Method and device for secure processing of embodied intelligence training data, medium and equipment

Abstract

The application discloses a kind of body intelligence practical training data security processing method, device, medium and equipment, method includes: obtaining operation semantic graph, identify the operation stage that body intelligence equipment currently is in, operation semantic graph includes task node, operation node, data node, the dependence weight between operation node and data node;From operation semantic graph, query current operation data set under operation stage, set includes the dependence weight between each operation node and each data node under operation stage;For each data frame in operation stage, according to the dependence weight between operation node and each data node under data frame, and the change rate of data frame, calculate the sensitivity of operation node to each data node under data frame, according to the sensitivity of operation node to each data node under data frame, generate time sequence sensitivity vector under data frame;Real-time sensor data corresponding to target sensitivity in time sequence sensitivity vector is encrypted.

Description

Technical Field

[0001] This application relates to the field of data security technology, and more specifically, to a method, apparatus, medium, and equipment for securely processing embodied intelligent training data. Background Technology

[0002] With the rapid development of embodied intelligence technology, the training tasks for robots in real physical environments are becoming increasingly complex, and the scale and sensitivity of the data collected during training are rising dramatically. Training data typically encompasses multimodal perception information, including but not limited to RGB images, depth maps, point clouds, force / torque sequences, joint states, and operational trajectories. This data not only records the physical layout and visual features of the experimental scene but also contains the robot's operational skill parameters, motion control strategies, and task execution logic, possessing high scene privacy attributes and core R&D value.

[0003] The training data contains scenario privacy and core operating parameters, which face multiple security risks such as leakage, tampering and misuse. Existing technologies mainly use a globally unified encryption method to protect the full amount of data. However, this method has the problem of coarse encryption granularity, which leads to redundant computational overhead. Summary of the Invention

[0004] This application provides a secure processing method, apparatus, medium, and device for embodied intelligent training data, which can solve the problem of redundant computational overhead caused by coarse encryption granularity when using a globally unified encryption method to protect the full amount of data.

[0005] The specific technical solution is as follows: In a first aspect, embodiments of this application provide a secure processing method for embodied intelligence training data, the method comprising: The operation semantic graph is obtained and the current operation stage of the embodied intelligent device is identified. The operation semantic graph is established based on historical embodied intelligent training task data. The operation semantic graph includes task nodes, operation nodes, data nodes, and dependency weights labeled on the edges connecting each operation node and each data node. The dependency weights are used to represent the degree of dependence of each operation node on each data node. Query the current operation data set under the operation stage from the operation semantic graph. The current operation data set includes the dependency weights between each operation node and each data node in the operation stage. For each data frame in the operation phase, the sensitivity of the operation node to each data node in the data frame is calculated based on the dependency weight between the operation node and each data node in the data frame and the rate of change of the data frame. Based on the sensitivity of the operation node to each data node in the data frame, a time-series sensitivity vector is generated for the data frame. Each time-series sensitivity vector in the operation phase is traversed, and the real-time sensor data corresponding to the target sensitivity in the time-series sensitivity vector is encrypted. The target sensitivity is the sensitivity in the time-series sensitivity vector that is greater than or equal to its own preset sensitivity threshold.

[0006] In one possible implementation, the real-time sensor data corresponding to the target sensitivity in the time-series sensitivity vector is encrypted, including: Calculate the semantic distance between each region in the real-time sensor data corresponding to the target sensitivity and the operation target of the operation stage, wherein each region includes at least a local region of an object; The semantic distance is processed based on the decay function to obtain the spatial encryption strength of each region; The corresponding region in the real-time sensor data is encrypted based on the spatial encryption strength, wherein the spatial encryption strength is positively correlated with the encryption level of the region.

[0007] In one possible implementation, calculating the semantic distance between each region in the real-time sensor data corresponding to the target sensitivity and the operational target of the operational phase includes: Calculate the physical spatial distance between the center point of each region in the real-time sensor data corresponding to the target sensitivity and the center of the operation target in the operation phase; The semantic isolation factor for each region is determined by judging whether each region belongs to the same physical rigid body as the operation target; For each region, the semantic distance is determined based on the physical spatial distance and the semantic isolation factor.

[0008] In one possible implementation, the operation semantic graph further includes: a sensitivity level labeled for the edge connecting each operation node to each data node; The preset sensitivity threshold is determined based on the sensitivity level.

[0009] In one possible implementation, the method further includes: Copy the real-time sensor data generated during the aforementioned operation phase to obtain real-time sensor backup data; The target dependency weight is queried from the current operation data set, and the real-time sensor backup data corresponding to the target dependency weight is determined as the data to be desensitized. The target dependency weight is a dependency weight that is greater than or equal to a preset weight threshold. The data to be de-identified is smoothed using statistical mean to obtain the de-identified data.

