End-cloud collaborative intention-driven AI host system

Abstract

The invention discloses an end-cloud collaborative intention-driven AI host system, and relates to the technical field of artificial intelligence and cloud computing, and the system comprises an intention semantic tensor mapping module which generates a three-dimensional semantic tensor through semantic analysis and structured conversion of multi-modal input, and achieves semantic disambiguation and structured expression of cross-modal information; the dynamic task unloading module dynamically optimizes unloading paths of tasks on an end side and a cloud side by constructing a Q-Learning reinforcement learning model, and realizes intelligent resource allocation of local execution of low-calculation-amount tasks and encryption of high-calculation-amount tasks and cloud processing; the end-cloud collaborative architecture module is used for realizing calculation load balancing, low-delay response and privacy protection through cloud and end-side collaborative reasoning in combination with a differential privacy synchronization and real-time hot update mechanism; and the privacy computing engine module adopts a double-engine security architecture, and realizes secret-state computing through cooperation of homomorphic encryption and a trusted execution environment. According to the method, high-precision semantic analysis, low-delay response and privacy protection can be realized.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and cloud computing technologies, and more specifically to an end-cloud collaborative intent-driven AI host system. Background Art

[0002] With the rapid development of artificial intelligence technology, AI systems are widely used in many fields such as intelligent interaction, medical diagnosis, and smart home. However, existing technologies have significant defects in intent analysis, real-time response, and privacy protection, making it difficult to meet application requirements in complex scenarios.

[0003] At the level of intent understanding, traditional natural language processing (NLP) models, such as recurrent neural networks (RNN) and their derivative variants LSTM, GRU, etc., mainly rely on single text modal information for command parsing. Due to the lack of cross-modal association capabilities, when faced with complex instructions containing multimodal information such as images, voice, and text, it is impossible to effectively integrate the semantic features of different modalities. For example, in intelligent customer service scenarios, users send voice instructions and graphic supplementary instructions at the same time. Traditional RNN models find it difficult to integrate multimodal information for accurate understanding, resulting in an error rate of more than 30% in parsing complex instructions. This makes it difficult for AI systems to accurately execute user intentions when handling complex tasks in actual applications, seriously affecting the interactive experience and task completion efficiency.

[0004] In terms of real-time response, existing AI systems mostly rely on centralized cloud computing. Take the ResNet-50 model, a commonly used model in the field of image recognition, for example. Due to its large number of model parameters and high computational complexity, terminal devices need to upload data to cloud servers for processing and then receive the results returned by the cloud. This process results in end-to-end latency exceeding 500ms. This delay cannot meet the real-time requirements of the system in scenarios such as autonomous driving and real-time game interaction, which require extremely high response speeds. For example, in autonomous driving scenarios, if the data collected by vehicle sensors undergoes a long cloud processing delay, the vehicle will not be able to respond to sudden road conditions in a timely manner, seriously threatening driving safety.

[0005] Privacy protection has also become a major pain point for existing AI systems. When processing sensitive data such as biometrics (such as fingerprints and facial data) and medical images, traditional AI systems typically require users to upload the data in plain text to the cloud for analysis and processing. This method exposes the data to the risk of being intercepted during transmission, and cloud storage also faces the risk of data leakage. For example, in the medical field, the leakage of sensitive information such as CT images and genetic data of patients will pose a serious threat to their privacy and safety.

[0006] Comparing technologies, pure cloud-based solutions, while offering powerful computing resources and storage capabilities, cannot guarantee real-time performance on the device side. For example, Google's Federated Learning Framework, while achieving some degree of data privacy protection, only supports small-scale models (1.4MB), making it difficult to effectively support the large-scale deep models required for complex tasks and failing to meet high-precision computing requirements. Local compression solutions, such as the lightweight model TinyBERT, use model compression to reduce computational and storage requirements. While this reduces reliance on cloud computing, it comes at the expense of semantic parsing accuracy, with its F1-score dropping to 82%, making it difficult to guarantee accurate task processing in real-world applications. Furthermore, traditional encryption technologies, such as the Advanced Encryption Standard (AES), can only encrypt data for storage and transmission and cannot support ciphertext computation. This makes it difficult to achieve compatibility between privacy protection and functional implementation in AI systems. Encrypted data cannot be directly used for model computation and must be decrypted before processing, exposing the data to privacy risks during the decryption phase.

[0007] In summary, the technical bottlenecks of existing AI systems in intent parsing, real-time response, and privacy protection limit their in-depth application in more key areas. Therefore, an innovative technical solution is urgently needed to achieve high-precision semantic parsing, low-latency response, and end-to-end privacy and security to meet the application requirements of AI systems in complex scenarios. Summary of the Invention

[0008] In response to the shortcomings of existing AI systems in intent analysis, real-time response and privacy protection, the present invention provides an end-cloud collaborative intent-driven AI host system to achieve high-precision semantic analysis, low-latency response and full-link privacy security.

