An asymmetric computing power privacy computing task dynamic offloading system and method
Patent Information
- Application Number
- CN202611116646.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-27
- Publication Date
- 2026-10-27
AI Technical Summary
当引入“任务动态卸载—远端执行—结果回传”这一新链路时，现有机制暴露出密钥主权与控制权分离、缺乏与卸载过程强耦合等缺陷，无法为动态卸载场景下的计算过程安全提供有效保障
[0041]（1）降低参与门槛：弱算力节点无需升级硬件即可安全参与隐私计算，从根本上打破了资源对等部署的刚性约束。
Smart Images
Figure CN122994025A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of privacy computing technology, and in particular to a dynamic offloading system and method for asymmetric computing power privacy computing tasks. Background Technology
[0002] With increasingly stringent data security regulations and a growing demand for cross-organizational data collaboration, privacy-preserving computing technologies such as federated learning and secure multi-party computation have become key means to break down "data silos." However, in actual deployments, participants often possess highly heterogeneous computing resources, ranging from powerful cloud servers to resource-constrained IoT devices. This asymmetric computing environment poses a significant challenge to existing privacy-preserving computing solutions.
[0003] Currently, mainstream privacy-preserving computation schemes, such as federated learning and secure multi-party computation, are typically based on a fundamental model of "local computation by each participant, aggregation by a central or coordinating party." In federated learning, each node performs gradient computation locally and uploads encrypted gradients to a central server via a secure aggregation protocol to complete joint modeling. Secure multi-party computation achieves collaborative computation through protocols such as secret sharing and obfuscated circuits. A common prerequisite for all these schemes is that each participant must possess the basic computing power to complete its local core computational tasks and a stable network connection.
[0004] To alleviate the problem of varying computing power across terminals, existing technologies have proposed optimization mechanisms for heterogeneous terminals, such as introducing proximal terms into the local objective function (i.e., proximal federated optimization) or improving convergence efficiency by selecting available nodes to participate in each round of training. However, these solutions only tolerate performance differences or select qualified nodes under heterogeneous conditions. Nodes with weak computing power still only participate "less" or "slowly." Once the local computing task exceeds their capabilities, they are excluded from collaboration or forced to outsource all raw data to high-performance nodes, thus losing their data sovereignty.
[0005] Another approach employs task splitting, edge-cloud collaboration, or split learning to offload computationally intensive tasks from weaker nodes to higher-performance nodes. For example, offloading decisions can be made by monitoring the computing resource status of terminal devices, or by dividing neural network layers between the client and server for collaborative training. However, these offloading decisions often rely on static presets or single-dimensional resource metrics (such as considering only CPU load), failing to perceive real-time dynamic changes in CPU / GPU utilization, memory pressure, network latency, etc., and thus unable to generate fine-grained, elastic allocation schemes. This leads to overloading of high-performance nodes while low-performance nodes remain idle. Furthermore, existing offloading research typically assumes the execution nodes are secure and trustworthy. After the weaker node offloads the computationally intensive subtask, the executing node can spy on the intermediate results, making the offloading process itself a new attack surface.
[0006] In terms of security protection, existing technologies mainly rely on cryptographic mechanisms such as secure aggregation, homomorphic encryption, and secret sharing to achieve data usability without visibility. These schemes are designed for scenarios with fixed-round aggregation, fixed participation relationships, and static key lifecycles, and their key management focuses on the fixed "upload-aggregation-decryption" link. When a new link of "dynamic task unloading-remote execution-result feedback" is introduced, existing mechanisms expose defects such as the separation of key sovereignty and control, and a lack of strong coupling with the unloading process, which cannot provide effective protection for the security of the computing process in dynamic unloading scenarios.