[0010] Secondly, embodiments of this application provide a secure processing device for embodied intelligent training data, the device comprising: The acquisition unit is used to acquire an operation semantic graph, which is established based on historical embodied intelligence training task data. The operation semantic graph includes task nodes, operation nodes, data nodes, and dependency weights labeled on the edges connecting each operation node and each data node. The dependency weights are used to represent the degree of dependence of each operation node on each data node. The identification unit is used to identify the current operating stage of the embodied intelligent device; The query unit is used to query the current operation data set under the operation stage from the operation semantic graph. The current operation data set includes the dependency weights between each operation node and each data node included in the operation stage. The calculation unit is used to calculate the sensitivity of the operation node to each data node in the data frame for each data frame in the operation phase, based on the dependency weight between the operation node and each data node in the data frame and the rate of change of the data frame. The generation unit is used to generate a timing sensitivity vector for the data frame based on the sensitivity of the operation node to each data node under the data frame. An encryption unit is used to traverse each time-series sensitivity vector in the operation phase and encrypt the real-time sensor data corresponding to the target sensitivity in the time-series sensitivity vector. The target sensitivity is the sensitivity in the time-series sensitivity vector that is greater than or equal to its own preset sensitivity threshold.

[0011] In one possible implementation, the encryption unit includes: The calculation module is used to calculate the semantic distance between each region in the real-time sensor data corresponding to the target sensitivity and the operation target of the operation stage, wherein each region includes at least a local region of an object; The attenuation module is used to process the semantic distance based on the attenuation function to obtain the spatial encryption strength of each region; An encryption module is used to encrypt the corresponding area in the real-time sensor data based on the spatial encryption strength, wherein the spatial encryption strength is positively correlated with the encryption level of the area.

[0012] In one possible implementation, the calculation module is used to calculate the physical spatial distance between the center point of each region in the real-time sensor data corresponding to the target sensitivity and the center of the operation target in the operation phase; determine the semantic isolation factor of each region by determining whether each region and the operation target belong to the same physical rigid body; and determine the semantic distance for each region based on the physical spatial distance and the semantic isolation factor.

[0013] In one possible implementation, the operation semantic graph further includes: a sensitivity level labeled for the edge connecting each operation node to each data node; The preset sensitivity threshold is determined based on the sensitivity level.

[0014] In one possible implementation, the device further includes: The copying unit is used to copy the real-time sensor data generated during the operation phase to obtain real-time sensor backup data. The query unit is further configured to query a target dependency weight from the current operation data set, wherein the target dependency weight is a dependency weight that is greater than or equal to a preset weight threshold. The determining unit is used to determine the real-time sensor backup data corresponding to the target dependency weight as the data to be desensitized; The smoothing unit is used to smooth the data to be de-identified using a statistical mean to obtain the de-identified data.

[0015] Thirdly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method as described in any possible implementation of the first aspect.

[0016] Fourthly, embodiments of this application provide an electronic device, which includes: One or more processors; The processor is coupled to a storage device for storing one or more programs; When one or more programs are executed by one or more processors, the electronic device performs the method as described in any possible implementation of the first aspect.

[0017] Fifthly, embodiments of this application provide a computer program product containing instructions that, when executed on a computer or processor, cause the computer or processor to perform the method described in any possible implementation of the first aspect.

[0018] The secure processing method, apparatus, medium, and equipment for embodied intelligent training data provided in this application abandon the one-size-fits-all protection model and instead construct a two-dimensional evaluation mechanism that integrates static prior knowledge and dynamic real-time status. Specifically, this mechanism, on the one hand, utilizes an operational semantic graph constructed from historical experience to quantify the basic dependence of different operational stages on various types of data; on the other hand, it captures instantaneous sensitive features in the physical interaction process by sensing the rate of change of the current data frame. The temporal sensitivity vector generated by the two collaboratively can accurately identify the core data elements that truly need protection at each moment. Based on this, this application embodiment only encrypts a small amount of critical data whose sensitivity exceeds the threshold, while allowing the vast majority of non-sensitive or redundant data to pass through in plaintext. Ultimately, this application embodiment significantly reduces the overall encryption computation burden of the system while effectively curbing the risk of data leakage and tampering, while ensuring the availability of training data and the real-time nature of the processing flow, achieving a reasonable balance between security protection and computational efficiency.

[0019] Furthermore, the technical effects achieved by the embodiments of this application may also include: 1. Through an additional spatial semantic awareness mechanism, this embodiment of the application can adaptively allocate differentiated encryption strengths within each frame of data that has been determined to be sensitive, based on the semantic distance between each region and the operation target: the core operation region remains weakly encrypted or plaintext is available, while the peripheral background region is protected by strong encryption. This mechanism achieves fine-grained hierarchical protection at the intra-frame spatial granularity, effectively compressing the computational area of ​​high-level encryption while ensuring visual servoing and control accuracy, reducing the overall encryption overhead of the system, and further improving the synergy between security protection and computational efficiency.

[0020] 2. In this embodiment, the sensitivity levels pre-marked in the operation semantic graph are directly used to determine the preset sensitivity thresholds corresponding to the data features under each operation node, so that the judgment criteria of the temporal sensitivity vector are aligned with the sensitivity level of the data itself, thereby improving the accuracy and rationality of the adaptive encryption process.

