention, which provides a method for dynamic scheduling of privacy-preserving computing resources for low-voltage systems, including: S1. Real-time acquisition of the physical status of multiple computing nodes in the low-voltage system, as well as the task attributes of at least one privacy computing task to be scheduled. It should be noted that this step is the data foundation of the entire dynamic scheduling method. The core objective of this step is to build a comprehensive and real-time environmental state representation for the subsequent deep reinforcement learning model. This state representation will provide a basis for scheduling decisions and ensure that the scheduling decisions can adapt to the constantly changing conditions of the weak current system. Furthermore, since the computing resources and energy status of nodes are equally important constraints in low-voltage systems, it is necessary to obtain the physical state of computing nodes to assess the computing task capacity that each node can undertake. Based on this, the following physical state indicators are mainly focused on in the solution of this invention: Central Processing Unit (CPU) Load: This metric indicates the current busyness of a compute node, usually expressed as a percentage. High CPU load means that the compute node has limited ability to process new tasks, which may cause delays in task execution. Available memory: This metric represents the amount of memory that a computing node can currently allocate. For some privacy computing tasks, especially those involving complex encryption operations, a large amount of memory may be required. Insufficient available memory will lead to task execution failure or frequent memory swapping in low-voltage systems, which will seriously affect performance. Remaining power level: This indicator is the core indicator of low-voltage systems. It is directly related to whether a computing node can complete a task. For example, if a time-consuming task is scheduled to a computing node with low power, the task will likely be interrupted due to the computing node shutting down, resulting in a waste of computing resources and time. Charge / discharge rate: This metric is designed to provide trend information on the dynamic changes in power levels. For example, a node that is charging (positive rate) is better suited to handle power-intensive tasks than a node that is discharging rapidly (negative rate), even if they are at the same remaining power level. Furthermore, monitoring agents deployed on each computing node periodically collect the aforementioned physical status indicators and report them to the scheduling center. During this process, in order to formally represent the computing node status, i.e., at time... For the first in a low-voltage system Each computing node has a physical state vector. It can be represented as: in, Representing the The computing nodes at time... CPU load (typically 0~100%). Representing the The computing nodes at time... The available memory size (e.g., in GB); Representing the The computing nodes at time... The remaining battery percentage (usually 0~100%). Representing the The computing nodes at time... The charge / discharge rate (e.g., in mA, with positive values ​​indicating charging and negative values ​​indicating discharging). Specifically, the monitoring agent can be a lightweight service running in the background of the node's operating system. It obtains information such as CPU and memory by calling APIs provided by the operating system (such as the / proc file system under Linux), and reads power and current data through the interface of the battery management chip. Furthermore, the scheduling center must first understand the basic requirements of the tasks to be processed in order to make appropriate matching. Typically, when a privacy computing task (e.g., a video stream analysis request from a smart camera) flows into the scheduling center, the task is stored in a scheduling queue. For each task in this queue, its task attributes must include at least the following: Task computational load estimate: This represents an estimate of the CPU cycles or number of standard operations required to execute a task. Task memory requirement: Indicates the minimum memory space required for the task to run; Input / output data size: This indicates the size of the raw data that the task needs to process and the size of the expected result data. This data size is directly related to network transmission overhead and the storage size of the computing nodes. Task deadline or delay requirement: This indicates the urgency of the task and is an important indicator for measuring Quality of Service (QoS). Task priority: The importance level of the task defined by the low-voltage system or the user; It should be noted that the above task attributes can be specified by the user when submitting the task, or estimated by analyzing similar historical tasks. Furthermore, since the cost (time and energy consumption) of data transmission in a multi-node collaborative low-voltage system cannot be ignored, it is also necessary to obtain the network topology status between nodes in order to achieve effective cross-node collaboration. Furthermore, the network topology status between nodes mainly includes the network latency between nodes and the available bandwidth between nodes; Specifically, inter-node network latency refers to the time required for a signal to travel back and forth between two nodes. For interactive tasks that require fast response, it is necessary to ensure low-latency transmission between the two nodes. Specifically, the available bandwidth between nodes refers to the current available data transmission rate of the data transmission link between two nodes. For tasks that require the transmission of a large amount of input data, high bandwidth needs to be guaranteed to shorten the preparation time before the task starts. It should be noted that the aforementioned network topology status can be dynamically obtained through periodic probing between nodes (such as ping tests and bandwidth tests), and this network topology status can be organized into a delay matrix. and a bandwidth matrix ,in Indicates at time From the From the computing node to the... The network latency of each computing node is similar. Indicates at time From the From the computing node to the... Available bandwidth of each computing node; S2. For the privacy computing task to be scheduled, a privacy entropy value representing the risk of data leakage of the task is calculated based on its data content and context information. It should be noted that this step is the core of achieving a dynamic balance between privacy protection and resource consumption. Unlike traditional methods that use a fixed, one-size-fits-all privacy protection level, this invention introduces the concept of privacy entropy, which aims to accurately and dynamically quantify