[0007] In summary, existing technical solutions face at least three pressing technical problems when dealing with asymmetric computing environments: (1) high participation threshold, with weak computing nodes excluded from collaboration due to their inability to complete local core computing; (2) rigid task deployment, unable to dynamically optimize scheduling based on multi-dimensional resource status; and (3) lack of security boundaries during computation unloading, and a key management mechanism that is strongly coupled with the unloading process. Summary of the Invention
[0008] (a) Purpose of the invention
[0009] To address the three technical shortcomings of existing technologies—high participation threshold, rigid task deployment, and lack of security boundaries in the unloading process—this invention aims to provide a method and system for dynamic unloading and secure collaborative deployment of privacy computing tasks in asymmetric computing environments, in order to achieve the following objectives:
[0010] First, it enables nodes with weak computing power to participate in privacy computing securely without upgrading their hardware, thus protecting their data sovereignty.
[0011] Second, it enables dynamic and elastic allocation of tasks among heterogeneous nodes, improving overall resource utilization efficiency.
[0012] Third, provide minimal, temporary computation authorization for each unloaded subtask to ensure the security boundaries of the entire computation unloading process.
[0013] (II) Technical Solution
[0014] To achieve the above-mentioned objectives, this invention provides a method for dynamic offloading and secure collaborative deployment of privacy computing tasks in an asymmetric computing environment, comprising the following steps:
[0015] Step 1: Task Analysis and Resource Awareness
[0016] The Task Management and Scheduling Center (TMSC) receives privacy-preserving computation tasks and parses them into a Directed Acyclic Graph (DAG) containing multiple atomic operators. Each atomic operator is an indivisible computational unit with computational complexity weights and data dependency transmission attributes. TMSC uses a topological sorting algorithm to generate a topological sequence of the DAG and calculates the earliest and latest start times for each atomic operator.
[0017] TMSC collects multi-dimensional resource status information of each participating node in real time, including CPU utilization, GPU utilization, memory pressure, network latency, available bandwidth, and the Trusted Execution Environment (TEE) support status and remote authentication report status of each node. TMSC establishes a two-way authentication TLS channel with each node based on a hardware root of trust.
[0018] Step 2: Dynamic Scheduling Decision
[0019] Based on the DAG and the multi-dimensional resource status information of each node, TMSC runs a multi-objective optimization scheduling algorithm with the goal of minimizing task completion time and maximizing resource utilization. Each atomic operator is assigned as a scheduling unit to the corresponding Computing Executor Node (CEN) to generate a scheduling mapping table.
[0020] Specifically, the multi-objective optimization scheduling algorithm adopts an improved ant colony algorithm, and its pheromone update rule is as follows: Where ρ is the volatility factor, The scheduling decision is determined by a linear combination of the reciprocal of the task completion time and resource utilization. The scheduling decision must satisfy the following constraints: for each computation executor node, the sum of the ratios of the computational amount of all atomic operators allocated to that node to the node's computational capacity does not exceed the time window corresponding to that node; and atomic operators involving sensitive data are only allocated to CENs with TEEs.
[0021] Step 3: Strategy Description Generation
[0022] The Key Synergy Service Node (KSSN) generates a policy description based on the scheduling mapping table and security policy. The policy description includes a policy identifier, a target CEN identifier, a time window, a list of allowed operator types, and a data identifier. The policy description is signed by the KSSN using its private key within its TEE.
[0023] Step 4: Temporary Token Generation
[0024] KSSN sends the policy description to the corresponding Data Owner Node (DON). After verifying the validity of KSSN's signature, DON generates a temporary computing token (TK) based on its master key (MK) and the policy description through a key encapsulation mechanism (KEM). The TK encapsulates the minimum functional key required to execute the corresponding atomic operator.
[0025] Specifically, the key encapsulation mechanism employs a hybrid public-key encryption standard, using the master key as the initial key material and the policy description as context information to generate a temporary symmetric key. The TK contains the temporary symmetric key encrypted using the KSSN public key, the policy description hash value, and the digital signature of DON.
