A data generation method for addressing silent data corruption issues

By constructing a training program sample instruction map and an adaptive injection strategy for large remote sensing models, the problem of low data generation efficiency caused by silent data corruption in existing technologies is solved. This enables efficient generation and analysis of silent data corruption, ensuring the reliability of large remote sensing model training.

CN123114933APending Publication Date: 2026-11-17STEJT GRID ELEKTRIK PAUER INZHINIRING RISERCH INSTITYUT KO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611278862.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies are inefficient at generating silent data corruption problems on GPUs, making it difficult to efficiently identify and protect against silent data corruption in the training of large remote sensing models, resulting in resource waste and training failure.

Method used

A training program sample instruction map for a large remote sensing model is constructed. The baseline sample library is updated by injecting RTL-level errors. A similarity calculation model is trained, and an adaptive injection strategy is used to generate silent data corruption problem data.

Benefits of technology

It improves the efficiency of generating data with silent data corruption issues, enhances the efficiency of analyzing the propagation of error patterns caused by silent data corruption issues during training, and ensures the reliability of large-scale, long-term remote sensing training tasks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123114933A_ABST
    Figure CN123114933A_ABST
Patent Text Reader

Abstract

This invention relates to the field of data processing technology and discloses a method for generating data related to silent data corruption. The method includes: constructing a sample instruction map of a training program for a large remote sensing model; updating a baseline sample library by injecting RTL-level errors into sample instruction fragments in the sample instruction map; training an initial model based on the sample instruction map to obtain a target model; calculating the target similarity between the instruction fragment to be processed and the baseline sample library using the target model; determining a risk level based on the target similarity; and using an injection strategy corresponding to the risk level to adaptively inject RTL-level errors into the instruction fragment to be processed, thereby obtaining silent data corruption problem data. This invention transforms the generation of silent data corruption problem data from an inefficient random injection method to a highly efficient and intelligent generation method based on data and models, improving the generation efficiency of silent data corruption problem data and the efficiency of analyzing error patterns caused by the propagation of silent data corruption problems during training.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and specifically to a data generation method for addressing the problem of silent data corruption. Background Technology

[0002] GPUs (Graphics Processing Units) are essential computing tools for training large-scale remote sensing models. However, during use, GPUs can experience transient failures due to factors such as particle impacts on silicon devices, potentially causing Silent Data Corruption (SDC) in the programs running on the GPU. SDC issues are difficult to detect in time during training, and the corrupted data contaminates the training process, ultimately leading to the failure of large-scale model training and resulting in a significant waste of GPU computing and power resources. To reduce the resource waste caused by SDC, it is necessary to study and analyze the potential harm of SDC to the model during the training of large-scale remote sensing models, identify weaknesses in the training program, and design corresponding protective measures.

[0003] Analyzing the impact of SDC (Signal Defect Conviction) issues on training requires acquiring a sufficient amount of SDC problem data. Current techniques inject faults into random locations in the instruction sequence during training on a GPU and compare the output after fault injection with the output of error-free execution to determine if an SDC problem has occurred. Data exhibiting SDC problems needs to be collected and repeated multiple times to accumulate sufficient SDC problem data.

[0004] However, the success rate of random injection to generate SDC problems is less than 1%. For large remote sensing models with billions of parameters, a large number of injections are required to obtain sufficient SDC problem data, which can take up to tens of days, severely limiting the efficiency of SDC problem analysis. Summary of the Invention

[0005] This invention provides a data generation method to address the problem of silent data corruption, thereby resolving one of the technical problems existing in the prior art.

[0006] In a first aspect, the present invention provides a data generation method for addressing the problem of silent data corruption, the method comprising: A sample instruction map of the training program for constructing a large remote sensing model; By injecting RTL-level errors into sample instruction fragments in the sample instruction graph, the baseline sample library that causes silent data corruption is updated. An initial similarity calculation model is trained based on the sample instruction map to obtain a target similarity calculation model; For each instruction fragment to be processed in the training program, a target similarity calculation model is used to calculate the target similarity between the instruction fragment to be processed and the benchmark sample library; Risk levels are determined based on target similarity, and injection strategies corresponding to the risk levels are adopted to perform adaptive RTL-level error injection on the instruction fragments to be processed, thereby obtaining the corresponding silent data corruption problem data.

[0007] This invention constructs a sample instruction map of the training program for a large-scale remote sensing model, systematically organizing the underlying instruction sequences into a graph structure rich in semantic information. Then, by injecting RTL-level errors into the sample instruction fragments in the sample instruction map to update the benchmark sample library, vulnerable pattern samples can be obtained in a targeted and systematic manner. Next, a similarity calculation model is trained based on the sample instruction map, enabling the model to assess the similarity of vulnerability among instruction fragments. For each instruction fragment to be processed in the training program, the trained model calculates its target similarity with the benchmark sample library, quantifying its potential vulnerability. Finally, the risk level is determined based on the target similarity, and a corresponding adaptive injection strategy is adopted to obtain silent data corruption problem data. By transforming the generation of silent data corruption problem data from an inefficient random injection method to a highly efficient and intelligent generation method based on data and models, the generation efficiency of silent data corruption problem data is improved, effectively enhancing the efficiency of analyzing error patterns caused by the propagation of silent data corruption problems during training, and providing technical support for ensuring the reliability of large-scale, long-term remote sensing training tasks.

[0008] In one alternative implementation, the sample instruction map of the training procedure for constructing a large remote sensing model includes: A program for extracting sample instructions from the training procedure of a large remote sensing model; By identifying the boundaries of the instruction segments in the sample instruction program, the sample instruction program is divided into multiple sample instruction segments, and each sample instruction segment is treated as a node. Each sample instruction fragment is divided into multiple dimensions to obtain the attribute features of each sample instruction fragment, which are used as the node attributes of the corresponding node. Edges are constructed based on the control flow and data dependencies between all sample instruction fragments to obtain the sample instruction graph.

[0009] This invention extracts instruction sequences and identifies instruction fragment boundaries, enabling the segmentation of continuous, unstructured GPU instruction sequences into logically independent, internally sequentially executed instruction fragments, which are then treated as nodes. The instruction fragments are divided in multiple dimensions, and each node is assigned attributes. By identifying control flow and data dependencies to construct edges, the resulting sample instruction graph provides a data foundation rich in structural information for subsequent steps.

[0010] In one alternative implementation, the baseline sample library that causes silent data corruption is updated by injecting RTL-level errors into sample instruction fragments in the sample instruction graph, including: Select multiple seed nodes from the sample instruction graph; Perform a parameterized scan for each RTL level error to obtain a list of parameter combinations; For each seed node, RTL-level errors are injected into the sample instruction fragment corresponding to the seed node according to the parameter combination list and executed to obtain the first instruction execution result; When the execution result of the first instruction shows a problem of silent data corruption, the sample instruction fragment is used as a benchmark sample, and the benchmark sample and its corresponding injection configuration are added to the benchmark sample library.

[0011] This invention systematically traverses all possible error manifestations by selecting representative seed nodes from the sample instruction graph and performing parameterized scanning on each type of RTL-level error to generate a complete list of parameter combinations. Then, for each seed node, it performs comprehensive injection and execution according to the parameter combination list, enabling efficient triggering of SDC (Signal Defective) issues in a targeted and systematic manner. Finally, by adding the sample instruction fragments verified as SDC and their injection configurations to the benchmark sample library, fully verified vulnerability patterns can be solidified in a structured form.

[0012] In one optional implementation, an initial similarity calculation model is trained based on the sample instruction graph to obtain a target similarity calculation model, including: Extract the sample feature vector of each node in the sample instruction graph, and enumerate all vector pairs formed by the sample feature vectors; An initial similarity calculation model is used to map each vector pair to an embedding vector and calculate the sample similarity. The data dimension of the embedding vector is lower than that of the sample feature vector. The initial similarity calculation model is trained based on vector pairs with sample labels to obtain the target similarity calculation model. The sample labels indicate whether the corresponding sample instruction fragments of each vector pair are similar.