[0009] The present invention provides an end-cloud collaborative intent-driven AI host system, which solves the above technical problems using the following technical solutions:

[0010] An end-cloud collaborative intent-driven AI host system, whose structure includes:

[0011] The intent semantic tensor mapping module generates a three-dimensional semantic tensor containing entities, actions, and spatiotemporal constraints through semantic parsing and structured transformation of multimodal input. The output semantic tensor directly provides task attribute definitions for the dynamic task offloading module, achieving semantic disambiguation and structured expression of cross-modal information, laying a semantic foundation for subsequent task allocation.

[0012] The dynamic task offloading module builds a Q-Learning reinforcement learning model constrained by latency, privacy level, and energy consumption. It dynamically optimizes the task offloading path based on the semantic tensor output by the intent semantic tensor mapping module. Tasks that meet the preset condition a are assigned to local execution, while tasks that meet the preset condition b are encrypted and handed over to the end-cloud collaborative architecture module for processing. At the same time, the privacy computing engine module is called to ensure data security. The offloading strategy is iterated based on the feedback from the privacy computing engine to achieve intelligent resource allocation, ensuring privacy security and energy efficiency.

[0013] The end-cloud collaborative architecture module, based on the task offloading decision results of the dynamic task offloading module, achieves computing load balancing and low-latency response through collaborative reasoning between large cloud models and lightweight end-side models, combined with differential privacy synchronization and real-time hot update mechanisms. When processing tasks requiring privacy protection, the encryption capabilities of the privacy computing engine module are automatically triggered to ensure data security during the end-cloud collaboration process.

[0014] The privacy computing engine module adopts a dual-engine security architecture. When the end-cloud collaborative architecture module calls a sensitive data processing task, it implements confidential computing through the collaboration of homomorphic encryption and a trusted execution environment, and ensures traceability of operations through blockchain evidence. After the calculation is completed, the privacy protection status is fed back to the dynamic task offloading module, providing an optimization basis for energy consumption and security levels for subsequent task allocation strategies, forming an end-to-end sensitive data security closed loop.

[0015] Optionally, the intent semantic tensor mapping module involved specifically includes:

[0016] The multi-source data fusion processing submodule receives multimodal inputs from voice, text, and sensor data, converting the unstructured data into a standardized sequence format. It also processes long sequence dependencies based on the Transformer-XL architecture and captures long-range correlations across modal information through an extended attention mechanism. The output standardized sequence directly provides input with pre-labeled task attributes for the knowledge graph disambiguation submodule, laying the data foundation for the subsequent definition of task attributes in the semantic tensor.

[0017] The knowledge graph disambiguation submodule is used to perform semantic disambiguation in conjunction with the domain knowledge graph. It uses a graph attention network to calculate the association weights between entities and graph nodes, and dynamically selects high-confidence semantic paths. Domain logic constraints are used to ensure the semantic accuracy of entities and actions in the parsing results. The disambiguated semantic elements provide structured input with task attribute priority for the tensor generation submodule, avoiding policy deviations in the dynamic task offloading module due to semantic ambiguity.

[0018] The tensor generation submodule is used to map the parsed semantic elements into three-dimensional tensor dimensions containing entities, actions, and spatiotemporal constraints. The output tensor retains the input information through cross-modal semantic associations and is then integrated into a structured JSON-LD format. The domain ontology is associated through contextual links to ensure that the generated semantic tensor can be directly parsed by the dynamic task offloading module as a decision basis for the task allocation strategy.

[0019] Preferably, the tensor generation submodule maps the parsed semantic elements into three-dimensional tensor dimensions containing entities, actions, and spatiotemporal constraints. This process specifically includes:

[0020] For the entity dimension, the subject and object in the input data are extracted and converted into low-dimensional semantic vectors through BERT word embedding;

[0021] For the action dimension, parse the behavioral instructions in the semantics and generate action encoding based on the domain action library;

[0022] For the spatiotemporal constraint dimension, time, space and conditional constraints are extracted and converted into standardized timestamps, geographic coordinate vectors and Boolean condition matrices.

[0023] Optionally, the dynamic task offloading module involved specifically includes:

[0024] The multi-objective decision model submodule uses the semantic tensor output by the intent semantic tensor mapping module as input, analyzes the task attributes, and constructs the constraints of the Q-Learning reinforcement learning model based on the real-time network status and device energy consumption status. It also dynamically selects the task offloading path based on the objective function of the Q-Learning reinforcement learning model and synchronizes the decision results to the task offloading execution submodule.

[0025] The task offloading execution submodule, based on the results of the multi-objective decision-making model, performs preprocessing on the client side and executes tasks that meet the preset condition a, ensuring local computing security through the trusted execution environment of the privacy computing engine module. It also offloads tasks that meet the preset condition b to the cloud, encrypts the data using the privacy computing engine module's homomorphic encryption, and transmits it to the cloud, inferring the 3DResNet-152 model.

[0026] The result feedback processing sub-module, after receiving the encrypted inference results returned by the cloud, calls the decryption capability of the privacy computing engine module to decrypt them on the end side and displays them through the secure rendering channel; the actual execution results are fed back to the Q-Learning reinforcement learning model, and the weight coefficient of the objective function is dynamically adjusted in combination with the security audit log provided by the privacy computing engine module; the Q-Learning reinforcement learning model with the updated weight coefficient re-determines the task offloading path, and synchronizes the optimized resource allocation strategy to the end-cloud collaborative architecture module to guide it to adjust the collaborative mode between the large cloud model and the lightweight end-side model.