[0026] Step 5: Token Distribution and Re-encryption
[0027] DON sends the encrypted computation data and TK to KSSN. KSSN verifies the validity of TK within its TEE, decrypts it using its private key to obtain a temporary symmetric key, and then re-encrypts the temporary symmetric key using the target CEN's TEE public key to generate a re-encryption token. KSSN routes the re-encryption token and encrypted data to the target CEN according to the CEN identifier in the policy description.
[0028] Step 6: Dense-state calculation execution and result return
[0029] After verifying the validity of the time window within its TEE, the target CEN uses its TEE private key to decrypt and re-encrypt the token to obtain a temporary function key. It then uses this function key to perform the corresponding cryptographic calculation on the encrypted data, obtains the calculation result, and returns it.
[0030] Step 7: Token Destruction
[0031] Upon completion of the task, CEN erases the function key and re-encryption token within its TEE using a secure erase command, and generates a destruction confirmation message containing a policy identifier and a TEE remote authentication report, which is then sent to KSSN. KSSN records the destruction confirmation message in its policy rule database for auditing purposes.
[0032] The method further includes a dynamic rescheduling step: when it is detected that the CPU utilization of CEN exceeds a preset threshold for a continuous period of time, or the network latency exceeds a preset threshold, or the TEE remote authentication report of CEN expires, TMSC triggers rescheduling; if rescheduling fails, the corresponding atomic operator is rolled back to DON for local execution, and KSSN is notified to revoke the corresponding token.
[0033] The method also includes a token reuse detection step: when KSSN detects that the same policy identifier is used by multiple CEN requests, KSSN generates a revocation list and broadcasts it to all DONs and CENs, triggering the regeneration of tokens.
[0034] This invention also provides a system for dynamic offloading and secure collaborative deployment of privacy computing tasks in an asymmetric computing environment, comprising:
[0035] TMSC is used to receive privacy computing tasks, parse the tasks into DAGs, collect multi-dimensional resource status information of each node in real time, run a multi-objective optimization scheduling algorithm to generate a scheduling mapping table, and trigger rescheduling.
[0036] Multiple DONs, each holding an MK, are used to generate a TK based on the policy description;
[0037] Multiple CENs, each CEN deploying a TEE, are used to receive re-encryption tokens and encrypted data, perform cryptographic calculations within the TEE, and destroy the tokens;
[0038] KSSN, with a TEE deployed, is used to generate policy descriptions based on the scheduling mapping table and security policies, verify the validity of temporary tokens generated by DON, re-encrypt the tokens within the TEE before routing them to the target CEN, and record the entire process of token generation, distribution and destruction.
[0039] (III) Beneficial Effects
[0040] Compared with the prior art, the present invention has the following advantages:
[0041] (1) Lowering the threshold for participation: weak computing nodes can participate in privacy computing securely without upgrading their hardware, which fundamentally breaks the rigid constraint of resource-equal deployment.
[0042] (2) Optimize resource utilization: Through multi-dimensional resource perception and DAG fine-grained scheduling, the task can be elastically allocated among heterogeneous nodes.
[0043] (3) Ensure uninstallation security: The DON master key never leaves the local machine. A temporary minimized token is issued for each uninstallation and is destroyed immediately after the task is completed.
[0044] (4) Enhance system robustness: The dynamic rescheduling mechanism enhances the system's fault tolerance in node failure scenarios. Attached Figure Description
[0045] Figure 1 This is a schematic diagram of the overall architecture of the privacy computing task dynamic unloading and secure collaborative deployment system of the present invention.
[0046] Figure 2 This is a flowchart illustrating the method for dynamically unloading and securely co-deploying privacy computing tasks according to the present invention.
[0047] Figure 3 This is a flowchart illustrating the process of generating and distributing temporary computation tokens in this invention. Detailed Implementation
[0048] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.