[0021] 3. In this embodiment, an independent backup is obtained by copying the real-time sensor data of the current operation stage. Target data with a weight higher than the preset weight threshold is selected based on the dependency weight in the operation semantic graph. Mean smoothing desensitization processing is performed only on the features closely related to the core operation in the backup. Thus, the permanent blurring of highly dependent and sensitive features is achieved without affecting the integrity of the original encrypted data, effectively preventing the accurate restoration and misuse of core process parameters. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0023] Figure 1 A flowchart illustrating a secure processing method for embodied intelligence training data provided in an embodiment of this application; Figure 2 This is a block diagram of a secure processing device for embodied intelligent training data provided in an embodiment of this application. Detailed Implementation

[0024] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0025] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The terms "comprising" and "having," and any variations thereof, in the embodiments and drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0026] This application presents a flowchart illustrating a secure processing method for embodied intelligent training data. This method can be applied to electronic or computer equipment, including but not limited to: embodied intelligent device bodies (such as industrial robotic arm control cabinets, humanoid robot onboard computing units), training data management servers (such as local laboratory servers or enterprise private cloud servers), edge computing devices (such as AI edge gateways, intelligent camera-related processors), general-purpose computer equipment (such as workstations or high-performance PCs for offline data analysis), and cloud computing platforms (such as virtual machines or container clusters deployed in public or private clouds). Figure 1 As shown, the method may include the following steps: S110: Acquire the operation semantic graph and identify the current operation stage of the embodied intelligent device.

[0027] The operation semantic graph is built based on historical embodied intelligence training task data. The operation semantic graph includes task nodes, operation nodes, data nodes, and dependency weights labeled on the edges connecting each operation node and each data node. The dependency weights are used to represent the degree of dependency of each operation node on each data node.

[0028] Historical embodied intelligence training task data includes multimodal sensor data, such as images, LiDAR point clouds, force sensing, and joint angles. Task nodes refer to the operational tasks of the embodied intelligent device (belonging to the top-level objectives), such as making an Americano. Operation nodes refer to the specific physical actions that perform the task, such as grasping a cup, moving it to the coffee machine, or pressing the dispensing button. Data nodes are the specific multimodal data instances that support each operation, such as the RGB depth map of the cup, the force feedback curve of the robotic arm's end effector, and the thermal image of the water outlet.

[0029] Dependency weights can be calculated using causal inference algorithms, including but not limited to Structural Causal Models (SCM) and Granger Causality Tests (GCT). Specific implementation logic may include conducting intervention experiments in multiple training sessions. For example, artificially injecting noise into data nodes (modifying the depth map or delaying force feedback) and observing the success rate of the operation nodes. For instance, if modifying the cup depth map causes the cup-grabbing success rate to plummet from 98% to 10%, it indicates a strong causal dependency of the grabbing action on the cup depth map, and the dependency weight should be assigned a higher value. Similarly, if modifying the water outlet temperature map doesn't affect the grabbing success rate but significantly impacts the water-dispensing success rate, then the grabbing action doesn't have a strong causal dependency on the water outlet temperature map, and the dependency weight should be assigned a lower value. However, the water-dispensing action has a strong causal dependency on the water outlet temperature map, and the dependency weight should be assigned a higher value.

[0030] The operation phase is a physical interaction sub-interval divided on a continuous time axis during the execution of training tasks by the embodied intelligent device; each operation phase corresponds to one or more operation nodes in the operation semantic graph, and the current operation phase of the embodied intelligent device is determined in real time by identifying the sensor data characteristics within the current time window during runtime.

[0031] In the specific implementation, multimodal sensor data streams from the embodied intelligent device are collected in real time, including body perception data and external environment perception data. A pre-trained temporal action segmentation classifier is used to perform sliding window inference on the data stream to detect boundary points between various physical interaction states, thereby dividing the continuous data stream into candidate intervals of operation stages with clear start and end times. Based on this, a corresponding stage label is assigned to each candidate interval by the classifier. The stage label is used to index the operation node with the same name in the operation semantic graph, thereby determining the operation stage that the embodied intelligent device is in at the current moment and its corresponding time window.

[0032] It should be noted that operation nodes and operation stages are not the same concept. The difference lies in the following: an operation node is a statically defined unit in the operation semantic graph, used to represent the type of operation skill and its data dependencies; an operation stage is a physical execution segment based on time-series data streams during the actual operation of an embodied intelligent device, possessing temporal attributes and real-time context. An operation stage identifies the type of action currently being executed, indexes the corresponding operation node in the operation graph, and thus obtains the data features and their dependency weights that need to be considered at that stage.

[0033] S120: Query the current operation data set under the operation stage from the operation semantic graph.

[0034] The current operational dataset includes the dependency weights between each operational node and each data node within the operational phase.