the data leakage risk inherent in each privacy computing task to be scheduled. The privacy entropy value is a scalar normalized to a specific range (e.g., [0,1]). The higher the value, the greater the sensitivity of the task data and the higher the potential leakage risk, thus requiring a more robust privacy protection scheme. Furthermore, the input data for privacy computing tasks is pre-detected to obtain the confidence level of personally identifiable information; Specifically, for a task to be scheduled, its input data (e.g., a video stream, an image, or an audio clip) first needs to be quickly scanned to identify any predefined personally identifiable information (PII) it may contain. These PII categories then form a set. ,For example, Represents the human face, Represents the license plate number. Represents specific keywords in speech, etc.; Furthermore, to ensure efficient execution in resource-constrained edge environments, this pre-detection step employs a lightweight, pre-trained deep learning model (such as a miniature version of MobileNet or YOLO) as the PII detector. This detector performs a forward pass through the input data and outputs a confidence vector. , This indicates that the first [number] element was detected in the input data. The confidence score of a PII category is as follows: for example, if a face appears clearly in the video, the confidence score of the corresponding face category may be close to 1; if it is just a blurry outline, the confidence score may be 0.3; if no PII category is detected, the confidence score is 0. It is important to emphasize that this PII detector can be pre-trained on public datasets (such as WIDER FACE for face detection and COCO for general object detection) and fine-tuned by combining labeled private datasets collected for specific low-voltage system scenarios (such as indoor monitoring and access control) to improve the recognition accuracy of specific PIIs (such as specific employee faces and company visitor license plates). The lightweighting of the model can be achieved through techniques such as knowledge distillation or network pruning. Furthermore, since different types of PII pose varying degrees of harm when leaked—for example, facial information is generally more sensitive than license plate numbers—this invention pre-determines a sensitivity weight for each PII category. , All sensitivity weights are stored in a weight vector. In this context, the standards are set by domain experts or in accordance with personal information privacy protection regulations. Furthermore, combining the confidence vector and weight vector obtained in the previous step, a basic privacy risk value is calculated using a pre-defined entropy calculation function. This entropy calculation function is a weighted sum of various PII risks, and its formal expression is as follows: in, Indicates privacy-preserving computation tasks to be scheduled. Basic privacy risk value; This represents the total number of predefined PII categories; Indicates the first Sensitivity weights for PII-like systems; This indicates that the first [item] was detected in the data of the privacy-preserving computation task to be scheduled. Confidence level of PII-like systems; It should be noted that, for this formula, the more types of PIIs the data of the privacy computing task to be scheduled contains, the higher the confidence level, and the greater the sensitivity weight of these PIIs, the higher the basic privacy risk value. Furthermore, since analyzing data content alone is insufficient to comprehensively assess risk, as the contextual information of task execution is equally crucial—for example, processing facial data of family members in a private residence carries a far lower risk than processing facial data of strangers in a public square—this invention also introduces a contextual risk correction factor. To dynamically adjust the base risk value; Specifically, this context information may include: Geographic location: The physical location where the task takes place (indoor, outdoor, public area, private area, etc.); Time: The time when the task occurs (daytime, nighttime, weekday, holiday, etc.); Data owner intent: Has the data owner explicitly authorized this processing? System security status: Has any security threat or attack been detected in the current low-voltage system? Specifically, this discrete or continuous contextual information is input into a pre-defined rule engine or a small machine learning model (e.g., a decision tree), and outputs a contextual risk correction factor. ,For example, The value range can be set to [0.5, 2.0] when the context is safe (such as in a private residence). A value of 0.7 indicates a reduced risk assessment; this is appropriate when the contextual risk is high (e.g., in crowded public places). A value of 1.5 indicates a higher risk assessment. Furthermore, by combining the basic privacy risk value with the contextual risk correction factor, the final privacy entropy value is calculated, resulting in: in, This is a pruning function used to restrict the calculation results to a preset range of privacy entropy values. Within (e.g., [0,1]), to ensure the normalization and stability of the output; It should be noted that, through the above processing steps, the present invention can dynamically generate a highly contextualized privacy entropy value for each privacy computing task to be scheduled. This value is not a static, rough estimate, but a measure that integrates the inherent sensitivity of the data and the risks of the external environment. S3. Based on physical state, task attributes and privacy entropy value, a pre-trained deep reinforcement learning model is used to output a scheduling decision. The scheduling decision specifies the selected privacy computing task, a target computing node, and a privacy protection scheme that matches the privacy entropy value. It should be noted that this step is the core intelligent decision-making part of the entire scheduling method. It aims to take the weak current system state and task risk obtained and processed in the aforementioned S1 and S2 steps as input, and autonomously explore and output an optimal scheduling decision in a complex, multi-dimensional scheduling decision space through a trained deep reinforcement learning (DRL) model. This decision aims to dynamically balance the three conflicting objectives of quality of service (QoS), system resource cost (especially energy consumption), and privacy protection level. Furthermore, due to the environmental state observed by the deep reinforcement learning model... It is the basis of its decision-making, so, at any given moment ,state It is a vector or tensor that contains all the relevant information obtained from steps S1 and S2, as follows: Node status: All The physical state of each computing node, including CPU load, available memory, remaining power, and charge / discharge