[0027] Preferably, the objective function expression of the Q-Learning reinforcement learning model involved is:

[0028] Minimize: α*T+β*(1 / P)+γ*E,

[0029] Where α, β, and γ are weight coefficients for initial configuration or subsequent calculation updates.

[0030] Optionally, the end-cloud collaborative architecture modules involved include:

[0031] The cloud-side submodule deploys the GPT-4 large model on a GPU cluster. This serves as a deep processing center for multimodal data and handles tasks assigned by the dynamic task offloading module. These tasks include integrating cross-modal data such as text, images, and videos, and outputting structured analysis results. Based on the federated learning framework, it aggregates differential privacy parameters uploaded on the client side and regularly iterates and optimizes the global model. Furthermore, it handles tasks that require large-scale knowledge graphs, returning encrypted results to the client side.

[0032] The device-side submodule runs the MobileBERT-Tiny lightweight model on edge nodes, executes tasks assigned by the dynamic task offloading module, and implements real-time execution of basic intents in offline scenarios. This includes: processing basic instructions; performing real-time noise reduction and normalization on sensor data to generate floating-point feature vectors, and triggering encryption processing in the privacy computing engine based on the privacy level; parsing control instructions and driving peripheral execution, with the results directly fed back to the user.

[0033] The collaborative sub-module dynamically balances the model parameters and computing loads on the cloud and the end side while protecting data privacy through differential privacy synchronization and real-time hot updates. This includes: receiving load balancing instructions from the dynamic task offloading module and adjusting the collaborative strategy between the large cloud model and the lightweight end-side model; using differential privacy technology to compress the end-side model parameters and securely upload them to the cloud for aggregation; and the real-time hot update mechanism dynamically adjusts the accuracy of collaborative reasoning and the strength of privacy protection based on the security audit results of the privacy computing engine module.

[0034] Optionally, the privacy computing engine modules involved specifically include:

[0035] The homomorphic encryption submodule, based on the CKKS homomorphic encryption scheme, performs element-by-element encryption on floating-point feature vectors transmitted by the end-cloud collaborative architecture module to generate a ciphertext tensor structure. It supports matrix multiplication and approximate calculation of nonlinear activation functions in a encrypted state. The encryption process is subject to the privacy level of the dynamic task offloading module. The encrypted data is transmitted to the cloud via a TLS1.3 channel, ensuring that the inference process is executed in a encrypted state throughout.

[0036] The trusted execution submodule builds an independent execution environment based on hardware-level security isolation technology, ensuring that sensitive data collected by the end-side submodule enters the trusted execution environment directly. The integrity of the environment is verified through the Secure Boot mechanism combined with a dynamic root trust chain, enabling full lifecycle key management, including HKDF-based key derivation, TPM-based key storage, and time-based key rotation. Based on the energy consumption constraints of the dynamic task offload module, the optimal decryption algorithm is selected and, upon completion, the actual security level and energy consumption data are fed back to the dynamic task offload module.

[0037] The blockchain evidence storage submodule, based on the Hyperledger Fabric consortium chain architecture, builds an operation record evidence storage system. It receives the decryption result metadata output by the trusted execution submodule, writes it to the ledger after signing it with ECDSA, and organizes the metadata using a Merkle tree structure to support efficient integrity verification. It regularly provides security audit reports to the dynamic task offloading module as a basis for optimizing task offloading strategies. It implements access control through smart contracts and automatically adjusts the evidence storage granularity according to the privacy level requirements of the dynamic task offloading module.

[0038] Further optionally, the specific operations performed by the privacy computing engine module include:

[0039] During data encryption and transmission, the client-side submodule extracts floating-point feature vectors. The homomorphic encryption submodule encrypts the feature vectors using the CKKS scheme and then transmits the encrypted data to the cloud via a TLS 1.3 encrypted channel. During this process, the dynamic task offloading module constrains encryption parameters based on the privacy level.

[0040] During the cloud-based encrypted computing phase, the cloud-based submodule performs deep learning model inference on encrypted data without decryption, and returns the results to the client in encrypted form. During this process, the trusted execution submodule monitors energy consumption and security status, and provides real-time feedback to the dynamic task offloading module.

[0041] In the end-side decryption and evidence storage link, the trusted execution sub-module decrypts the ciphertext inference results returned by the cloud in an independent execution environment. The decrypted analysis result metadata will be signed by the blockchain evidence sub-module using the ECDSA elliptic curve signature algorithm and then written into the Hyperledger Fabric ledger to form a traceable inference record chain. During this process, the blockchain evidence sub-module will regularly provide security audit reports to the dynamic task offloading module, providing a basis for the dynamic task offloading module to optimize the task offloading strategy.