[0035] After identifying the current operation stage and its corresponding stage label, the operation semantic graph is searched using the stage label as the query keyword to match the operation node in the graph that corresponds to the stage label. Then, the edges connecting the operation node to each data node in the operation semantic graph are read, and the pre-stored dependency weight values ​​on each edge are extracted to obtain a set of dependency weights of the operation node to all associated data nodes. This set serves as the current operation data set under the operation stage and can be used as the data filtering basis for subsequent time-series sensitivity calculation.

[0036] S130: For each data frame in the operation phase, calculate the sensitivity of the operation node to each data node in the data frame based on the dependency weight between the operation node and each data node in the data frame, as well as the rate of change of the data frame, and generate the time sensitivity vector of the data frame based on the sensitivity of the operation node to each data node in the data frame.

[0037] Sensitivity = Dependency Weight * Rate of Change of Data Frame. The rate of change of data frame is the gradient or information entropy of the real-time sensor data of each feature dimension under that data frame at that moment. This feature dimension corresponds to the data dimension represented by the data node.

[0038] For each data frame, the dependency weight of the operation node on each data node in the data frame is multiplied by the data change rate of the corresponding data node at this time to obtain the sensitivity of the operation node to each data node, and all sensitivity values ​​are combined to form the temporal sensitivity vector of the data frame.

[0039] S140: Traverse each time-series sensitivity vector in the operation phase and encrypt the real-time sensor data corresponding to the target sensitivity in the time-series sensitivity vector.

[0040] Here, target sensitivity is defined as the sensitivity in the time-series sensitivity vector that is greater than or equal to its corresponding preset sensitivity threshold. The preset sensitivity threshold corresponding to each sensitivity element in the time-series sensitivity vector can be the same or different, and can be determined based on practical experience.

[0041] This application embodiment can perform full encryption processing on the real-time sensor data corresponding to the target sensitivity in the time-series sensitivity vector, or perform partial encryption on a portion of the real-time sensor data, while not performing encryption processing on the real-time sensor data corresponding to the sensitivity in the time-series sensitivity vector other than the target sensitivity.

[0042] The secure processing method for embodied intelligent training data provided in this application abandons the one-size-fits-all protection model and instead constructs a two-dimensional evaluation mechanism that integrates static prior knowledge and dynamic real-time status. Specifically, this mechanism, on the one hand, utilizes an operational semantic graph constructed from historical experience to quantify the basic dependence of different operational stages on various types of data; on the other hand, it captures instantaneous sensitive features in the physical interaction process by sensing the rate of change of the current data frame. The temporal sensitivity vector generated by the two collaboratively can accurately identify the core data elements that truly need protection at each moment. Based on this, this application embodiment only encrypts a small amount of critical data whose sensitivity exceeds the threshold, while allowing the vast majority of non-sensitive or redundant data to pass through in plaintext. Ultimately, this application embodiment significantly reduces the overall encryption computation burden of the system while effectively curbing the risk of data leakage and tampering, while ensuring the availability of training data and the real-time nature of the processing flow, achieving a reasonable balance between security protection and computational efficiency.

[0043] In one possible implementation, applying indiscriminate encryption to all regions within a frame after precisely locking the sensitive frame would waste computational resources on redundant background information and could also interfere with visual servoing and control accuracy by excessively blurring the core operational target (such as a workpiece or reference surface). To address this technical problem, another embodiment of this application provides the following method:

[0044] The semantic distance between each region in the real-time sensor data corresponding to the target sensitivity and the operation target in the operation stage is calculated, and each region includes at least a local area of ​​an object; the semantic distance is processed based on the attenuation function to obtain the spatial encryption strength of each region; the corresponding region in the real-time sensor data is encrypted based on the spatial encryption strength, wherein the spatial encryption strength is positively correlated with the encryption level of the region.

[0045] The semantic distance calculation method may include: calculating the physical spatial distance between the center point of each region in the real-time sensor data corresponding to the target sensitivity and the center of the operation target in the operation phase; determining the semantic isolation factor of each region by judging whether each region and the operation target belong to the same physical rigid body; and determining the semantic distance for each region based on the physical spatial distance and the semantic isolation factor.

[0046] The formula for calculating semantic distance, based on physical spatial distance and semantic isolation factor, can include: ; In the formula, Indicates semantic distance. Indicates physical spatial distance. This represents the semantic isolation factor.

[0047] If the area to be judged and the target of the operation belong to the same physical rigid body. A preset minimum value can be used, which can be a constant greater than 0 and less than 1; if the region to be judged and the target do not belong to the same physical rigid body, You can select a preset maximum value, which can be a constant greater than 1.

[0048] The decay function includes, but is not limited to, the exponential decay function and the sigmoid function. For each data region in each frame after semantic distance calculation, the semantic distance is mapped to a normalized spatial encryption strength based on the decay function. This decay function is configured to approach saturated encryption strength as the semantic distance increases and approach zero encryption strength as the semantic distance decreases, so that the encryption strength presents a smooth increasing spatial distribution along the direction away from the center of the operation target. The spatial encryption strength is positively correlated with the final encryption level applied to each region. That is, the higher the spatial encryption strength, the stronger the encryption algorithm or the stricter the encryption parameters called in the corresponding region. Conversely, a lower strength encryption method or plaintext is used, thus forming a hierarchical encryption protection that gradually strengthens from the core operation region to the outer background region in the spatial dimension.