rate; Network status: The network topology status between nodes, including latency and available bandwidth matrix; Task status: Currently in the queue of tasks awaiting scheduling The attributes of each task include computational load, memory requirements, data size, deadline, etc., as well as the privacy entropy value corresponding to each task calculated through the S2 step; Furthermore, since the model's output is a scheduling decision, i.e., an action... This action is a triple that defines "which task to schedule, on which node, and with what privacy protection scheme to execute it." Therefore, an action can be formalized. Represented as: in, This represents the index of the task selected from the queue of tasks to be scheduled. This represents the index of the compute node selected as the target execution node. Represents the index of the selected scheme from a predefined set of privacy protection schemes; Furthermore, for this predefined set of schemes, the set includes at least two of the following schemes: schemes without privacy protection, schemes based on differential privacy, schemes executed in a trusted execution environment, and schemes based on homomorphic encryption; wherein each scheme Each of these is associated with a preset protection strength. (For example, a value between [0,1], where higher values ​​provide stronger protection) and a computational cost model. (The model estimates the additional CPU, memory, and energy costs associated with this approach.) Furthermore, for deep reinforcement learning models, this invention preferably employs an actor-critic architecture, specifically including: Policy network: Receives the current state As input, the output is a probability distribution over all possible scheduling decisions (actions). In actual decision-making, the model will select an action based on this probability distribution, with the goal of learning an optimal policy to maximize long-term cumulative rewards. Value network: same receiving state As input, output a scalar value. The value network, or state-value function, is used to evaluate the expected total of future rewards that can be obtained by following the current policy in the current state (i.e., the long-term value of the state). The role of the value network is to "judge" the quality of the decisions made by the policy network and guide the updating of the policy network. It should be noted that the probability distribution of the policy network when generating decisions is affected by the evaluation results of the value network, for example, by calculating the advantage function. Policy networks increase the probability of actions leading to higher value and decrease the probability of actions leading to lower value. It is an action-value function; Furthermore, the model's pre-training process involves extensive interaction with a simulated environment and receiving a carefully designed composite reward signal. To accomplish; Specifically, the simulation environment includes: A task generator, for example, follows a Poisson process to simulate the random arrival of new tasks, whose attributes (such as computational cost and data size) are generated by sampling from a predefined distribution; A node state simulator that dynamically updates the physical state of each computing node at each time step based on the node's initial state (power consumption, CPU load, etc.) and the resources consumed in executing the task (calculated according to the cost model). For example, power consumption can be modeled as a function related to CPU utilization and task duration. A network state simulator is used to simulate the dynamic changes in network latency and bandwidth between nodes. For example, network congestion can be simulated by introducing random fluctuations or functions based on background traffic models. In each interaction, the model selects an action (scheduling decision), and the simulation environment updates the state of the nodes and tasks in the scheduling queue according to the action, and calculates the corresponding composite reward signal to feed back to the agent. Through sufficient exploration and learning in this controllable and repeatable simulation environment, the model policy can converge. Specifically, this composite reward signal It takes into account three aspects: in, Service quality rewards are positive rewards, primarily based on task completion, such as whether the task is completed on time. If completed on time, a larger positive reward is given; if completed late, a negative reward or zero reward is given. This represents a resource cost penalty, which is a negative reward (penalty item). The penalty is a weighted sum of the electricity, computing resources, and network bandwidth consumed in executing the scheduling decision. In particular, the weight of electricity consumption is set to the highest to highlight the sensitivity of the low-voltage system to energy consumption. ,in, These represent the electricity consumed, computing resources used, and network bandwidth consumed, respectively. These are the weights of the corresponding power consumption, computing resources, and network bandwidth, and To be the largest; The privacy matching reward is a positive or negative reward used to incentivize the model to learn to select a privacy protection scheme that exactly matches the task's privacy entropy value. Furthermore, regarding this privacy-matching reward, ,in, It is the predefined scheme selected above. The strength of protection, It is the privacy entropy value; Specifically, if the protection is insufficient ( If the difference is 0, the model will receive a negative reward, and the magnitude of this penalty will vary with the difference. The increase in size accelerates as the number of people increases, aiming to severely punish the lack of privacy protection; Specifically, if protection is adequate or excessive ( If the protection strength exceeds the privacy entropy value, the model will receive a positive reward, but the size of this reward will vary depending on the difference between the protection strength and the privacy entropy value. The value decreases as it increases, which is intended to encourage models to avoid unnecessary waste of resources while meeting privacy requirements (because stronger protection schemes are usually more expensive). It should be noted that by continuously optimizing the model during training to maximize the long-term expectation of this composite reward signal, the deep reinforcement learning model can eventually learn how to dynamically balance task performance, energy consumption, and privacy in complex low-voltage system environments, thereby receiving real-time system states. Then, it can quickly output a high-quality, multi-objective optimized scheduling decision. ; S4. Based on the scheduling decision, the selected privacy computing task is executed on the target computing node using a