[0042] The present invention provides an intention-driven AI host system with end-cloud collaboration, which has the following beneficial effects compared with the prior art:

[0043] This invention can achieve high-precision semantic parsing, low-latency response, and full-link privacy security. It is suitable for high-security scenarios such as medical diagnosis and financial transactions. It solves the problems of large intention understanding deviation, high terminal latency, and privacy leakage risks in existing technologies, and improves system response efficiency and compliance.

[0044] The present invention improves the accuracy of cross-modal instruction parsing and eliminates semantic ambiguity through the intent semantic tensor mapping module, reduces energy consumption through dynamic task offloading and modular resource allocation, reduces end-side latency and balances computing load through the end-cloud collaborative architecture module, and ensures full-link data security through the privacy computing engine module. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Attachment Figure 1 It is a module connection block diagram of the system described in an embodiment of the present invention. DETAILED DESCRIPTION

[0046] In order to make the technical solution, the technical problems solved and the technical effects of the present invention more clear, the technical solution of the present invention is clearly and completely described below in conjunction with specific embodiments.

[0047] Example:

[0048] Reference Attachment Figure 1 This embodiment proposes an end-cloud collaborative intent-driven AI host system, the structure of which includes:

[0049] The intent semantic tensor mapping module generates a three-dimensional semantic tensor containing entities, actions, and spatiotemporal constraints through semantic parsing and structured transformation of multimodal input. The output semantic tensor directly provides task attribute definitions for the dynamic task offloading module, achieving semantic disambiguation and structured expression of cross-modal information, laying a semantic foundation for subsequent task allocation.

[0050] The dynamic task offloading module builds a Q-Learning reinforcement learning model constrained by latency, privacy level, and energy consumption. It dynamically optimizes the task offloading path based on the semantic tensor output by the intent semantic tensor mapping module. Tasks that meet the preset condition a are assigned to local execution, while tasks that meet the preset condition b are encrypted and handed over to the end-cloud collaborative architecture module for processing. At the same time, the privacy computing engine module is called to ensure data security. The offloading strategy is iterated based on the feedback results of the privacy computing engine to achieve intelligent resource allocation, ensure privacy security, and ensure energy efficiency.

[0051] The end-cloud collaborative architecture module, based on the task offloading decision results of the dynamic task offloading module, achieves computing load balancing and low-latency response through collaborative reasoning between large cloud models and lightweight end-side models, combined with differential privacy synchronization and real-time hot update mechanisms. When processing tasks requiring privacy protection, the encryption capabilities of the privacy computing engine module are automatically triggered to ensure data security during the end-cloud collaboration process.

[0052] The privacy computing engine module adopts a dual-engine security architecture. When the end-cloud collaborative architecture module calls a sensitive data processing task, it implements confidential computing through the collaboration of homomorphic encryption and a trusted execution environment, and ensures traceability of operations through blockchain evidence. After the calculation is completed, the privacy protection status is fed back to the dynamic task offloading module, providing an optimization basis for energy consumption and security levels for subsequent task allocation strategies, forming an end-to-end sensitive data security closed loop.

[0053] It should be noted that a computational complexity threshold is pre-set. When the actual computational complexity threshold of a task is less than the set threshold (i.e., the task meets the preset condition a), the task is called a low-computation task and is subsequently assigned to the local server for execution. When the actual computational complexity threshold of a task is greater than the set threshold (i.e., the task meets the preset condition b), the task is called a high-computation task and is subsequently offloaded to the cloud.

[0054] In this embodiment, the intention semantic tensor mapping module involved specifically includes:

[0055] The multi-source data fusion processing submodule receives multimodal inputs from voice, text, and sensor data, converting the unstructured data into a standardized sequence format. It also processes long sequence dependencies based on the Transformer-XL architecture and captures long-range correlations across modal information through an extended attention mechanism. The output standardized sequence directly provides input with pre-labeled task attributes for the knowledge graph disambiguation submodule, laying the data foundation for the subsequent definition of task attributes in the semantic tensor.

[0056] The knowledge graph disambiguation submodule is used to perform semantic disambiguation in conjunction with the domain knowledge graph. It uses a graph attention network to calculate the association weights between entities and graph nodes, and dynamically selects high-confidence semantic paths. Domain logic constraints are used to ensure the semantic accuracy of entities and actions in the parsing results. The disambiguated semantic elements provide structured input with task attribute priority for the tensor generation submodule, avoiding policy deviations in the dynamic task offloading module due to semantic ambiguity.

[0057] The tensor generation submodule is used to map the parsed semantic elements into three-dimensional tensor dimensions containing entities, actions, and spatiotemporal constraints. The output tensor retains the input information through cross-modal semantic associations and is then integrated into a structured JSON-LD format. The domain ontology is associated through contextual links to ensure that the generated semantic tensor can be directly parsed by the dynamic task offloading module as a decision basis for the task allocation strategy.

[0058] The tensor generation submodule involved maps the parsed semantic elements into three-dimensional tensor dimensions containing entities, actions, and spatiotemporal constraints. This process specifically includes:

[0059] For the entity dimension, the subject and object in the input data are extracted and converted into low-dimensional semantic vectors through BERT word embedding;

[0060] For the action dimension, parse the behavioral instructions in the semantics and generate action encoding based on the domain action library;

[0061] For the spatiotemporal constraint dimension, time, space and conditional constraints are extracted and converted into standardized timestamps, geographic coordinate vectors and Boolean condition matrices.