[0049] Through an additional spatial semantic awareness mechanism, this embodiment of the application can adaptively allocate differentiated encryption strengths within each frame of data that has been determined to be sensitive, based on the semantic distance between each region and the operation target: the core operation region remains weakly encrypted or plaintext is available, while the peripheral background region is protected by strong encryption. This mechanism achieves fine-grained hierarchical protection at the intra-frame spatial granularity, effectively compressing the computational area of ​​high-level encryption while ensuring visual servoing and control accuracy, reducing the overall encryption overhead of the system, and further improving the synergy between security protection and computational efficiency.

[0050] In one possible implementation, to further improve the accuracy and rationality of the adaptive encryption process, the operation semantic graph also includes: a sensitivity level labeled on the edge connecting each operation node and each data node; the preset sensitivity threshold is determined based on the sensitivity level.

[0051] Specifically, the sensitivity level can be determined based on task coreness, data sensitivity, and industry compliance rules: For each data node under the current operation node, the system weights and aggregates the task coreness and data sensitivity to obtain a comprehensive score. The comprehensive score is usually normalized to the range of 0-1, and then compared with a preset level mapping table to obtain a basic sensitivity level (e.g., L1-L5). The comprehensive score and the basic sensitivity level are positively correlated. After obtaining the basic sensitivity level, the system calls the industry compliance rule library for a rule-by-rule matching check: If the current data node does not trigger any compliance rules, the basic sensitivity level is directly output as the final sensitivity level; if the current data node triggers a compliance rule (e.g., the data contains facial information), the sensitivity level is forcibly set to the level specified by that rule.

[0052] Task coreness measures the importance of the current operation node within the overall training task. It reflects whether the failure of this operation would render the entire task unusable. High coreness indicates the operation is a critical path node and is indispensable. For example, the "insertion" operation in "precision shaft and hole assembly" is extremely core-critical; failure in this operation renders the entire workpiece unusable. Low coreness indicates the operation is auxiliary or substitutable. For example, the "equipment self-check" before the task begins or the "cleaning the workbench" after completion, even with brief delays or failures, does not affect the achievement of the main task objective. Task coreness is typically predefined by the topological relationships between task nodes and operation nodes in the operation semantic graph, or it can be derived statistically from historical task execution records.

[0053] Data sensitivity measures the importance of information carried by a data node in terms of privacy protection or commercial value. It reflects the extent of security damage that would result from the leakage or misuse of this data. High sensitivity data contains information that can directly or indirectly identify individuals (such as facial or voiceprint data), high-precision process parameters (such as precision force control curves), or confidential scene layouts (such as cloud-based observation maps within a laboratory). Leakage of this data could lead to privacy violations or the leakage of core technologies. Low sensitivity data consists of general environmental information or low-precision auxiliary measurements (such as ambient temperature or rough location information), and even public disclosure would not pose a substantial security risk. Data sensitivity can be predefined based on data type, precision level, and semantic content, and can also be dynamically updated through rules during training.

[0054] Industry compliance rules are legally binding or mandatory industry standards that define the minimum security requirements that data must meet during collection, storage, transmission, and use. These rules typically originate from national laws and regulations, industry standards (such as HIPAA in the medical field and confidentiality management regulations in the industrial field), or internal corporate compliance policies. Industry compliance rules are mandatory and non-negotiable, and their typical form is: "Any data involving XX type must have a sensitivity level no lower than X level" or "XX type data must be protected with XX strength of encryption." For example, facial data falls under biometric information under the Personal Information Protection Law, and its sensitivity level must be set to the highest level.

[0055] The preset sensitivity threshold is determined based on the sensitivity level, with an inverse mapping relationship between the two: the higher the sensitivity level, the more critical the data node is in terms of privacy protection or technological value, and the lower the corresponding preset sensitivity threshold. This allows encryption protection to be triggered even when the data has a low rate of change, achieving more stringent security. Conversely, the lower the sensitivity level, the higher the corresponding preset sensitivity threshold. Encryption is only triggered when the data undergoes drastic changes, avoiding unnecessary computational overhead on non-critical data. In practice, a corresponding threshold value can be pre-configured for each sensitivity level, forming a level-threshold mapping table. At runtime, the corresponding preset sensitivity threshold is obtained by looking up the table based on the sensitivity level of the current data node.

[0056] As can be seen from the above scheme, the embodiments of this application directly use the pre-annotated sensitivity levels in the operation semantic graph to determine the preset sensitivity thresholds corresponding to the data features under each operation node, so that the judgment criteria of the temporal sensitivity vector is aligned with the sensitivity level of the data itself, thereby improving the accuracy and rationality of the adaptive encryption process.