[0049] Example 1: System Deployment and Task Initialization
[0050] This embodiment uses the training of a federated learning model across geographically dispersed medical institutions as its application scenario. The data owner nodes are 10 community hospitals (ordinary PCs), each possessing local patient electronic medical record data, aiming to jointly train a disease prediction model. The computation executor nodes are 2 large tertiary hospitals (GPU servers supporting Intel SGX TEE). The task management and scheduling center is deployed by a third-party technology platform of the medical consortium, and the key coordination service node is deployed on the TEE server (supporting AMD SEV) of the consortium's arbitration institution.
[0051] The system deployment is carried out as follows: a lightweight agent program, including an encryption library, communication module and resource status collector, is installed on all data owner nodes; a compute execution agent is installed on all compute executor nodes, which can interact with the Intel SGX Enclave and integrate a homomorphic encryption library; the task management and scheduling center and the key coordination service node are deployed as independent services on their respective servers.
[0052] Key initialization: Upon initial startup, each data owner node generates a unique master key using its hardware TPM endorsement key and a user-input random seed. This master key is securely stored in a local secure area and never leaves the node. The key coordination service node generates a coordination key pair, completes TEE remote authentication with all computation executor nodes, and establishes a trusted channel. A two-way authenticated TLS encrypted channel is established between all nodes.
[0053] Administrators submit federated learning training tasks through the task management and scheduling center, along with a security policy: "All dense computations must be performed on the computation executor node that has a TEE."
[0054] The multi-dimensional resource-aware scheduling engine of the task management and scheduling center sends status query commands to all nodes, and each node reports information using the UDP protocol through a lightweight proxy. After receiving the data, the task management and scheduling center smooths it using a weighted moving average method.
[0055] Example 2: Task Splitting and Scheduling Decisions
[0056] The task parser in the task management and scheduling center parses the federated learning training rounds into a DAG, which contains 13 nodes: nodes 1 to 10 correspond to the local gradient calculation tasks of each data owner node; node 11 is used to perform homomorphic addition aggregation on the encrypted gradients of the 1st to 5th DONs; node 12 is used to perform homomorphic addition aggregation on the encrypted gradients of the 6th to 10th DONs; and node 13 is used to perform addition aggregation on the aggregation results of nodes 11 and 12.
[0057] The scheduling engine runs an improved ant colony algorithm and outputs a scheduling mapping table based on the current resource status: Nodes 1 to 10 are assigned to their respective data owner nodes for local execution; Node 11 is assigned to computation executor node A; Node 12 is assigned to computation executor node B; and Node 13 is assigned to computation executor node A.
[0058] Example 3: Token Generation and Distribution
[0059] The task management and scheduling center sends the scheduling decision to the key coordination service node. The key coordination service node generates a policy description for subtask node 11, which includes the policy identifier, the target computation executor node identifier, the time window, the allowed operator types, and the data identifier, and signs it using the private key within the TEE.
[0060] The key coordination service node sends the policy description to the corresponding data owner node. After verifying the signature, each data owner node uses the master key and the policy description to generate a temporary computation token using a hybrid public-key encryption scheme. This token encapsulates the minimal function key required to perform homomorphic addition and is only valid for the target computation executor node within a specified time window.
[0061] Each data owner node sends the encrypted gradient data and temporary token to the key coordination service node. After verifying the identity and policy conditions of the computation executor node within its TEE, the key coordination service node uses the coordination key to perform proxy re-encryption of the temporary token, making it decryptable by the TEE private key of the target computation executor node. Then, it routes the re-encrypted token and encrypted data to the target computation executor node.
[0062] Example 4: Dense-state computation execution and result return
[0063] Each data owner node calculates the gradient locally, encrypts it using its own master key, and then uploads it to the designated storage in the task management and scheduling center.
[0064] The computation executor node A decrypts the temporary token within its TEE to obtain the temporary function key, and performs homomorphic addition on the cryptographic gradient of the corresponding data owner node to obtain an intermediate result. Subsequently, it decrypts another token, and adds the intermediate result to the intermediate result received from the computation executor node B to obtain the final aggregated ciphertext.