[0013] This invention extracts the sample feature vector of each node in the sample instruction graph and enumerates all vector pairs. An initial similarity calculation model is then used to map each vector pair to an embedding vector and calculate sample similarity. Finally, the model is trained based on the sample labels to obtain a target similarity calculation model with stable and reliable discrimination capabilities.

[0014] In one optional implementation, extracting the sample feature vector for each node in the sample instruction graph includes: For each node in the sample instruction map, obtain the static and dynamic features of the sample instruction fragment corresponding to the node; Determine the in-degree and out-degree of the node in the sample instruction graph; Enumerate all node pairs in the sample instruction graph, construct the shortest path between each pair of nodes, and calculate the betweenness centrality of the nodes based on the number of shortest paths passing through the nodes and the total number of shortest paths for all node pairs. Identify critical path nodes in the sample instruction graph and calculate the shortest distance from each node to the critical path node. The in-degree, out-degree, betweenness centrality, and shortest distance of a node are used as the sample graph features of the node. The sample graph features, static features, and dynamic features are concatenated to obtain the sample feature vector of the node.

[0015] This invention constructs the graph features of nodes by determining their degree centrality, betweenness centrality, and shortest distance. These features are then combined with static and dynamic features to form a sample feature vector that comprehensively reflects the structural roles and behavioral characteristics of nodes. This provides an input representation for the subsequent similarity calculation model, enabling the model to accurately assess the vulnerability similarity between instruction fragments from multiple dimensions.

[0016] In one optional implementation, for each target instruction fragment to be processed in the training program, a target similarity calculation model is used to calculate the target similarity between the target instruction fragment to be processed and the target sample library, including: For each instruction segment to be processed in the training program, extract the first feature vector of the instruction segment to be processed, and extract the second feature vector of each benchmark sample in the benchmark sample library; A target similarity calculation model is adopted, based on the first feature vector and the second feature vector of each benchmark sample, to calculate the vulnerability similarity between the target instruction fragment and each benchmark sample in the benchmark sample library; The maximum value among all vulnerability similarities is taken as the target similarity.

[0017] This invention employs a target similarity calculation model to calculate the target similarity between the instruction fragment to be processed and the benchmark sample library. It can fully utilize the vulnerability discrimination knowledge learned by the model from a large number of training samples to achieve intelligent assessment based on data and model, reflect the vulnerability level of the instruction fragment to be processed, and provide a quantitative basis for subsequent risk level classification and adaptive injection strategy selection.

[0018] In one optional implementation, a risk level is determined based on target similarity, and an injection strategy corresponding to the risk level is adopted to perform adaptive RTL-level error injection on the target instruction fragment to be processed, thereby obtaining the corresponding silent data corruption problem data, including: When the risk level is Level 1, a verification RTL-level error injection is performed on the instruction fragment to be processed to obtain the corresponding silent data corruption problem data. When the risk level is level 2, boundary exploration RTL-level error injection is performed on the instruction fragment to be processed to obtain the corresponding silent data corruption problem data. Level 1 is higher than level 2. When the risk level is level three, exploratory RTL-level error injection is performed on the instruction fragment to be processed to obtain the corresponding silent data corruption problem data. Level two is higher than level three.

[0019] This invention maps target similarity to risk level, which can quantitatively assess the probability of the instruction fragment to be processed generating SDC problem, and uses the injection strategy corresponding to the risk level for injection, thereby improving the generation efficiency of SDC problem data.

[0020] In one optional implementation, when the risk level is Level 1, a verification RTL-level error injection is performed on the instruction fragment to be processed to obtain the corresponding silent data corruption problem data, including: When the risk level is the first level, the injection configuration of the benchmark sample corresponding to the target similarity is adopted to inject RTL-level errors into the instruction fragment to be processed and execute it to obtain the second instruction execution result; When the execution result of the second instruction shows a silent data corruption problem, the injected instruction fragment to be processed is used as the silent data corruption problem data, and the instruction fragment to be processed and its injection configuration are added to the benchmark sample library. When the execution result of the second instruction does not show the problem of silent data corruption, the fine-tuning configuration of the injection configuration of the benchmark sample corresponding to the target similarity is adopted, and the instruction fragment to be processed is injected with RTL level error again and executed to obtain the result of the third instruction; When the third instruction result shows a silent data corruption problem, the instruction fragment to be processed after re-injection is used as the silent data corruption problem data, and the instruction fragment to be processed and its re-injection configuration are added to the benchmark sample library.

[0021] This invention employs a confirmatory injection strategy at the first level, directly reusing the injection configuration of the most similar benchmark sample for rapid verification. This enables the identification of vulnerabilities in high-risk segments at the lowest cost. If the reuse is successful, the new sample and its injection configuration are added to the benchmark sample library, continuously enriching the library of known vulnerability patterns.

[0022] In one optional implementation, when the risk level is level two, boundary-exploratory RTL-level error injection is performed on the instruction fragment to be processed to obtain the corresponding silent data corruption problem data, including: When the risk level is level two, the number of injections is determined based on the target similarity. Select the parameter combination for the number of injections from the parameter combination list, perform layered injection of the instruction fragment to be processed, and execute it after each layer of injection to obtain the execution result of the fourth instruction. When the execution result of the fourth instruction shows a silent data corruption problem, the injected instruction fragment to be processed is used as the silent data corruption problem data, and the instruction fragment to be processed and its injection configuration are added to the benchmark sample library. If the execution result of the fourth instruction does not show a problem of silent data corruption, continue to inject the instruction fragments to be processed in layers until the injection count is reached.

[0023] This invention employs a boundary exploration injection strategy at the second level, selecting a corresponding number of parameter combinations from the parameter combination list for layered injection. This allows the discovery of critical conditions from normal operation to the occurrence of SDC problems. Samples that successfully trigger SDC problems and their injection configurations are added to the benchmark sample library, continuously enriching the known vulnerability pattern library.

[0024] In one optional implementation, when the risk level is level three, exploratory RTL-level error injection is performed on the instruction fragment to be processed to obtain the corresponding silent data corruption problem data, including: When the risk level is level 3, random RTL-level error injection is performed on the instruction fragment to be processed and executed to obtain the execution result of the fifth instruction. The random injection is repeated until the preset injection density is reached. When the execution result of any fifth instruction exhibits a silent data corruption problem, the injected instruction fragment corresponding to the execution result of the fifth instruction is taken as the silent data corruption problem data, and the instruction fragment to be processed and its injection configuration are added to the benchmark sample library. Obtain the neighbor nodes of the instruction fragment to be processed from the instruction graph where the instruction fragment to be processed is located, corresponding to the execution result of the fifth instruction. Based on the instruction fragment to be processed corresponding to the neighbor node, return to the step of using the target similarity calculation model to calculate the target similarity between the instruction fragment to be processed and the benchmark sample library.

[0025] This invention employs an exploratory injection strategy at the third level, using random injection to discover SDC (Signal Disturbance Controller) issues, and adds the newly discovered vulnerability patterns and their injection configurations to the baseline sample library. Then, by utilizing the structural relationships of the instruction graph, neighboring nodes are processed, improving the efficiency of vulnerability discovery in low-similarity regions.

[0026] Secondly, the present invention provides a data generation apparatus for addressing the problem of silent data corruption, the apparatus comprising: The building block is used to construct a sample instruction map of the training program for large remote sensing models; The update module is used to update the benchmark sample library that causes silent data corruption by injecting RTL-level errors into sample instruction fragments in the sample instruction graph. The training module is used to train an initial similarity calculation model based on the sample instruction map to obtain the target similarity calculation model; The calculation module is used to calculate the target similarity between the instruction fragment to be processed and the benchmark sample library for each instruction fragment to be processed in the training program, using the target similarity calculation model. The generation module is used to determine the risk level based on the target similarity, and then use the injection strategy corresponding to the risk level to perform adaptive RTL-level error injection on the instruction fragment to be processed, so as to obtain the corresponding silent data corruption problem data.

[0027] Thirdly, the present invention provides an electronic device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the data generation method for solving the silent data corruption problem described in the first aspect or any corresponding embodiment.