privacy protection scheme. It should be noted that, since the deep reinforcement learning model has already output a triplet scheduling decision that includes task, node, and privacy scheme, Therefore, the core objective of this step is to transform this abstract decision into a concrete operation in the physical world, ensuring that the selected task can be executed correctly and efficiently on the target computing node in accordance with the specified privacy protection scheme. Furthermore, the scheduling center generates scheduling decisions. Then, the decision is immediately assigned to the target computing node. This process is typically accomplished through a control channel, where the assigned instruction packet contains at least: Task identifier: explicitly indicates the task that needs to be performed; Task input data: A pointer to the location where the input data is stored; Privacy scheme identifier: Clearly informs the compute node which predefined privacy protection scheme should be used; Furthermore, the task execution process is initiated by receiving and parsing this instruction packet through an agent program deployed on the target computing node; Furthermore, the node agent invokes the corresponding execution logic based on the received privacy scheme identifier. Since the predefined scheme set contains schemes with different protection strengths and resource overheads, the node agent only needs to select one of them to execute according to the instruction packet. Specifically, the implementation methods for each plan are as follows: When choosing the option without privacy protection: This corresponds to the level of protection. The case where the value is 0 is determined by the deep reinforcement learning model in step S3, usually due to the privacy entropy value of the task. Extremely low, and does not involve any sensitive data; the node agent will directly execute the original computing tasks in its computing environment without any additional privacy processing operations. This process follows the most efficient and resource-lowest execution method, ensuring the maximization of the performance of the weak current system without privacy risks. When choosing a differential privacy-based approach: this approach is typically chosen only when quantifiable privacy guarantees are provided in the statistical results. Deep reinforcement learning models will adjust the approach based on the privacy entropy value of the task. Choose one with appropriate protection strength Differential privacy schemes typically offer protection strength that is correlated with differential privacy budgets. Directly related (the higher the protection strength, the smaller the differential privacy budget); during task execution, the node agent applies a differential privacy mechanism to its output or intermediate computation process. That is, the node agent injects a precisely calculated random noise into the actual computation result. A mechanism that satisfies differential privacy must comply with the following constraints: in, and These are two adjacent datasets that differ by only one record. In the dataset The output after applying differential privacy mechanism, It is any set of possible outputs. Differential privacy budgeting controls the strength of privacy protection. The smaller the value, the stronger the privacy protection, but the greater the injected noise and the lower the data availability. This is the failure probability, usually a very small number, representing the probability that the mechanism deviates from differential privacy; finally, the target node will proceed according to the preset parameters of the selected scheme. and The parameters are used to generate noise and add it to the result, then the noisy result is returned. When choosing a scheme that executes in a trusted execution environment: this scheme is typically chosen to handle tasks with high privacy entropy values. The task requires the target computing node to possess a hardware-level Trusted Execution Environment (TEE), such as Intel SGX or ARM TrustZone. The node agent first creates a secure enclave within the computing node's TEE, and then securely loads the task's computation logic and the input data requiring protection into this enclave. It should be explained that a TEE is a secure computing unit implemented using hardware isolation technology, primarily used to protect the execution environment of sensitive data and code. During this process, the entire computation of the task is completed within this memory-encrypted enclave, isolated from the host operating system. External entities (including the operating system with the highest privileges) cannot spy on or tamper with the code and data within the enclave. After computation, the result can be encrypted and transmitted, or remotely verified within the enclave to confirm the integrity and confidentiality of the computation to the requesting party. When choosing a homomorphic encryption-based scheme: this scheme is typically chosen in scenarios with the highest privacy requirements (i.e., (Approximately 1), and requires computation to be completed under end-to-end data encryption; in this mode, the task's input data has been encrypted by the data owner using a homomorphic encryption public key before being submitted to the scheduling center, so the target computing node receives the input data and the task's homomorphic computation logic in ciphertext form; at this time, the computing node can complete the processing of the encrypted data by performing special homomorphic operations (such as addition and multiplication on the ciphertext), and the node has no access to the plaintext data throughout the entire process; finally, the encrypted result generated by the computation is returned to the data owner, and only the owner holding the private key can decrypt it to obtain the final plaintext result; It should be noted that while this approach provides top-level privacy protection, it also incurs the greatest computational and communication overhead. Deep reinforcement learning models will only choose this approach after weighing the costs and determining that such overhead is necessary and that the computing nodes are capable of handling it. Furthermore, once the task is completed, the target computing node returns the processing result (which may be plaintext, noisy result, or ciphertext) to the scheduling center or the task requester. At the same time, the node agent reports the actual resource consumption of this task execution (such as CPU time, peak memory, actual power consumption, etc.) to the scheduling center. This feedback data can be used to verify and correct the computational cost model of each privacy scheme in step S3, and can be used as empirical data for incremental training or fine-tuning of future deep reinforcement learning models, thereby achieving continuous dynamic optimization scheduling.