[0065] The computation executor node B decrypts the corresponding token within its TEE, performs homomorphic addition on the cryptographic gradient of the remaining data owner nodes, obtains the intermediate result, and sends it to the computation executor node A.
[0066] The computation executor node A returns the final aggregated ciphertext to the task management and scheduling center, which then broadcasts it to all data owner nodes. Each data owner node uses its own master key to decrypt and obtain the model gradient update for this round.
[0067] Example 5: Token Destruction and Anomaly Handling
[0068] Upon completion of the task, each compute executor node erases the temporary function key within its TEE using a secure erase command and sends a destruction confirmation containing the TEE remote authentication report to the key coordination service node. The key coordination service node marks the corresponding token as "destroyed" and records it in the audit log.
[0069] During task execution, if the GPU utilization of a certain computing executor node exceeds 90% for more than 5 seconds, the task management and scheduling center immediately triggers rescheduling, transferring some atomic operators to other available computing executor nodes. If rescheduling fails, execution will revert to the local execution on the data owner node, and the key coordination service node will be notified to revoke the corresponding token.
[0070] If the key coordination service node detects that the same policy identifier is requested by multiple computation executor nodes, it determines it as a token reuse attack, immediately generates a revocation list and broadcasts it to all nodes, triggering the affected data owner nodes to regenerate new tokens.
[0071] Example 6: Verification of Technical Effects
[0072] Verification through the above embodiments shows that the present invention enables weak computing nodes to participate in privacy computing tasks securely without upgrading hardware, realizes dynamic and elastic offloading of privacy computing tasks, and ensures the security boundary of the entire computing offloading process through a temporary token mechanism strongly coupled with the offloading process.
[0073] In summary, this invention, through its innovative system architecture and core protocol, has achieved significant technological advancements in lowering the barrier to entry, improving resource utilization efficiency, strengthening security boundaries, and enhancing system robustness. It can be deployed in privacy computing scenarios such as federated learning and secure multi-party computation, is suitable for cross-institutional data collaboration environments with heterogeneous computing power, and has industrial applicability.
Claims
1. A method for dynamic offloading and secure collaborative deployment of privacy computing tasks in an asymmetric computing power environment, characterized in that, Includes the following steps: (1) The task management and scheduling center receives the privacy computing task, parses the task into a directed acyclic graph containing multiple atomic operators, and collects the multidimensional resource status information of each participating node in real time; (2) The task management and scheduling center runs a multi-objective optimization scheduling algorithm based on the directed acyclic graph and the multi-dimensional resource status information of each node, with the goal of minimizing task completion time and maximizing resource utilization. Each atomic operator is assigned as a scheduling unit to the corresponding computing executor node, and a scheduling mapping table is generated. (3) The key coordination service node generates a policy description based on the scheduling mapping table and the security policy, and sends the policy description to the data owner node; (4) After the data owner node verifies the policy description, it generates a temporary computing token based on its master key and the policy description through a key encapsulation mechanism. The temporary computing token encapsulates the minimum functional key required to execute the corresponding atomic operator. (5) The data owner node sends the encrypted computation data and the temporary computation token to the key collaboration service node. After verifying the validity of the temporary computation token in the trusted execution environment, the key collaboration service node uses the collaboration key of the key collaboration service node to re-encrypt the temporary computation token to generate a re-encrypted token that can only be decrypted by the target computation executor node, and routes the re-encrypted token and the encrypted computation data to the target computation executor node. (6) The target computation executor node decrypts the re-encryption token in its trusted execution environment to obtain the minimized function key, and uses the function key to perform the corresponding encrypted computation on the encrypted computation data to obtain the computation result; (7) The target computation executor node returns the computation result and destroys the function key and related token information in its trusted execution environment after the task is completed.
2. The method according to claim 1, characterized in that, The multidimensional resource status information includes CPU utilization, GPU utilization, memory pressure, network latency, available bandwidth, as well as the trusted execution environment support status and remote authentication report status of each node.