[0028] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute the data generation method for the silent data corruption problem described in the first aspect or any corresponding embodiment thereof. Attached Figure Description

[0029] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0030] Figure 1 This is a flowchart of a data generation method for addressing the problem of silent data corruption according to an embodiment of the present invention; Figure 2 This is a structural block diagram of a data generation apparatus for addressing the problem of silent data corruption according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0033] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0034] According to an embodiment of the present invention, a data generation method embodiment for solving the problem of silent data corruption is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0035] This embodiment provides a data generation method for addressing the problem of silent data corruption. Figure 1 This is a flowchart of a data generation method for addressing the silent data corruption problem according to an embodiment of the present invention, such as... Figure 1 As shown, the process includes the following steps: Step S101: Construct a sample instruction map of the training program for the remote sensing large model.

[0036] Specifically, the training procedure for the remote sensing large model includes multiple instruction programs, which are used to construct a sample instruction map, systematically organizing the unstructured low-level instruction sequences into a graph structure rich in semantic information.

[0037] Step S102: Update the baseline sample library that causes silent data corruption by injecting RTL-level errors into sample instruction fragments in the sample instruction graph.

[0038] Specifically, GPU RTL (Register Transfer Level) errors include four types: opcode errors, control flow errors, parallel management errors, and resource management errors. Among these, parallel management errors are the most likely to cause SDC (Single Thread Disorder) problems because they only cause execution thread misalignment without leading to program crashes. Therefore, this embodiment of the invention focuses on injecting parallel management errors as the primary injection target. During execution, parallel management errors are injected into sample instruction fragments in the sample instruction graph. The results of the injected instructions are then used to filter out samples that cause SDC problems, thereby updating the benchmark sample library. This allows for the rapid accumulation of a batch of known vulnerable pattern samples, i.e., known and reproducible SDC problem samples, based on a small amount of targeted injection. It should be noted that if this is the first time constructing a sample instruction graph for training, the benchmark sample library is empty and is continuously updated as the training process progresses.

[0039] Step S103: Train an initial similarity calculation model based on the sample instruction map to obtain the target similarity calculation model.

[0040] Specifically, an initial similarity calculation model is trained through the sample instruction graph to obtain a target model capable of assessing the similarity between two instruction fragments at the vulnerability level, providing a basis for transferring the injection experience of known SDC samples to unknown fragments.

[0041] Step S104: For each instruction segment to be processed in the training program, the target similarity calculation model is used to calculate the target similarity between the instruction segment to be processed and the benchmark sample library.

[0042] Specifically, the instruction fragment to be processed belongs to an instruction program that did not participate in the construction of the sample instruction map, and its vulnerability is unknown. By using a trained target similarity calculation model, the target similarity between the instruction fragment to be processed and the benchmark sample library can be calculated, which can quantify the potential probability of the instruction fragment to be processed generating SDC problems.

[0043] Step S105: Determine the risk level based on target similarity, adopt the injection strategy corresponding to the risk level, and perform adaptive RTL-level error injection on the instruction fragment to be processed to obtain the corresponding silent data corruption problem data.

[0044] Specifically, risk levels are assigned based on the range of target similarity, indicating the likelihood that the instruction fragment to be processed will cause a SDC (Safety, Distress, and Corruption) problem. The higher the target similarity, the closer the instruction fragment to a known vulnerability pattern is, and the higher its risk level. For example: when the target similarity is ≥90%, the risk level is classified as Level 1, i.e., a high vulnerability similarity fragment; when 50% < target similarity < 90%, the risk level is classified as Level 2, i.e., a medium vulnerability similarity fragment; and when the target similarity is ≤50%, the risk level is classified as Level 3, i.e., a low vulnerability similarity fragment.

[0045] Based on different risk levels, corresponding injection strategies are selected to perform adaptive RTL-level error injection on the instruction fragments to be processed, thereby obtaining the corresponding silent data corruption problem data. Compared with the random injection in traditional methods, the adaptive injection method based on similarity in this embodiment of the invention improves the generation efficiency of SDC problem data, thereby enhancing the efficiency of SDC problem analysis.

[0046] This invention constructs a sample instruction map of the training program for a large-scale remote sensing model, systematically organizing the underlying instruction sequences into a graph structure rich in semantic information. Then, by injecting RTL-level errors into the sample instruction fragments in the sample instruction map to update the benchmark sample library, vulnerable pattern samples can be obtained in a targeted and systematic manner. Next, a similarity calculation model is trained based on the sample instruction map, enabling the model to assess the similarity of vulnerability among instruction fragments. For each instruction fragment to be processed in the training program, the trained model calculates its target similarity with the benchmark sample library, quantifying its potential vulnerability. Finally, the risk level is determined based on the target similarity, and a corresponding adaptive injection strategy is adopted to obtain silent data corruption problem data. By transforming the generation of silent data corruption problem data from an inefficient random injection method to a highly efficient and intelligent generation method based on data and models, the generation efficiency of silent data corruption problem data is improved, effectively enhancing the efficiency of analyzing error patterns caused by the propagation of silent data corruption problems during training, and providing technical support for ensuring the reliability of large-scale, long-term remote sensing training tasks.

[0047] This embodiment provides a data generation method for addressing the problem of silent data corruption, which specifically includes the following steps: Step S201: Construct a sample instruction map of the training program for the remote sensing large model.

[0048] Specifically, step S201 includes: Step S2011: Extract sample instruction program from the training program of the remote sensing large model.

[0049] Specifically, the training program is used to train a large-scale remote sensing model. It utilizes relevant tools to extract SASS-level instruction programs from the compiled CUDA kernel function file. Since the training program for a large-scale remote sensing model typically contains multiple instruction programs, a sample instruction program for generating SDC problem data is selected from these.

[0050] Step S2012: By identifying the instruction segment boundaries of the sample instruction program, the sample instruction program is divided into multiple sample instruction segments, and each sample instruction segment is treated as a node.

[0051] Specifically, the boundary of an instruction segment is defined by both the start point and the end point. The start point includes the entry point of the instruction program (the first instruction) and the target address of any branch instructions (i.e., the instruction to which a jump is made). The end point is usually the location of a branch instruction (e.g., BRA unconditional jump, JMP jump, EXIT exit, etc.), or the instruction preceding a jump to a target address. During execution, the entire sample instruction program is traversed, and all start and end points are marked based on branch instructions and their target addresses. The continuous sequence of instructions between adjacent start and end points is extracted into an instruction segment. Each instruction segment is executed sequentially; there are no branch jumps. Control flow only transfers to other segments after the last instruction of the segment has been executed. Each instruction segment is treated as a node.

[0052] Step S2013: Perform multi-dimensional segmentation on each sample instruction fragment to obtain the attribute features of each sample instruction fragment, which are used as the node attributes of the corresponding node.