[0062] In this embodiment, the dynamic task offloading module specifically includes:

[0063] The multi-objective decision model submodule uses the semantic tensor output by the intent semantic tensor mapping module as input, analyzes the task attributes, and constructs the constraints of the Q-Learning reinforcement learning model based on the real-time network status and device energy consumption status. It also dynamically selects the task offloading path based on the objective function of the Q-Learning reinforcement learning model and synchronizes the decision results to the task offloading execution submodule.

[0064] The task offloading execution submodule, based on the results of the multi-objective decision-making model, performs preprocessing on the client side and executes tasks that meet the preset condition a (i.e., low-computation tasks), ensuring local computing security through the trusted execution environment of the privacy computing engine module. It also offloads tasks that meet the preset condition b (i.e., high-computation tasks) to the cloud, encrypts the data using the privacy computing engine module's homomorphic encryption, and transmits it to the cloud, inferring the 3DResNet-152 model.

[0065] The result feedback processing sub-module, after receiving the encrypted inference results returned by the cloud, calls the decryption capability of the privacy computing engine module to decrypt them on the end side and displays them through the secure rendering channel; the actual execution results are fed back to the Q-Learning reinforcement learning model, and the weight coefficient of the objective function is dynamically adjusted in combination with the security audit log provided by the privacy computing engine module; the Q-Learning reinforcement learning model with the updated weight coefficient re-determines the task offloading path, and synchronizes the optimized resource allocation strategy to the end-cloud collaborative architecture module to guide it to adjust the collaborative mode between the large cloud model and the lightweight end-side model.

[0066] The objective function expression of the Q-Learning reinforcement learning model involved is:

[0067] Minimize: α*T+β*(1 / P)+γ*E,

[0068] Where α, β, and γ are weight coefficients for initial configuration or subsequent calculation updates.

[0069] In this embodiment, the end-cloud collaboration architecture modules involved specifically include:

[0070] The cloud-side submodule deploys the GPT-4 large model on a GPU cluster. This serves as a deep processing center for multimodal data and handles tasks assigned by the dynamic task offloading module. These tasks include integrating cross-modal data such as text, images, and videos, and outputting structured analysis results. Based on the federated learning framework, it aggregates differential privacy parameters uploaded on the client side and regularly iterates and optimizes the global model. Furthermore, it handles tasks that require large-scale knowledge graphs, returning encrypted results to the client side.

[0071] The device-side submodule runs the MobileBERT-Tiny lightweight model on edge nodes, executes tasks assigned by the dynamic task offloading module, and implements real-time execution of basic intents in offline scenarios. This includes: processing basic instructions; performing real-time noise reduction and normalization on sensor data to generate floating-point feature vectors, and triggering encryption processing in the privacy computing engine based on the privacy level; parsing control instructions and driving peripheral execution, with the results directly fed back to the user.

[0072] The collaborative sub-module dynamically balances the model parameters and computing loads on the cloud and the end side while protecting data privacy through differential privacy synchronization and real-time hot updates. This includes: receiving load balancing instructions from the dynamic task offloading module and adjusting the collaborative strategy between the large cloud model and the lightweight end-side model; using differential privacy technology to compress the end-side model parameters and securely upload them to the cloud for aggregation; and the real-time hot update mechanism dynamically adjusts the accuracy of collaborative reasoning and the strength of privacy protection based on the security audit results of the privacy computing engine module.

[0073] In this embodiment, the privacy computing engine module involved specifically includes:

[0074] The homomorphic encryption submodule, based on the CKKS homomorphic encryption scheme, performs element-by-element encryption on floating-point feature vectors transmitted by the end-cloud collaborative architecture module to generate a ciphertext tensor structure. It supports matrix multiplication and approximate calculation of nonlinear activation functions in a encrypted state. The encryption process is subject to the privacy level of the dynamic task offloading module. The encrypted data is transmitted to the cloud via a TLS1.3 channel, ensuring that the inference process is executed in a encrypted state throughout.

[0075] The trusted execution submodule builds an independent execution environment based on hardware-level security isolation technology, ensuring that sensitive data collected by the end-side submodule enters the trusted execution environment directly. The integrity of the environment is verified through the Secure Boot mechanism combined with a dynamic root trust chain, enabling full lifecycle key management, including HKDF-based key derivation, TPM-based key storage, and time-based key rotation. Based on the energy consumption constraints of the dynamic task offload module, the optimal decryption algorithm is selected and, upon completion, the actual security level and energy consumption data are fed back to the dynamic task offload module.

[0076] The blockchain evidence storage submodule, based on the Hyperledger Fabric consortium chain architecture, builds an operation record evidence storage system. It receives the decryption result metadata output by the trusted execution submodule, writes it to the ledger after signing it with ECDSA, and organizes the metadata using a Merkle tree structure to support efficient integrity verification. It regularly provides security audit reports to the dynamic task offloading module as a basis for optimizing task offloading strategies. It implements access control through smart contracts and automatically adjusts the evidence storage granularity according to the privacy level requirements of the dynamic task offloading module.