[0025] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0026] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0027] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0028] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0029] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0030] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A privacy protection computing resource dynamic scheduling method for a weak electric system, characterized in that, The method comprises: obtaining physical states of a plurality of computing nodes in a weak electricity system and task attributes of at least one to-be-scheduled privacy computing task in real time; calculating a privacy entropy value representing a data leakage risk of the to-be-scheduled privacy computing task based on data content and context information of the to-be-scheduled privacy computing task; outputting a scheduling decision by using a pre-trained deep reinforcement learning model based on the physical states, the task attributes, and the privacy entropy value, the scheduling decision specifying a selected privacy computing task, a target computing node, and a privacy protection scheme matching the privacy entropy value; executing the selected privacy computing task on the target computing node by using the privacy protection scheme according to the scheduling decision.

2. The privacy-preserving dynamic scheduling method for weak electric system oriented computing resources according to claim 1, wherein, The physical states of the computing nodes at least include: a central processing unit load, available memory, remaining power level, and / or charging and discharging rate of the nodes.

3. The privacy-preserving dynamic scheduling method for weak electric system oriented computing resources according to claim 1, wherein, In addition to the real-time obtaining of the physical states of the plurality of computing nodes in the weak electricity system and the task attributes of the at least one to-be-scheduled privacy computing task, the method further comprises: obtaining network topology states between the nodes in the weak electricity system, the network topology states including network delay and available bandwidth between the nodes.