[0053] Specifically, the attribute features of the sample instruction fragments are obtained by dividing them according to the following dimensions: (1) Idempotency: Instruction segments are divided into idempotent segments and non-idempotent segments. Idempotent segments refer to instruction sequences whose final value after repeated execution of the same input operands any number of times is completely equivalent to that after execution only once, provided that the instruction is executed correctly. Its essence is the functional semantic attribute of the instruction sequence, which depends on the ideal premise of "correct instruction execution" and is not directly related to physical faults at the hardware level. Non-idempotent segments refer to instruction segments whose repeated execution will lead to state superposition or cumulative changes (such as accumulation operations). This dimension is used to characterize the sensitivity of instruction segments to repeated execution at the logical level, providing basic features at the functional semantic level for subsequent node attribute analysis. (2) Functional category: Based on the core operation of the instruction fragment, it is classified into computationally intensive (e.g. matrix multiplication-accumulation operation on Tensor Core), memory access intensive (e.g. data loading, weight update), control flow (e.g. condition jump, loop), communication intensive (e.g. collection communication primitives such as Reduce Scatter, AllGather, etc.) or synchronous operation; (3) Criticality Score: This assesses the importance of the instruction fragment on the critical training path. A higher score indicates a greater impact after an SDC problem occurs. First, nodes on the gradient calculation and backpropagation paths are identified, such as nodes that accumulate weight gradients. Once an SDC problem occurs, its impact will continue to accumulate and cause great harm. Nodes that calculate temporary variables are usually not on the above paths and have a relatively smaller impact. Then, breadth-first search or Dijkstra's algorithm is used to calculate the shortest distance (number of hops) from the node to each node on the above path. The closer the distance, the closer it is to the core calculation. The minimum value of the shortest distance from the node to each node on the above path is normalized (the closer the distance, the closer the score is to 1.0; the farther the distance, the closer the score is to 0.0), and weighted according to the functional category of the instruction fragment to obtain the criticality score. (4) Execution context: The specific module of the remote sensing big model to which the instruction fragment belongs, such as the QKV calculation part of the Attention module, the MLP layer, and the LayerNorm layer, is used to analyze the sensitivity differences of different modules to the SDC problem; By dividing the data into the above dimensions, we can obtain the attribute features of each sample instruction fragment, which are used as the node attributes of the corresponding node of the sample instruction fragment. These attributes include: idempotency (True / False); functional category (computation-intensive, memory-intensive, control-flow, communication-intensive, or synchronous operation); criticality score (0.0-1.0); and execution context (e.g., Attention.QK_MatMul, MLP.Layer1, LayerNorm.Backward, etc.).

[0054] Step S2014: Construct edges based on the control flow and data dependencies between all sample instruction fragments to obtain the sample instruction graph.

[0055] Specifically, edges represent the relationships between instruction segments, generally including two types: one is a control flow edge representing the execution flow. By analyzing the jump, branch, and other instructions at the end of each instruction segment, the subsequent instruction segment to be executed is determined. If the last instruction of an instruction segment A transfers control to instruction segment B through branching, jumping, or sequential execution, then there exists a directed control flow edge from A to B. The other type is a data dependency edge representing data flow dependence. By analyzing the read and write operations of each instruction segment on registers and memory, the data transfer path is identified. If the output of instruction segment A is the input of instruction segment B, then there exists a directed data dependency edge from A to B.

[0056] Finally, the discrete instruction fragments are organized into a graph structure that reflects the actual execution logic of the program, namely the sample instruction graph. It is a directed graph consisting of nodes, edges and node attributes, formally described as G=(V,E,A), where V represents the set of nodes, E represents the set of edges, and A represents the set of node attributes.

[0057] Step S202: Update the baseline sample library that causes silent data corruption by injecting RTL-level errors into sample instruction fragments in the sample instruction graph.

[0058] Specifically, step S202 includes: Step S2021: Select multiple seed nodes from the sample instruction graph.

[0059] Specifically, multiple seed nodes are selected from the sample instruction graph. The selection criteria include: nodes located on the critical path (such as weight gradient accumulation, QKV calculation of the Attention module, etc.), and some nodes selected from different functional categories.

[0060] Step S2022: Perform a parameterized scan for each type of RTL-level error to obtain a list of parameter combinations.

[0061] Specifically, to systematically explore the behavior of seed nodes under different hardware failure conditions, it is necessary to inject them with a refined RTL-level error model. This embodiment of the invention selects RTL-level errors most prone to causing SDC problems, including the following three types of parallel management errors: (1) Warp scheduler error: simulate thread scheduling misalignment by reversing thread or Warp identifier; it can also simulate more specific errors, such as Warp scheduling priority error (causing high-priority Warp that should be executed first to be delayed) and Warp merging or splitting error; (2) Thread bundle error: The voting result of the thread in the simulated Warp on the branch judgment is incorrectly flipped due to a transient fault, causing the entire Warp to go to the wrong branch path; (3) Instruction dispatch error: The simulated instruction dispatch unit sends an instruction to another mismatched functional unit in the SM, for example, sending an integer arithmetic instruction to the floating-point arithmetic unit.

[0062] For each RTL-level error type, its parameter space is first determined (e.g., for Warp scheduler errors, parameters include the bit sequence number of the flipped thread ID (bits 0-4) and the flip direction (0→1 or 1→0). Then, the Cartesian product of all parameters is enumerated to generate all possible parameter combinations. This yields a list of parameter combinations for all RTL-level errors.

[0063] Step S2023: For each seed node, inject RTL-level errors into the sample instruction fragment corresponding to the seed node according to the parameter combination list and execute it to obtain the first instruction execution result.

[0064] Specifically, for each seed node, each parameter combination in the parameter combination list is traversed, and injection is performed at multiple key instruction positions within the sample instruction fragment corresponding to that seed node. After each parameter combination injection is completed, the instruction fragment is executed to obtain the first instruction execution result. Through this step, the SDC problem can be triggered efficiently in a targeted and systematic manner.

[0065] Step S2024: When the execution result of the first instruction shows a silent data corruption problem, the sample instruction fragment is used as a benchmark sample, and the benchmark sample and its corresponding injection configuration are added to the benchmark sample library.

[0066] Specifically, to filter out valid samples that truly generate SDC issues, the execution result of the first instruction corresponding to each parameter combination needs to be compared with the golden output result. The golden output result refers to the standard execution result obtained by executing the original instruction fragment without error injection. By comparison, if the execution result of the first instruction under a certain parameter combination is inconsistent with its golden output result and the program does not crash, then the execution result is determined to exhibit an SDC issue. In this case, the original sample instruction fragment corresponding to that parameter combination is used as the baseline sample, and its injection configuration (including error type, injection location, injection parameter combination, etc.) is added to the baseline sample library, thereby solidifying the verified vulnerability pattern as a baseline.

[0067] Step S203: Train an initial similarity calculation model based on the sample instruction map to obtain the target similarity calculation model.

[0068] Specifically, step S203 includes: Step S2031: Extract the sample feature vector of each node in the sample instruction graph, and enumerate all vector pairs formed by the sample feature vectors.

[0069] Specifically, step S2031 above, which retrieves the sample feature vector of each node in the sample instruction graph, includes: Step a1: For each node in the sample instruction map, obtain the static and dynamic features of the sample instruction fragment corresponding to the node.

[0070] Specifically, for each node in the sample instruction graph, the static and dynamic features of its corresponding sample instruction fragment are obtained. Static features refer to the inherent attributes of the instruction that are independent of the runtime environment, including instruction type histograms (proportion of FP32, TF32, and INT8 instructions), register usage pressure, shared memory usage, and instruction-level parallelism. Dynamic features refer to the behavioral characteristics exhibited by the instruction fragment during actual execution, including memory access locality, warp execution efficiency, and the proportion of active thread bundles.

[0071] Step a2: Determine the in-degree and out-degree of the node in the sample instruction graph.

[0072] Specifically, for each node, the number of edges pointing to that node is taken as its in-degree, and the number of edges originating from that node and pointing to other nodes is taken as its out-degree. The degree centrality of a node is described by determining its in-degree and out-degree. In data dependencies, a node with a high in-degree indicates that it depends on many preceding computation results, while a node with a high out-degree indicates that its computation results are used by many subsequent steps, and may be a computational bottleneck or a key data distribution point.

[0073] Step a3: Enumerate all node pairs in the sample instruction graph, construct the shortest path between each pair of nodes, and calculate the betweenness centrality of the nodes based on the number of shortest paths passing through the nodes and the total number of shortest paths for all node pairs.

[0074] Specifically, all node pairs in the sample instruction graph are enumerated. For each node pair, the shortest path between the two nodes is constructed, which is the path with the fewest edges connecting the two nodes. The number of shortest paths passing through the node and the total number of shortest paths for all node pairs are counted. The ratio of the two is used as the betweenness centrality of the node. The higher the value, the more likely the node is to be a key hub for information flow, and its damage may lead to widespread computational errors, making it more vulnerable.

[0075] Step a4: Identify the critical path nodes in the sample instruction graph and calculate the shortest distance from the node to the critical path node.