[0077] The specific operations performed by the privacy computing engine module include:

[0078] During data encryption and transmission, the client-side submodule extracts floating-point feature vectors. The homomorphic encryption submodule encrypts the feature vectors using the CKKS scheme and then transmits the encrypted data to the cloud via a TLS 1.3 encrypted channel. During this process, the dynamic task offloading module constrains encryption parameters based on the privacy level.

[0079] During the cloud-based encrypted computing phase, the cloud-based submodule performs deep learning model inference on encrypted data without decryption, and returns the results to the client in encrypted form. During this process, the trusted execution submodule monitors energy consumption and security status, and provides real-time feedback to the dynamic task offloading module.

[0080] In the end-side decryption and evidence storage link, the trusted execution sub-module decrypts the ciphertext inference results returned by the cloud in an independent execution environment. The decrypted analysis result metadata will be signed by the blockchain evidence sub-module using the ECDSA elliptic curve signature algorithm and then written into the Hyperledger Fabric ledger to form a traceable inference record chain. During this process, the blockchain evidence sub-module will regularly provide security audit reports to the dynamic task offloading module, providing a basis for the dynamic task offloading module to optimize the task offloading strategy.

[0081] Based on the detailed description of the embodiment of the medical imaging diagnosis workflow, the technical implementation and interaction logic of the modules described in this embodiment are described:

[0082] 1. The doctor speaks instructions, which are converted into text by the speech recognition module and transmitted to the intent semantic tensor mapping module;

[0083] 2. In the intent semantic tensor mapping module, the multi-source data fusion processing submodule uses Transformer-XL to extract keywords and convert them into standardized sequences with task attributes. The knowledge graph disambiguation submodule associates the SNOMED-CT atlas and generates high-confidence semantic paths after disambiguation. The tensor generation submodule constructs a three-dimensional semantic tensor containing entities, actions, and spatiotemporal constraints, converts it into JSON-LD, and associates it with the medical ontology for decision-making by the dynamic task offloading module.

[0084] 3. The dynamic task offloading module parses the semantic tensor and combines it with the end-side resources.

[0085] i) Decision-making for low-computation tasks (such as CT image window width adjustment and ROI segmentation) is performed on the device side: Instructions are sent to the device-cloud collaborative architecture module. The device-side submodule of the device-cloud collaborative architecture module adjusts the CT window width in an independent execution environment, segments nodules using the lightweight MobileBERT-Tiny, and generates feature vectors that are transmitted to the privacy computing engine module for encryption. The device-side submodule then reports resource consumption to the dynamic task offloading module.

[0086] ii) Offloading high-computation tasks (such as malignancy classification and gene sequence analysis) to the cloud: The offloading path is optimized using the Q-Learning reinforcement learning model. The feature vector is encrypted using the homomorphic encryption submodule (CKKS scheme) of the privacy computing engine and then transmitted to the cloud submodule of the end-cloud collaborative architecture module via the TLS 1.3 channel. The cloud submodule relies on the GPT-4 large model deployed on the GPU cluster to perform deep learning inference in a secure state and generate malignancy probability prediction results. After the cloud-based inference is completed, the encrypted results are returned to the end-side, received by the end-side submodule of the end-cloud collaborative architecture module, and the cloud computing resource usage (such as GPU time consumption and network latency) and encryption transmission status are simultaneously reported to the dynamic task offloading module.

[0087] 4. The homomorphic encryption submodule of the privacy computing engine module encrypts the feature vector using CKKS and transmits it to the cloud submodule via TLS1.3. The cloud submodule performs dense inference using the GPT-4 large model, generates probabilities, and adds differential privacy noise. The blockchain evidence storage submodule of the privacy computing engine module signs the inference metadata and writes it to the Hyperledger Fabric ledger. The cloud submodule reports cloud resource consumption to the dynamic task offloading module.

[0088] 5. The trusted execution submodule of the privacy computing engine module decrypts the results in an independent execution environment, verifies the integrity, and generates diagnostic recommendations. The end-side submodule of the end-cloud collaborative architecture module displays the results through a secure rendering channel and transmits the diagnostic metadata to the blockchain for storage. The blockchain storage submodule uploads the results to the chain, generates a hash value, and synchronizes it to the dynamic task offloading module. The trusted execution submodule feeds back the decryption energy consumption and security level to the dynamic task offloading module for optimizing subsequent strategies.

[0089] In summary, the end-cloud collaborative intent-driven AI host system of the present invention can achieve high-precision semantic analysis, low-latency response and full-link privacy security, and solve the problems of large intention understanding deviation, high end-side latency, and privacy leakage risk in the existing technology.

[0090] The above specific examples are used to illustrate the principles and implementation methods of the present invention in detail. These examples are only used to help understand the core technical content of the present invention. Based on the above specific embodiments of the present invention, any improvements and modifications made by those skilled in the art without departing from the principles of the present invention should fall within the scope of patent protection of the present invention.