[0057] In one possible implementation, in the security protection of training data, encrypted data is mainly used for archiving and authorized backtracking, ensuring that the original sensor data can be accurately restored when needed to support process reproduction or accident investigation; while de-identified data is used for unauthorized fine-grained access scenarios such as model training and algorithm verification, preventing the precise extraction and misuse of core process parameters while retaining macro-trend characteristics. Therefore, this solution, based on encryption protection, further performs independent irreversible de-identification processing on core dependency data. Specifically, first, real-time sensor data generated during the operation phase is copied to obtain real-time sensor backup data; then, the target dependency weight is queried from the current operation data set, and the real-time sensor backup data corresponding to the target dependency weight is determined as the data to be de-identified, with the target dependency weight being a dependency weight greater than or equal to a preset weight threshold; finally, statistical mean is used to smooth the data to be de-identified to obtain the de-identified data.

[0058] The preset weight threshold can be dynamically determined based on the statistical distribution characteristics of all dependent weights in the current operation dataset at the current operation stage, or determined by a system-preset unified empirical threshold. In the dynamic determination method, an adaptive threshold calculation method can be used, including but not limited to: calculating the mean and standard deviation of all dependent weights in the current operation dataset and setting the preset weight threshold as the sum of the mean and one standard deviation; or using the percentile method, setting the preset weight threshold as the upper quartile (i.e., the 75th percentile) of all dependent weights to ensure that only high-dependent-weight data at the top are included in the desensitization scope. In the preset empirical threshold method, a fixed weight threshold (such as 0.7 or 0.8) can be preset based on the historical sensitivity statistics of the task type or operation node, and remain unchanged during operation or be adjusted by the operator.

[0059] After identifying the data to be anonymized, statistical mean smoothing can be used. The specific implementation process includes: for the force curve or time-series numerical sequence of the data to be anonymized, a sliding window mean filter is used to locally smooth the data. Specifically, taking the current sampling point as the center, the values ​​of a preset number of adjacent sampling points (e.g., 2 to 5 each) are taken. The arithmetic mean of all sampling point values ​​within the window is calculated, and this average value replaces the original value of the current sampling point. This process is repeated along the time sequence, traversing each sampling point in the data to be anonymized, thereby smoothing out sharp peaks and subtle fluctuations in the original data. The width of the sliding window can be preset according to the data sampling frequency and the required anonymization intensity. Higher sampling frequencies or higher desired anonymization intensity result in a larger window width for a more significant smoothing effect; conversely, a smaller window width preserves more of the original features. After statistical mean smoothing, the high-precision numerical information in the original data is permanently obscured, making it impossible to recover the accurate original values ​​before desensitization. This effectively prevents the precise extraction and misuse of core process parameters without affecting the macroscopic trend features required for model training.

[0060] For example, in the precision shaft and hole assembly stage, the contact force curve collected by the force sensor contains core process parameters such as peak time and amplitude. If the dependence weight of this force feature is 0.95 and greater than the preset weight threshold of 0.80, then the force backup data is used as the data to be desensitized. The average value of the force sampled values ​​in the window is taken by the sliding window mean filtering, and the peak amplitude in the original sequence is smoothed from 5.278N to 5.2N. At the same time, the overall upward trend of the curve and the approximate position of the peak are preserved, so that the desensitized data can still be used to train the force control strategy model, but the precise process parameter of 5.278N cannot be restored.

[0061] This application embodiment obtains an independent backup by copying real-time sensor data of the current operation stage, and selects target data with a higher than preset weight threshold based on the dependency weight in the operation semantic graph. Only the features closely related to the core operation in the backup are subjected to mean smoothing desensitization processing, thereby achieving permanent blurring of highly dependent and sensitive features without affecting the integrity of the original encrypted data, effectively preventing the accurate restoration and abuse of core process parameters.

[0062] This application embodiment achieves comprehensive security protection for embodied intelligence training data from the temporal, spatial, and feature dimensions through the collaborative processing of temporal adaptive encryption, spatial adaptive encryption, and causal feature filtering desensitization. Based on this, the system integrates the temporal adaptive encryption strategy, spatial adaptive encryption strategy, and causal feature filtering strategy into an encryption strategy configuration file. Combined with future action sequence prediction information provided by the action planner, it pre-generates and caches encryption strategy configurations for subsequent time periods, thereby achieving zero-latency real-time encryption processing during the execution phase. Simultaneously, the system inputs the encrypted data into a shadow training environment for pre-training verification. By comparing the change in model convergence accuracy before and after encryption, it dynamically monitors the impact of the current encryption strategy on algorithm performance. When the detected accuracy decrease exceeds a preset safety threshold, a strategy fallback mechanism is triggered, adaptively reducing the weight of the semantic isolation factor to appropriately relax the spatial encryption strength. This restores model accuracy while ensuring data security, maintaining a dynamic balance between security protection and algorithm usability, forming a closed-loop adaptive optimization mechanism.