[0076] Specifically, the backpropagation calculation of the loss function and the weight gradient update is the critical path in the training task of the large remote sensing model. Starting from the node representing the loss function calculation, the critical path is identified by traversing the data dependency path in reverse, and nodes located on the critical path are designated as critical path nodes. Damage to any critical path node will directly affect the correctness of gradient calculation during backpropagation. Referring to the criticality score calculation method in step S2023, the shortest distance from the node to the critical path node is calculated. The smaller the value, the closer the node is to the core calculation, and the greater the potential impact of the SDC problem.

[0077] Step a5: Use the in-degree, out-degree, betweenness centrality, and shortest distance of the node as the sample graph features of the node. Concatenate the sample graph features, static features, and dynamic features to obtain the sample feature vector of the node.

[0078] Specifically, the in-degree, out-degree, betweenness centrality, and shortest distance of a node are used as sample graph features to describe the node's position and importance in the sample instruction graph. Concatenating these features with static and dynamic features generates a high-dimensional vector for each node that comprehensively reflects its structural role and behavioral characteristics.

[0079] Alternatively, as an alternative to the above-mentioned sample feature vector extraction method, the nodes and edges of the sample instruction graph can be used as input, and the static and dynamic features of the nodes can be combined to output the sample feature vector of each node through graph embedding algorithm or graph neural network training.

[0080] Step S2032: Using the initial similarity calculation model, each vector pair is mapped to an embedding vector, and the sample similarity is calculated. The data dimension of the embedding vector is lower than that of the sample feature vector.

[0081] Specifically, the initial similarity calculation model uses a Siamese neural network, consisting of two structurally identical sub-networks and a distance calculation module. The two sample feature vectors from each vector pair are input into the two sub-networks respectively, mapping the high-dimensional sample feature vectors to low-dimensional embedding vectors. Then, the distance calculation module receives the embedding vectors from the two sub-networks and calculates the cosine similarity or Euclidean distance between the two embedding vectors as the sample similarity between them.

[0082] Step S2033: Train an initial similarity calculation model based on vector pairs with sample labels to obtain a target similarity calculation model. The sample labels indicate whether the corresponding sample instruction fragments of each vector pair are similar.

[0083] Specifically, firstly, an unsupervised clustering algorithm is used to initially filter out similar vector pairs from all vector pairs. Then, professional technicians confirm each selected vector pair to obtain sample labels. If the two are considered similar, the sample label is marked as 1; if they are considered dissimilar, the sample label is marked as 0. Since manual similarity judgment is very costly, supervised learning is performed using vector pairs with sample labels to train an initial similarity calculation model, thereby obtaining a target similarity calculation model with stable and reliable similarity judgment capabilities, which can then be used for rapid judgment.

[0084] Step S204: For each instruction segment to be processed in the training program, the target similarity calculation model is used to calculate the target similarity between the instruction segment to be processed and the benchmark sample library.

[0085] Specifically, step S204 includes: Step S2041: For each instruction segment to be processed in the training program, extract the first feature vector of the instruction segment to be processed, and extract the second feature vector of each benchmark sample in the benchmark sample library.

[0086] Specifically, for each instruction fragment to be processed in the training program of the large remote sensing model, it did not participate in the training process of the similarity calculation model, that is, it did not participate in the construction of the sample instruction map, and its vulnerability is unknown. Referring to the method of constructing the sample instruction map, an instruction map of the instruction program to which the instruction fragment to be processed is located is constructed. Then, referring to the method of extracting sample feature vectors in step S2031, the first feature vector of the instruction fragment to be processed is extracted using the map to which the instruction fragment to be processed is located. Similarly, for each benchmark sample in the benchmark sample library, its second feature vector is extracted. If the benchmark sample has already participated in feature extraction during the training process of the similarity calculation model, it can be directly reused without further extraction.

[0087] Step S2042: Using a target similarity calculation model, based on the first feature vector and the second feature vector of each benchmark sample, calculate the vulnerability similarity between the target instruction fragment and each benchmark sample in the benchmark sample library.

[0088] Specifically, the first feature vector of the instruction fragment to be processed and the second feature vector of each benchmark sample are input into the target similarity calculation model to obtain the vulnerability similarity between the two.

[0089] Step S2043: Take the maximum value among all vulnerability similarities as the target similarity.

[0090] Specifically, the maximum value among all vulnerability similarities is taken as the target similarity between the instruction fragment to be processed and the benchmark sample library, reflecting the vulnerability of the instruction fragment to be processed.

[0091] Step S205: Determine the risk level based on target similarity, adopt the injection strategy corresponding to the risk level, and perform adaptive RTL-level error injection on the instruction fragment to be processed to obtain the corresponding silent data corruption problem data.

[0092] Specifically, step S205 above employs an injection strategy corresponding to the risk level to perform adaptive RTL-level error injection on the instruction fragment to be processed, obtaining the corresponding silent data corruption problem data, including: Step S2051: When the risk level is Level 1, perform a verification RTL-level error injection on the instruction fragment to be processed to obtain the corresponding silent data corruption problem data.

[0093] Specifically, step S2051 includes: Step b1: When the risk level is level 1, the injection configuration of the benchmark sample corresponding to the target similarity is used to perform RTL-level error injection on the instruction fragment to be processed and execute it to obtain the second instruction execution result.

[0094] Specifically, when the risk level is level one, it means that the vulnerability pattern of the instruction fragment to be processed is highly similar to that of the benchmark sample corresponding to the target similarity. Therefore, the injection configuration of the benchmark sample is used to inject errors into the instruction fragment to be processed and execute it to obtain the second instruction execution result.

[0095] Step b2: When the execution result of the second instruction shows a silent data corruption problem, the injected instruction fragment to be processed is used as silent data corruption problem data, and the instruction fragment to be processed and its injection configuration are added to the benchmark sample library.

[0096] Specifically, the execution result of the second instruction is compared with its golden output result. When the execution result of the second instruction exhibits an SDC problem, the injected instruction fragment is treated as silent data corruption problem data, and the original instruction fragment to be processed and its injection configuration are added to the benchmark sample library. Optionally, the vulnerability mode of the instruction fragment to be processed can be marked as high risk, indicating that a protection method needs to be designed.

[0097] Step b3: If the execution result of the second instruction does not show a problem of silent data corruption, the fine-tuning configuration of the injection configuration of the benchmark sample corresponding to the target similarity is adopted to perform RTL-level error injection on the instruction fragment to be processed again and execute it to obtain the third instruction result.

[0098] Specifically, when directly reusing the injection configuration of the most similar benchmark sample fails to trigger the SDC issue, it indicates that the instruction fragment to be processed has subtle differences from the benchmark sample, and its variant configuration needs to be tried to further verify the vulnerability. Therefore, fine-tuning is performed based on the injection configuration of the benchmark sample corresponding to the target similarity. For example, in the case of a Warp scheduler error, attempts are made to flip different bits of the thread ID (e.g., changing bit 4 to bit 3), or to change the injection position (e.g., changing from the 3rd instruction to the 5th instruction). The fine-tuned variant configuration is then used to re-inject and execute the instruction fragment to be processed, resulting in the third instruction.

[0099] Step b4: When the result of the third instruction shows a silent data corruption problem, the instruction fragment to be processed after re-injection is taken as silent data corruption problem data, and the instruction fragment to be processed and its re-injection configuration are added to the benchmark sample library.

[0100] Specifically, the execution result of the third instruction is compared with its golden output result. When the execution result of the third instruction exhibits an SDC problem, the re-injected instruction fragment is treated as silent data corruption problem data, and the original instruction fragment to be processed and its injection configuration are added to the benchmark sample library. Optionally, the vulnerability mode of the instruction fragment to be processed can be marked as high risk, indicating that a protection method needs to be designed.

[0101] In some optional implementations, if the execution result of the third instruction still fails to trigger the SDC problem, it indicates that the instruction fragment to be processed is substantially different from the vulnerability patterns in the benchmark sample library, and its risk level is downgraded to the second level, and it is subsequently processed according to the injection strategy corresponding to the second level.

[0102] Step S2052: When the risk level is Level 2, perform boundary exploration RTL-level error injection on the instruction segment to be processed to obtain the corresponding silent data corruption problem data. Level 1 is higher than Level 2.