4. The privacy-preserving dynamic scheduling method for weak electric system oriented as claimed in claim 1, wherein, The calculating of the privacy entropy value representing the data leakage risk of the to-be-scheduled privacy computing task based on the data content and the context information of the to-be-scheduled privacy computing task further comprises: pre-detecting input data of the privacy computing task to obtain a confidence of one or more pre-defined personally identifiable information contained in the input data; obtaining a basic privacy risk value by using a pre-set entropy calculation function according to the confidence and a pre-set sensitivity weight of each personally identifiable information; determining a context risk correction factor according to the context information of the privacy computing task; and correcting the basic privacy risk value by using the context risk correction factor to obtain the privacy entropy value.

5. The privacy-preserving dynamic scheduling method for weak electric system-oriented computing resources according to claim 1 or 4, wherein, The deep reinforcement learning model is trained by receiving a composite reward signal including a service quality reward, a resource cost penalty, and a privacy matching degree reward.

6. The privacy-preserving dynamic scheduling method for weak electric system oriented as claimed in claim 5, wherein, The determination of the privacy matching degree reward comprises: comparing a pre-set protection strength of the privacy protection scheme specified in the scheduling decision with the privacy entropy value, and determining a reward value according to a comparison result: if the protection strength is lower than the privacy entropy value, determining the reward value as a negative reward, and the penalty amplitude of the negative reward is accelerated to increase with an increase of a difference between the two values; if the protection strength is not lower than the privacy entropy value, determining the reward value as a positive reward, and the size of the positive reward decreases with an increase of a difference between the protection strength and the privacy entropy value.

7. The privacy-preserving dynamic scheduling of computing resources for a low-power system oriented method according to claim 1, wherein, The deep reinforcement learning model comprises: a policy network and a value network; the policy network is used to output a probability distribution on all possible scheduling decisions according to the physical states, the task attributes, and the privacy entropy value of the current weak electricity system; the value network is used to evaluate a long-term value of a current state according to the physical states, the task attributes, and the privacy entropy value of the current weak electricity system. The probability distribution of the scheduling decision of the policy network is affected by the long-term value evaluated by the value network when generating the scheduling decision.

8. The privacy-preserving dynamic scheduling method for weak electric system oriented as claimed in claim 5, wherein, The resource cost penalty in the composite reward signal is determined according to a weighted sum of the power, computing resource and / or network bandwidth consumed by executing the scheduling decision, wherein the weight of the power consumption is set to be the highest.

9. The privacy-preserving dynamic scheduling of computing resources for a low-power system oriented method according to claim 1, wherein, The privacy protection scheme is selected from a predefined scheme set, each scheme in the set having a preset protection strength and an associated computing cost model.

10. The privacy-preserving dynamic scheduling method for weak electric system oriented as claimed in claim 9, wherein, The predefined scheme set at least includes the following two schemes: a scheme without privacy protection, a scheme based on differential privacy, a scheme executed in a trusted execution environment, and a scheme based on homomorphic encryption.

Images