[0063] Based on the above method embodiments, another embodiment of this application provides a secure processing device for embodied intelligent training data, such as... Figure 2 As shown, the device includes: The acquisition unit 210 is used to acquire an operation semantic graph, which is established based on historical embodied intelligence training task data. The operation semantic graph includes task nodes, operation nodes, data nodes, and dependency weights labeled on the edges connecting each operation node and each data node. The dependency weights are used to represent the degree of dependency of each operation node on each data node. The identification unit 220 is used to identify the current operating stage of the embodied smart device; Query unit 230 is used to query the current operation data set under the operation stage from the operation semantic graph, wherein the current operation data set includes the dependency weight between each operation node and each data node included in the operation stage; The calculation unit 240 is used to calculate the sensitivity of the operation node to each data node in the data frame for each data frame in the operation phase, based on the dependency weight between the operation node and each data node in the data frame and the rate of change of the data frame. The generation unit 250 is used to generate a timing sensitivity vector under the data frame based on the sensitivity of the operation node to each data node under the data frame. The encryption unit 260 is used to traverse each time-series sensitivity vector in the operation phase and encrypt the real-time sensor data corresponding to the target sensitivity in the time-series sensitivity vector. The target sensitivity is the sensitivity in the time-series sensitivity vector that is greater than or equal to its own preset sensitivity threshold.

[0064] In one possible implementation, the encryption unit 260 includes: The calculation module is used to calculate the semantic distance between each region in the real-time sensor data corresponding to the target sensitivity and the operation target of the operation stage, wherein each region includes at least a local region of an object; The attenuation module is used to process the semantic distance based on the attenuation function to obtain the spatial encryption strength of each region; An encryption module is used to encrypt the corresponding area in the real-time sensor data based on the spatial encryption strength, wherein the spatial encryption strength is positively correlated with the encryption level of the area.

[0065] In one possible implementation, the calculation module is used to calculate the physical spatial distance between the center point of each region in the real-time sensor data corresponding to the target sensitivity and the center of the operation target in the operation phase; determine the semantic isolation factor of each region by determining whether each region and the operation target belong to the same physical rigid body; and determine the semantic distance for each region based on the physical spatial distance and the semantic isolation factor.

[0066] In one possible implementation, the operation semantic graph further includes: a sensitivity level labeled for the edge connecting each operation node to each data node; The preset sensitivity threshold is determined based on the sensitivity level.

[0067] In one possible implementation, the device further includes: The copying unit is used to copy the real-time sensor data generated during the operation phase to obtain real-time sensor backup data. The query unit 230 is further configured to query a target dependency weight from the current operation data set, wherein the target dependency weight is a dependency weight that is greater than or equal to a preset weight threshold. The determining unit is used to determine the real-time sensor backup data corresponding to the target dependency weight as the data to be desensitized; The smoothing unit is used to smooth the data to be de-identified using a statistical mean to obtain the de-identified data.

[0068] The secure processing device for embodied intelligent training data provided in this application abandons the one-size-fits-all protection model and instead constructs a two-dimensional evaluation mechanism that integrates static prior knowledge and dynamic real-time status. Specifically, this mechanism, on the one hand, utilizes an operational semantic graph constructed from historical experience to quantify the basic dependence of different operational stages on various types of data; on the other hand, it captures instantaneous sensitive features in the physical interaction process by sensing the rate of change of the current data frame. The temporal sensitivity vector generated by the two collaboratively can accurately identify the core data elements that truly need protection at each moment. Based on this, this application embodiment only encrypts a small amount of critical data whose sensitivity exceeds the threshold, while allowing the vast majority of non-sensitive or redundant data to pass through in plaintext. Ultimately, this application embodiment significantly reduces the overall encryption computation burden of the system while effectively curbing the risk of data leakage and tampering, while ensuring the availability of training data and the real-time nature of the processing flow, achieving a reasonable balance between security protection and computational efficiency.

[0069] Based on the above method embodiments, another embodiment of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in any of the above embodiments.

[0070] Based on the above method embodiments, another embodiment of this application provides an electronic device or computer device, including: One or more processors; The processor is coupled to a storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the electronic device or computer device performs the method as described in any of the above embodiments.

[0071] Based on the above embodiments, another embodiment of this application provides a computer program product, which includes instructions that, when executed on a computer or processor, cause the computer or processor to perform the method described in any of the above embodiments.

[0072] The above-described apparatus embodiments correspond to the method embodiments and have the same technical effects. For detailed descriptions, please refer to the method embodiments. The apparatus embodiments are derived from the method embodiments; detailed descriptions can be found in the method embodiments section, and will not be repeated here. Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application.