[0103] Specifically, step S2052 includes: Step c1: When the risk level is level 2, determine the number of injections based on the target similarity.

[0104] Specifically, when the risk level is Level 2, it indicates that the vulnerability patterns of the instruction fragment to be processed are moderately similar to those of the benchmark sample corresponding to the target similarity, and the injection configuration cannot be directly reused; its vulnerability boundary needs to be explored. First, the total potential injection count for the instruction to be processed is calculated, representing the number of possible parameter combinations for erroneous injections within an instruction fragment. This can be calculated by multiplying the number of injection locations, the types of error models, and the number of parameter combinations for each error model. The number of injection locations is the number of instructions in the instruction fragment that can serve as injection points, such as memory access instructions and computation instructions; the number of error types is taken as 3 in this embodiment. The injection density is determined based on the direct proportionality between injection density and target similarity. The product of the injection density and the total potential injection count is calculated as the injection count.

[0105] Step c2: Select the parameter combination for the number of injections from the parameter combination list, perform layered injection of the instruction fragment to be processed, and execute it after each layer of injection to obtain the execution result of the fourth instruction.

[0106] Specifically, in step S2022, a list of parameter combinations has been generated. The number of injection times and parameter combinations are selected according to priority. The key instructions of the instruction fragment to be processed are injected and executed one by one to obtain the execution result of the fourth instruction.

[0107] Step c3: When the execution result of the fourth instruction shows a silent data corruption problem, the injected instruction fragment to be processed is taken as silent data corruption problem data, and the instruction fragment to be processed and its injection configuration are added to the benchmark sample library.

[0108] Specifically, the execution result of the fourth instruction, when compared with its golden output, exhibits a silent data corruption problem, indicating that it has reached the critical condition for SDC (Single Data Conversion) issues, i.e., the vulnerability boundary. For example, the SDC problem is only triggered when the most significant bit of the thread ID is flipped and the injection occurs on a specific instruction. The injected instruction fragment corresponding to the execution result of the fourth instruction is used as the silent data corruption problem data, and the original instruction fragment and its injection configuration are added to the benchmark sample library.

[0109] Step c4: If the execution result of the fourth instruction does not show a problem of silent data corruption, continue to inject the instruction fragment to be processed in layers until the injection number is reached.

[0110] Specifically, after comparing the execution result of the fourth instruction with its golden output result, no silent data corruption problem was found. The original instruction fragment to be processed was injected layer by layer until the number of injections was reached, so that the vulnerability boundary of the instruction fragment to be processed could be fully explored within the set test density.

[0111] Step S2053: When the risk level is level 3, exploratory RTL-level error injection is performed on the instruction fragment to be processed to obtain the corresponding silent data corruption problem data. Level 2 is higher than level 3.

[0112] Specifically, step S2053 includes: Step d1: When the risk level is level 3, random RTL-level error injection is performed on the instruction fragment to be processed and executed to obtain the execution result of the fifth instruction. The random injection is repeated until the preset injection density is reached.

[0113] Specifically, when the risk level is level three, it indicates that the vulnerability patterns of the instruction fragment to be processed are low compared to the baseline sample corresponding to the target similarity. In this case, an exploratory injection strategy is adopted. First, an error is randomly selected from the parallel management errors, a set of parameters is randomly selected from its corresponding combinations, and an instruction is randomly selected from the instruction fragment to be processed for injection and execution, resulting in the execution result of the fifth instruction. The number of injections is then accumulated. The above random injection process is repeated until the ratio of the number of injections to the total number of potential injections reaches a preset injection density, which is a relatively low fixed value.

[0114] Step d2: When the execution result of any fifth instruction exhibits a silent data corruption problem, the injected instruction fragment corresponding to the execution result of the fifth instruction is taken as the silent data corruption problem data, and the instruction fragment to be processed and its injection configuration are added to the benchmark sample library.

[0115] Specifically, during the random exploration process, if any fifth instruction execution result shows a silent data corruption problem after comparison with its golden output result, the injected instruction fragment corresponding to the fifth instruction execution result is taken as silent data corruption problem data, and the original instruction fragment to be processed and its injection configuration are added to the benchmark sample library.

[0116] Step d3: Obtain the neighbor nodes of the instruction fragment to be processed from the instruction graph where the instruction fragment to be processed is located, corresponding to the execution result of the fifth instruction.

[0117] Specifically, in the instruction graph where the instruction fragment to be processed corresponds to the execution result of the fifth instruction is located, the first-order and second-order neighbor nodes of the node corresponding to the instruction fragment to be processed are obtained. The instruction fragments corresponding to these neighbor nodes are closely related to the current instruction fragment to be processed in terms of function and logic, and therefore also have potential vulnerabilities.

[0118] Step d4: Based on the instruction fragment to be processed corresponding to the neighbor node, return to the step of calculating the target similarity between the instruction fragment to be processed and the target of the benchmark sample library using the target similarity calculation model.

[0119] Specifically, the process returns to step 204 to recalculate the target similarity between the instruction fragments corresponding to neighboring nodes and the updated benchmark sample library. Since the benchmark sample library now includes the instruction fragments that cause the SDC problem, the target similarity of these neighboring nodes is likely to jump from low to medium or even high. Based on the target similarity, these neighboring nodes will automatically enter the aforementioned boundary exploratory injection or verification injection process, thereby achieving concentrated and efficient mining of newly discovered vulnerable regions.

[0120] In some optional implementations, after completing adaptive RTL-level error injection for all instruction fragments to be processed, the SDC (Simplified Data Concern) sensitivity information for each instruction fragment is recorded, including: whether an SDC problem occurs, under what injection configuration the SDC problem occurs, and the frequency of SDC problem occurrence. Then, this SDC sensitivity information is correlated back to the corresponding specific modules and code segments in the remote sensing large model. A visual SDC vulnerability heatmap is generated with the module or code segment as the horizontal axis and the vulnerability level (such as SDC problem occurrence rate and risk level) as the vertical axis or color depth, intuitively showing "which modules and code segments are susceptible to hardware transient failures."

[0121] This invention constructs a sample instruction map of the training program for a large-scale remote sensing model, systematically organizing the underlying instruction sequences into a graph structure rich in semantic information. Then, by injecting RTL-level errors into the sample instruction fragments in the sample instruction map to update the benchmark sample library, vulnerable pattern samples can be obtained in a targeted and systematic manner. Next, a similarity calculation model is trained based on the sample instruction map, enabling the model to assess the similarity of vulnerability among instruction fragments. For each instruction fragment to be processed in the training program, the trained model calculates its target similarity with the benchmark sample library, quantifying its potential vulnerability. Finally, the risk level is determined based on the target similarity, and a corresponding adaptive injection strategy is adopted to obtain silent data corruption problem data. By transforming the generation of silent data corruption problem data from an inefficient random injection method to a highly efficient and intelligent generation method based on data and models, the generation efficiency of silent data corruption problem data is improved, effectively enhancing the efficiency of analyzing error patterns caused by the propagation of silent data corruption problems during training, and providing technical support for ensuring the reliability of large-scale, long-term remote sensing training tasks.

[0122] The present invention has the following beneficial effects: (1) Unlike random software instruction level bit flip error injection, this invention simulates injection by mapping specific error types (especially parallel management errors) of the underlying hardware (RTL level) to upper-level software instructions, making the injection closer to the physical layer and improving the efficiency and targeting of generating SDC problems; (2) Construct a multi-dimensional instruction map of GPU instruction sequences. Based on this instruction map, extract the feature vectors of instruction fragments and establish a model to calculate the vulnerability similarity between any instruction fragment and the known SDC benchmark sample library. (3) A non-uniform intelligent injection strategy based on vulnerability similarity is proposed, which aims to maximize the efficiency of discovering new vulnerabilities while avoiding wasting computational resources on known vulnerability patterns. (4) Based on the above method, a fine-grained SDC vulnerability heatmap is generated. Based on this heatmap, targeted code protection, redundant computation insertion or optimized training strategies (such as dynamically adjusting the checkpoint frequency) can be designed, thus forming a complete closed loop from vulnerability discovery to hardening protection.