Claims

1. An end-cloud collaborative intent-driven AI host system, characterized by: Its structure includes: The intent semantic tensor mapping module generates a three-dimensional semantic tensor containing entities, actions, and spatiotemporal constraints through semantic parsing and structured transformation of multimodal input. The output semantic tensor directly provides task attribute definitions for the dynamic task offloading module, achieving semantic disambiguation and structured expression of cross-modal information, laying a semantic foundation for subsequent task allocation. The dynamic task offloading module builds a Q-Learning reinforcement learning model constrained by latency, privacy level, and energy consumption. It dynamically optimizes the task offloading path based on the semantic tensor output by the intent semantic tensor mapping module. Tasks that meet the preset condition a are assigned to local execution, while tasks that meet the preset condition b are encrypted and handed over to the end-cloud collaborative architecture module for processing. At the same time, the privacy computing engine module is called to ensure data security. The offloading strategy is iterated based on the feedback from the privacy computing engine to achieve intelligent resource allocation, ensuring privacy security and energy efficiency. The end-cloud collaborative architecture module, based on the task offloading decision results of the dynamic task offloading module, achieves computing load balancing and low-latency response through collaborative reasoning between large cloud models and lightweight end-side models, combined with differential privacy synchronization and real-time hot update mechanisms. When processing tasks requiring privacy protection, the encryption capabilities of the privacy computing engine module are automatically triggered to ensure data security during the end-cloud collaboration process. The privacy computing engine module adopts a dual-engine security architecture. When the end-cloud collaborative architecture module calls a sensitive data processing task, it implements confidential computing through the collaboration of homomorphic encryption and a trusted execution environment, and ensures traceability of operations through blockchain evidence. After the calculation is completed, the privacy protection status is fed back to the dynamic task offloading module, providing an optimization basis for energy consumption and security levels for subsequent task allocation strategies, forming an end-to-end sensitive data security closed loop.

2. The end-cloud collaborative intent-driven AI host system according to claim 1, characterized in that: The intention semantic tensor mapping module specifically includes: The multi-source data fusion processing submodule receives multimodal inputs from voice, text, and sensor data, converting the unstructured data into a standardized sequence format. It also processes long sequence dependencies based on the Transformer-XL architecture and captures long-range correlations across modal information through an extended attention mechanism. The output standardized sequence directly provides input with pre-labeled task attributes for the knowledge graph disambiguation submodule, laying the data foundation for the subsequent definition of task attributes in the semantic tensor. The knowledge graph disambiguation submodule is used to perform semantic disambiguation in conjunction with the domain knowledge graph. It uses a graph attention network to calculate the association weights between entities and graph nodes, and dynamically selects high-confidence semantic paths. Domain logic constraints are used to ensure the semantic accuracy of entities and actions in the parsing results. The disambiguated semantic elements provide structured input with task attribute priority for the tensor generation submodule, avoiding policy deviations in the dynamic task offloading module due to semantic ambiguity. The tensor generation submodule is used to map the parsed semantic elements into three-dimensional tensor dimensions containing entities, actions, and spatiotemporal constraints. The output tensor retains the input information through cross-modal semantic associations and is then integrated into a structured JSON-LD format. The domain ontology is associated through contextual links to ensure that the generated semantic tensor can be directly parsed by the dynamic task offloading module as a decision basis for the task allocation strategy.

3. The end-cloud collaborative intent-driven AI host system according to claim 2, characterized in that: The tensor generation submodule maps the parsed semantic elements into three-dimensional tensor dimensions containing entities, actions, and spatiotemporal constraints. This process specifically includes: For the entity dimension, the subject and object in the input data are extracted and converted into low-dimensional semantic vectors through BERT word embedding; For the action dimension, parse the behavioral instructions in the semantics and generate action encoding based on the domain action library; For the spatiotemporal constraint dimension, time, space and conditional constraints are extracted and converted into standardized timestamps, geographic coordinate vectors and Boolean condition matrices.

4. The end-cloud collaborative intent-driven AI host system according to claim 1, characterized in that: The dynamic task offloading module specifically includes: The multi-objective decision model submodule uses the semantic tensor output by the intent semantic tensor mapping module as input, analyzes the task attributes, and constructs the constraints of the Q-Learning reinforcement learning model based on the real-time network status and device energy consumption status. It also dynamically selects the task offloading path based on the objective function of the Q-Learning reinforcement learning model and synchronizes the decision results to the task offloading execution submodule. The task offloading execution submodule, based on the results of the multi-objective decision-making model, performs preprocessing on the client side and executes tasks that meet the preset condition a, ensuring local computing security through the trusted execution environment of the privacy computing engine module. It also offloads tasks that meet the preset condition b to the cloud, encrypts the data using the privacy computing engine module's homomorphic encryption, and transmits it to the cloud, inferring the 3DResNet-152 model. The result feedback processing sub-module, after receiving the encrypted inference results returned by the cloud, calls the decryption capability of the privacy computing engine module to decrypt them on the end side and displays them through the secure rendering channel; the actual execution results are fed back to the Q-Learning reinforcement learning model, and the weight coefficient of the objective function is dynamically adjusted in combination with the security audit log provided by the privacy computing engine module; the Q-Learning reinforcement learning model with the updated weight coefficient re-determines the task offloading path, and synchronizes the optimized resource allocation strategy to the end-cloud collaborative architecture module to guide it to adjust the collaborative mode between the large cloud model and the lightweight end-side model.