[0073] Those skilled in the art will understand that the modules in the apparatus of the embodiments can be distributed in the apparatus of the embodiments as described in the embodiments, or they can be located in one or more devices different from this embodiment with corresponding changes. The modules of the above embodiments can be combined into one module, or they can be further divided into multiple sub-modules.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for securely processing embodied intelligent training data, characterized in that, The method includes: The operation semantic graph is obtained and the current operation stage of the embodied intelligent device is identified. The operation semantic graph is established based on historical embodied intelligent training task data. The operation semantic graph includes task nodes, operation nodes, data nodes, and dependency weights labeled on the edges connecting each operation node and each data node. The dependency weights are used to represent the degree of dependence of each operation node on each data node. Query the current operation data set under the operation stage from the operation semantic graph. The current operation data set includes the dependency weights between each operation node and each data node in the operation stage. For each data frame in the operation phase, the sensitivity of the operation node to each data node in the data frame is calculated based on the dependency weight between the operation node and each data node in the data frame and the rate of change of the data frame. Based on the sensitivity of the operation node to each data node in the data frame, a time-series sensitivity vector is generated for the data frame. Each time-series sensitivity vector in the operation phase is traversed, and the real-time sensor data corresponding to the target sensitivity in the time-series sensitivity vector is encrypted. The target sensitivity is the sensitivity in the time-series sensitivity vector that is greater than or equal to its own preset sensitivity threshold.

2. The method according to claim 1, characterized in that, Encrypting the real-time sensor data corresponding to the target sensitivity in the time-series sensitivity vector includes: Calculate the semantic distance between each region in the real-time sensor data corresponding to the target sensitivity and the operation target of the operation stage, wherein each region includes at least a local region of an object; The semantic distance is processed based on the decay function to obtain the spatial encryption strength of each region; The corresponding region in the real-time sensor data is encrypted based on the spatial encryption strength, wherein the spatial encryption strength is positively correlated with the encryption level of the region.

3. The method according to claim 2, characterized in that, Calculating the semantic distance between each region in the real-time sensor data corresponding to the target sensitivity and the operational target of the operational stage includes: Calculate the physical spatial distance between the center point of each region in the real-time sensor data corresponding to the target sensitivity and the center of the operation target in the operation phase; The semantic isolation factor for each region is determined by judging whether each region belongs to the same physical rigid body as the operation target; For each region, the semantic distance is determined based on the physical spatial distance and the semantic isolation factor.

4. The method according to claim 1, characterized in that, The operation semantic graph also includes: the sensitivity level labeled for the edges connecting each operation node to each data node; The preset sensitivity threshold is determined based on the sensitivity level.

5. The method according to any one of claims 1-4, characterized in that, The method further includes: Copy the real-time sensor data generated during the aforementioned operation phase to obtain real-time sensor backup data; The target dependency weight is queried from the current operation data set, and the real-time sensor backup data corresponding to the target dependency weight is determined as the data to be desensitized. The target dependency weight is a dependency weight that is greater than or equal to a preset weight threshold. The data to be de-identified is smoothed using statistical mean to obtain the de-identified data.

6. A secure processing device for embodied intelligent training data, characterized in that, The device includes: The acquisition unit is used to acquire an operation semantic graph, which is established based on historical embodied intelligence training task data. The operation semantic graph includes task nodes, operation nodes, data nodes, and dependency weights labeled on the edges connecting each operation node and each data node. The dependency weights are used to represent the degree of dependence of each operation node on each data node. The identification unit is used to identify the current operating stage of the embodied intelligent device; The query unit is used to query the current operation data set under the operation stage from the operation semantic graph. The current operation data set includes the dependency weights between each operation node and each data node included in the operation stage. The calculation unit is used to calculate the sensitivity of the operation node to each data node in the data frame for each data frame in the operation phase, based on the dependency weight between the operation node and each data node in the data frame and the rate of change of the data frame. The generation unit is used to generate a timing sensitivity vector for the data frame based on the sensitivity of the operation node to each data node under the data frame. An encryption unit is used to traverse each time-series sensitivity vector in the operation phase and encrypt the real-time sensor data corresponding to the target sensitivity in the time-series sensitivity vector. The target sensitivity is the sensitivity in the time-series sensitivity vector that is greater than or equal to its own preset sensitivity threshold.

7. The apparatus according to claim 6, characterized in that, The encryption unit includes: The calculation module is used to calculate the semantic distance between each region in the real-time sensor data corresponding to the target sensitivity and the operation target of the operation stage, wherein each region includes at least a local region of an object; The attenuation module is used to process the semantic distance based on the attenuation function to obtain the spatial encryption strength of each region; An encryption module is used to encrypt the corresponding area in the real-time sensor data based on the spatial encryption strength, wherein the spatial encryption strength is positively correlated with the encryption level of the area.

8. The apparatus according to claim 7, characterized in that, The calculation module is used to calculate the physical spatial distance between the center point of each region in the real-time sensor data corresponding to the target sensitivity and the center of the operation target in the operation stage; determine the semantic isolation factor of each region by judging whether each region and the operation target belong to the same physical rigid body; and determine the semantic distance for each region based on the physical spatial distance and the semantic isolation factor.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-5.

10. An electronic device, characterized in that, The electronic device includes: One or more processors; The processor is coupled to a storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the electronic device performs the method as described in any one of claims 1-5.

Images