[0123] This embodiment also provides a data generation apparatus for addressing the problem of silent data corruption. This apparatus is used to implement the above embodiments and preferred embodiments, and details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the apparatus described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0124] This embodiment provides a data generation device for addressing the problem of silent data corruption, such as... Figure 2 As shown, it includes: Module 210 is used to construct a sample instruction map of the training program for a large remote sensing model.

[0125] Update module 220 is used to update the baseline sample library that causes silent data corruption problems by injecting RTL-level errors into sample instruction fragments in the sample instruction graph.

[0126] Training module 230 is used to train an initial similarity calculation model based on the sample instruction map to obtain a target similarity calculation model.

[0127] The calculation module 240 is used to calculate the target similarity between the instruction segment to be processed and the benchmark sample library for each instruction segment to be processed in the training program, using a target similarity calculation model.

[0128] The generation module 250 is used to determine the risk level based on the target similarity, and to perform adaptive RTL-level error injection on the instruction fragment to be processed using the injection strategy corresponding to the risk level, so as to obtain the corresponding silent data corruption problem data.

[0129] In some alternative implementations, construction module 210 includes: The first extraction unit is used to extract sample instruction programs from the training program of the remote sensing large model.

[0130] The first partitioning unit is used to divide the sample instruction program into multiple sample instruction segments by identifying the instruction segment boundaries of the sample instruction program, and to treat each sample instruction segment as a node.

[0131] The second partitioning unit is used to perform multi-dimensional partitioning on each sample instruction fragment to obtain the attribute features of each sample instruction fragment, which are used as the node attributes of the corresponding node.

[0132] The building unit is used to construct edges based on the control flow and data dependencies between all sample instruction fragments to obtain the sample instruction graph.

[0133] In some alternative implementations, update module 220 includes: The filtering unit is used to filter multiple seed nodes from the sample instruction graph.

[0134] The scanning unit is used to perform parameterized scanning for each RTL level error to obtain a list of parameter combinations.

[0135] The injection unit is used to inject RTL-level errors into the sample instruction fragment corresponding to each seed node according to the parameter combination list and execute it to obtain the first instruction execution result.

[0136] The storage unit is used to add the sample instruction fragment as a baseline sample and its corresponding injection configuration to the baseline sample library when the execution result of the first instruction shows a silent data corruption problem.

[0137] In some alternative implementations, training module 230 includes: The second extraction unit is used to extract the sample feature vector of each node in the sample instruction graph and enumerate all vector pairs formed by the sample feature vectors.

[0138] The first calculation unit is used to map each pair of vectors into an embedding vector using an initial similarity calculation model, and to calculate the sample similarity. The data dimension of the embedding vector is lower than that of the sample feature vector.

[0139] The training unit is used to train an initial similarity calculation model based on vector pairs with sample labels to obtain a target similarity calculation model. The sample labels indicate whether the corresponding sample instruction fragments of each vector pair are similar.

[0140] In some alternative implementations, the second extraction unit includes: The acquisition sub-unit is used to acquire the static and dynamic features of the sample instruction fragment corresponding to each node in the sample instruction map.

[0141] The first determining sub-unit is used to determine the in-degree and out-degree of a node in the sample instruction graph.

[0142] The first computational subunit is used to enumerate all node pairs in the sample instruction graph, construct the shortest path between each pair of nodes, and calculate the betweenness centrality of nodes based on the number of shortest paths passing through nodes and the total number of shortest paths for all node pairs.

[0143] The second computational subunit is used to identify critical path nodes in the sample instruction graph and calculate the shortest distance from the node to the critical path node.

[0144] The second determining unit is used to take the in-degree, out-degree, betweenness centrality, and shortest distance of the node as the sample graph features of the node, and to concatenate the sample graph features, static features, and dynamic features to obtain the sample feature vector of the node.

[0145] In some alternative implementations, the computing module 240 includes: The third extraction unit is used to extract the first feature vector of each instruction segment to be processed in the training program, and to extract the second feature vector of each benchmark sample in the benchmark sample library.

[0146] The second calculation unit is used to calculate the vulnerability similarity between the target instruction fragment and each benchmark sample in the benchmark sample library by adopting the target similarity calculation model, based on the first feature vector and the second feature vector of each benchmark sample.

[0147] A unit is defined as the maximum value among all vulnerability similarities, which is used as the target similarity.

[0148] In some alternative implementations, the generation module 250 includes: The first generation unit is used to perform verification RTL-level error injection on the instruction fragment to be processed when the risk level is the first level, so as to obtain the corresponding silent data corruption problem data.

[0149] The second generation unit is used to perform boundary exploration RTL-level error injection on the instruction fragment to be processed when the risk level is the second level, so as to obtain the corresponding silent data corruption problem data. The first level is higher than the second level.

[0150] The third generation unit is used to perform exploratory RTL-level error injection on the instruction fragment to be processed when the risk level is level three, so as to obtain the corresponding silent data corruption problem data. Level two is higher than level three.

[0151] In some alternative implementations, the first generating unit includes: The first injection subunit is used to inject RTL-level errors into the instruction fragment to be processed and execute it when the risk level is the first level, using the injection configuration of the benchmark sample corresponding to the target similarity, to obtain the second instruction execution result.

[0152] The first generation subunit is used to treat the injected instruction fragment as silent data corruption problem data when the execution result of the second instruction is a silent data corruption problem, and to add the instruction fragment to be processed and its injection configuration to the benchmark sample library.

[0153] The second injection subunit is used to fine-tune the injection configuration of the benchmark sample corresponding to the target similarity when the execution result of the second instruction does not show a silent data corruption problem. It then performs RTL-level error injection on the instruction fragment to be processed again and executes it to obtain the third instruction result.

[0154] The second generation subunit is used to treat the re-injected instruction fragment as the silent data corruption problem data when the result of the third instruction is a silent data corruption problem, and to add the instruction fragment to be processed and its re-injection configuration to the benchmark sample library.

[0155] In some alternative implementations, the second generating unit includes: A sub-unit is defined to determine the number of injections based on target similarity when the risk level is level two.

[0156] The third injection subunit is used to select the parameter combination for the number of injections from the parameter combination list, perform layered injection of the instruction fragment to be processed, and execute it after each layer of injection to obtain the fourth instruction execution result.

[0157] The third generation subunit is used to treat the injected instruction fragment as silent data corruption problem data when the execution result of the fourth instruction is a silent data corruption problem, and to add the instruction fragment to be processed and its injection configuration to the benchmark sample library.

[0158] The fourth injection subunit is used to continue injecting the instruction fragment to be processed in layers when the execution result of the fourth instruction does not show a problem of silent data corruption, until the injection number is reached.

[0159] In some alternative implementations, the third generation unit includes: The fifth injection subunit is used to perform random RTL-level error injection and execution on the instruction fragment to be processed when the risk level is level three, so as to obtain the execution result of the fifth instruction. The random injection is repeated until the preset injection density is reached.

[0160] The fourth generation subunit is used to take the injected instruction fragment corresponding to the execution result of the fifth instruction as the silent data corruption problem data when the execution result of any fifth instruction shows a silent data corruption problem, and add the instruction fragment to be processed and its injection configuration to the benchmark sample library.

[0161] The acquisition sub-unit is used to acquire the neighbor nodes of the instruction segment to be processed from the instruction graph where the instruction segment to be processed is located, corresponding to the execution result of the fifth instruction.

[0162] The fifth generation subunit is used to return to the step of calculating the target similarity between the instruction fragment to be processed and the benchmark sample library based on the instruction fragment to be processed corresponding to the neighbor node.

[0163] The data generation apparatus for the silent data corruption problem provided in this embodiment of the invention can execute the data generation method for the silent data corruption problem provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects for executing the method. Further functional descriptions of the above modules and units are the same as in the corresponding embodiments described above, and will not be repeated here.