5. The end-cloud collaborative intent-driven AI host system according to claim 4, characterized in that: The objective function expression of the Q-Learning reinforcement learning model is: Minimize: α*T+β*(1 / P)+γ*E, Where α, β, and γ are weight coefficients for initial configuration or subsequent calculation updates.

6. The end-cloud collaborative intent-driven AI host system according to claim 1, characterized in that: The end-cloud collaborative architecture module specifically includes: The cloud-side submodule deploys the GPT-4 large model on a GPU cluster. This serves as a deep processing center for multimodal data and handles tasks assigned by the dynamic task offloading module. These tasks include integrating cross-modal data such as text, images, and videos, and outputting structured analysis results. Based on the federated learning framework, it aggregates differential privacy parameters uploaded on the client side and regularly iterates and optimizes the global model. Furthermore, it handles tasks that require large-scale knowledge graphs, returning encrypted results to the client side. The device-side submodule runs the MobileBERT-Tiny lightweight model on edge nodes, executes tasks assigned by the dynamic task offloading module, and implements real-time execution of basic intents in offline scenarios. This includes: processing basic instructions; performing real-time noise reduction and normalization on sensor data to generate floating-point feature vectors, and triggering encryption processing in the privacy computing engine based on the privacy level; parsing control instructions and driving peripheral execution, with the results directly fed back to the user. The collaborative sub-module dynamically balances the model parameters and computing loads on the cloud and the end side while protecting data privacy through differential privacy synchronization and real-time hot updates. This includes: receiving load balancing instructions from the dynamic task offloading module and adjusting the collaborative strategy between the large cloud model and the lightweight end-side model; using differential privacy technology to compress the end-side model parameters and securely upload them to the cloud for aggregation; and the real-time hot update mechanism dynamically adjusts the accuracy of collaborative reasoning and the strength of privacy protection based on the security audit results of the privacy computing engine module.

7. The end-cloud collaborative intent-driven AI host system according to claim 1, characterized in that: The privacy computing engine module specifically includes: The homomorphic encryption submodule, based on the CKKS homomorphic encryption scheme, performs element-by-element encryption on floating-point feature vectors transmitted by the end-cloud collaborative architecture module to generate a ciphertext tensor structure. It supports matrix multiplication and approximate calculation of nonlinear activation functions in a encrypted state. The encryption process is subject to the privacy level of the dynamic task offloading module. The encrypted data is transmitted to the cloud via a TLS1.3 channel, ensuring that the inference process is executed in a encrypted state throughout. The trusted execution submodule builds an independent execution environment based on hardware-level security isolation technology, ensuring that sensitive data collected by the end-side submodule enters the trusted execution environment directly. The integrity of the environment is verified through the Secure Boot mechanism combined with a dynamic root trust chain, enabling full lifecycle key management, including HKDF-based key derivation, TPM-based key storage, and time-based key rotation. Based on the energy consumption constraints of the dynamic task offload module, the optimal decryption algorithm is selected and, upon completion, the actual security level and energy consumption data are fed back to the dynamic task offload module. The blockchain evidence storage submodule, based on the Hyperledger Fabric consortium chain architecture, builds an operation record evidence storage system. It receives the decryption result metadata output by the trusted execution submodule, writes it to the ledger after signing it with ECDSA, and organizes the metadata using a Merkle tree structure to support efficient integrity verification. It regularly provides security audit reports to the dynamic task offloading module as a basis for optimizing task offloading strategies. It implements access control through smart contracts and automatically adjusts the evidence storage granularity according to the privacy level requirements of the dynamic task offloading module.

8. The end-cloud collaborative intent-driven AI host system according to claim 7, characterized in that: The specific operations performed by the privacy computing engine module include: During data encryption and transmission, the client-side submodule extracts floating-point feature vectors. The homomorphic encryption submodule encrypts the feature vectors using the CKKS scheme and then transmits the encrypted data to the cloud via a TLS 1.3 encrypted channel. During this process, the dynamic task offloading module constrains encryption parameters based on the privacy level. During the cloud-based encrypted computing phase, the cloud-based submodule performs deep learning model inference on encrypted data without decryption, and returns the results to the client in encrypted form. During this process, the trusted execution submodule monitors energy consumption and security status, and provides real-time feedback to the dynamic task offloading module. In the end-side decryption and evidence storage link, the trusted execution sub-module decrypts the ciphertext inference results returned by the cloud in an independent execution environment. The decrypted analysis result metadata will be signed by the blockchain evidence sub-module using the ECDSA elliptic curve signature algorithm and then written into the Hyperledger Fabric ledger to form a traceable inference record chain. During this process, the blockchain evidence sub-module will regularly provide security audit reports to the dynamic task offloading module, providing a basis for the dynamic task offloading module to optimize the task offloading strategy.

Images