3. The method according to claim 1, characterized in that, The multi-objective optimization scheduling algorithm in step (2) adopts an improved ant colony algorithm, and its pheromone update rule is as follows: Where ρ is the volatility factor, The scheduling is determined by a linear combination of the reciprocal of the task completion time and the resource utilization rate; the scheduling satisfies the following constraints: for each computation executor node, the sum of the ratio of the computational amount of all atomic operators allocated to that node to the computational capacity of that node does not exceed the time window corresponding to that node; and atomic operators involving sensitive data are only allocated to computation executor nodes with a trusted execution environment.
4. The method according to claim 1, characterized in that, The policy description includes a policy identifier, a target computation executor node identifier, a time window, a list of allowed operator types, and a data identifier. The policy description is signed by the key coordination service node using its private key within its trusted execution environment.
5. The method according to claim 1, characterized in that, The key encapsulation mechanism in step (4) uses hybrid public key encryption, with the master key as the initial key material and the policy description as context information to generate a temporary symmetric key; the temporary computing token contains the temporary symmetric key encrypted with the public key of the key collaboration service node, the policy description hash value, and the digital signature of the data owner node.
6. The method according to claim 1, characterized in that, The key collaboration service node re-encrypting the temporary computing token in step (5) specifically includes: the key collaboration service node using its private key to decrypt and obtain the temporary symmetric key in the trusted execution environment, and then using the trusted execution environment public key of the target computing executor node to re-encrypt the temporary symmetric key.
7. The method according to claim 1, characterized in that, The destruction of the function key by the target computation executor node within the trusted execution environment specifically includes: erasing the function key and re-encrypting the token through a secure erasure instruction, and generating a destruction confirmation message containing a policy identifier and a trusted execution environment remote authentication report, which is then sent to the key collaboration service node; the key collaboration service node records the destruction confirmation message in the policy rule database for auditing purposes.
8. The method according to claim 1, characterized in that, It also includes a dynamic rescheduling step: when the CPU utilization of the computing executor node exceeds a preset threshold for a continuous period of time, or the network latency exceeds a preset threshold, or the remote authentication report of the trusted execution environment of the computing executor node expires, the task management and scheduling center triggers rescheduling; if the rescheduling fails, the corresponding atomic operator is rolled back to the data owner node for local execution, and the key coordination service node is notified to revoke the corresponding temporary token.
9. The method according to claim 1, characterized in that, It also includes a token reuse detection step: when the key coordination service node detects that the same policy identifier is requested to be used by multiple computation executor nodes, the key coordination service node generates a revocation list and broadcasts it to all data owner nodes and computation executor nodes through the task management and scheduling center, triggering all affected data owner nodes to regenerate temporary tokens.
10. A system for dynamic offloading and secure collaborative deployment of privacy computing tasks in an asymmetric computing power environment, characterized in that, include: The task management and scheduling center is used to receive privacy computing tasks, parse the tasks into a directed acyclic graph containing multiple atomic operators, collect multi-dimensional resource status information of each participating node in real time, run a multi-objective optimization scheduling algorithm to generate a scheduling mapping table, and trigger rescheduling when a sudden change in node status is detected. Multiple data owner nodes, each holding a master key, are used to generate a temporary computing token through a key encapsulation mechanism according to the policy description issued by the key coordination service node, and send the encrypted computing data and the token to the key coordination service node; Multiple computation executor nodes, each deployed with a trusted execution environment, are used to receive the re-encryption token and encrypted computation data routed by the key coordination service node, decrypt the data within the trusted execution environment to obtain the minimized function key, perform encrypted computation, and destroy the function key after the task is completed. The key coordination service node is deployed with a trusted execution environment. It is used to generate policy descriptions based on the scheduling mapping table and security policy, verify the validity of the temporary token generated by the data owner node, re-encrypt the token in the trusted execution environment and route it to the target computing executor node, and record the entire process of token generation, distribution and destruction.