[0164] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0165] The following is a detailed reference. Figure 3 The diagram illustrates a structural schematic suitable for implementing an electronic device according to embodiments of the present invention. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 301, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 302 or a program loaded from memory 308 into random access memory (RAM) 303. The RAM 303 also stores various programs and data required for the operation of the electronic device. The processor 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0166] Typically, the following devices can be connected to I / O interface 305: input devices 306 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 307 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 308 including, for example, magnetic tapes, hard disks, etc.; and communication devices 309. Communication device 309 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 3 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0167] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 309, or installed from a memory 308, or installed from a ROM 302. When the computer program is executed by the processor 301, it performs the functions defined in the data generation method for solving the silent data corruption problem according to embodiments of the present invention.

[0168] Figure 3 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0169] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the data generation method for solving the silent data corruption problem shown in the above embodiments is implemented.

[0170] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A data generation method for silencing data corruption problems, characterized by, The method includes: A sample instruction map of the training program for constructing a large remote sensing model; By injecting RTL-level errors into sample instruction fragments in the sample instruction map, the baseline sample library that causes silent data corruption is updated. An initial similarity calculation model is trained based on the sample instruction map to obtain a target similarity calculation model; For each instruction segment to be processed in the training program, the target similarity calculation model is used to calculate the target similarity between the instruction segment to be processed and the benchmark sample library; Based on the target similarity, the risk level is determined, and the injection strategy corresponding to the risk level is adopted to perform adaptive RTL-level error injection on the instruction fragment to be processed, so as to obtain the corresponding silent data corruption problem data.

2. The method of claim 1, wherein, The sample instruction map of the training program for constructing the large remote sensing model includes: Extract sample instruction program from the training program of the aforementioned large remote sensing model; By identifying the instruction segment boundaries of the sample instruction program, the sample instruction program is divided into multiple sample instruction segments, and each sample instruction segment is treated as a node. Each sample instruction fragment is divided into multiple dimensions to obtain the attribute features of each sample instruction fragment, which are used as the node attributes of the corresponding node. Edges are constructed based on the control flow and data dependencies between all sample instruction fragments to obtain the sample instruction graph.

3. The method of claim 2, wherein, The step of updating the baseline sample library that causes silent data corruption by injecting RTL-level errors into sample instruction fragments in the sample instruction map includes: Multiple seed nodes are selected from the sample instruction map; Perform a parameterized scan for each RTL level error to obtain a list of parameter combinations; For each seed node, RTL-level errors are injected into the sample instruction fragment corresponding to the seed node according to the parameter combination list and executed to obtain the first instruction execution result; When the execution result of the first instruction shows a silent data corruption problem, the sample instruction fragment is used as a benchmark sample, and the benchmark sample and its corresponding injection configuration are added to the benchmark sample library.

4. The method of claim 3, wherein, The process of training an initial similarity calculation model based on the sample instruction map to obtain a target similarity calculation model includes: Extract the sample feature vector of each node in the sample instruction map, and enumerate all vector pairs formed by the sample feature vectors; Using the initial similarity calculation model, each vector pair is mapped to an embedding vector, and the sample similarity is calculated. The data dimension of the embedding vector is lower than that of the sample feature vector. An initial similarity calculation model is trained based on vector pairs with sample labels to obtain a target similarity calculation model. The sample labels indicate whether the corresponding sample instruction fragments of each vector pair are similar.

5. The method of claim 4, wherein, The extraction of the sample feature vector for each node in the sample instruction map includes: For each node in the sample instruction map, obtain the static and dynamic features of the sample instruction fragment corresponding to the node; Determine the in-degree and out-degree of the node in the sample instruction graph; Enumerate all node pairs in the sample instruction graph, construct the shortest path between each pair of nodes, and calculate the betweenness centrality of the node based on the number of shortest paths passing through the node and the total number of shortest paths for all node pairs. Identify the key path nodes in the sample instruction graph and calculate the shortest distance from the node to the key path node; The in-degree, out-degree, betweenness centrality, and shortest distance of the node are used as the sample graph features of the node. The sample graph features, the static features, and the dynamic features are concatenated to obtain the sample feature vector of the node.

6. The method of claim 4, wherein, For each target instruction fragment to be processed in the training program, the target similarity calculation model is used to calculate the target similarity between the target instruction fragment to be processed and the target in the benchmark sample library, including: For each instruction segment to be processed in the training program, the first feature vector of the instruction segment to be processed is extracted, and the second feature vector of each benchmark sample in the benchmark sample library is extracted. Using the target similarity calculation model, based on the first feature vector and the second feature vector of each benchmark sample, the vulnerability similarity between the target instruction fragment and each benchmark sample in the benchmark sample library is calculated. The maximum value among all vulnerability similarities is taken as the target similarity.

7. The method of claim 6, wherein, The method employs the injection strategy corresponding to the risk level to perform adaptive RTL-level error injection on the target instruction fragment to be processed, thereby obtaining the corresponding silent data corruption problem data, including: When the risk level is the first level, a verification RTL-level error injection is performed on the instruction fragment to be processed to obtain the corresponding silent data corruption problem data. When the risk level is the second level, boundary exploration RTL-level error injection is performed on the instruction fragment to be processed to obtain the corresponding silent data corruption problem data. The first level is higher than the second level. When the risk level is level three, exploratory RTL-level error injection is performed on the instruction fragment to be processed to obtain the corresponding silent data corruption problem data. The second level is higher than the third level.

8. The method of claim 7, wherein, When the risk level is Level 1, a verification RTL-level error injection is performed on the instruction fragment to be processed to obtain the corresponding silent data corruption problem data, including: When the risk level is the first level, the injection configuration of the benchmark sample corresponding to the target similarity is used to inject RTL-level errors into the instruction fragment to be processed and execute it to obtain the second instruction execution result; When the execution result of the second instruction shows a silent data corruption problem, the injected instruction fragment to be processed is used as silent data corruption problem data, and the instruction fragment to be processed and its injection configuration are added to the benchmark sample library; When the execution result of the second instruction does not show a problem of silent data corruption, the fine-tuning configuration of the injection configuration of the benchmark sample corresponding to the similarity of the target is adopted to inject RTL-level errors into the instruction fragment to be processed again and execute it to obtain the third instruction result; When the result of the third instruction is a silent data corruption problem, the instruction fragment to be processed after re-injection is taken as silent data corruption problem data, and the instruction fragment to be processed and its re-injection configuration are added to the benchmark sample library.

9. The method of claim 7, wherein, When the risk level is level two, boundary exploration RTL-level error injection is performed on the instruction fragment to be processed to obtain the corresponding silent data corruption problem data, including: When the risk level is level two, the number of injections is determined based on the target similarity; Select the parameter combination for the number of injections from the parameter combination list, perform layered injection on the instruction fragment to be processed, and execute it after each layer of injection to obtain the fourth instruction execution result; When the execution result of the fourth instruction shows a silent data corruption problem, the injected instruction fragment to be processed is used as silent data corruption problem data, and the instruction fragment to be processed and its injection configuration are added to the benchmark sample library; If the execution result of the fourth instruction does not show a problem of silent data corruption, the layered injection of the instruction fragment to be processed continues until the injection count is reached.

10. The method of claim 7, wherein, When the risk level is level three, exploratory RTL-level error injection is performed on the instruction fragment to be processed to obtain the corresponding silent data corruption problem data, including: When the risk level is level three, random RTL-level error injection is performed on the instruction fragment to be processed and executed to obtain the fifth instruction execution result. The random injection is repeated until the preset injection density is reached. When any fifth instruction execution result exhibits a silent data corruption problem, the injected instruction fragment corresponding to the fifth instruction execution result is taken as silent data corruption problem data, and the instruction fragment to be processed and its injection configuration are added to the benchmark sample library; Obtain the neighbor nodes of the instruction fragment to be processed from the instruction graph where the instruction fragment to be processed is located, corresponding to the execution result of the fifth instruction; Based on the instruction fragment to be processed corresponding to the neighbor node, return to the step of calculating the target similarity between the instruction fragment to be processed and the benchmark sample library using the target similarity calculation model.