CN109063775A - Instruction SDC fragility prediction technique based on shot and long term memory network - Google Patents
Instruction SDC fragility prediction technique based on shot and long term memory network Download PDFInfo
- Publication number
- CN109063775A CN109063775A CN201810893739.8A CN201810893739A CN109063775A CN 109063775 A CN109063775 A CN 109063775A CN 201810893739 A CN201810893739 A CN 201810893739A CN 109063775 A CN109063775 A CN 109063775A
- Authority
- CN
- China
- Prior art keywords
- instruction
- sdc
- fragility
- feature
- shot
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Debugging And Monitoring (AREA)
Abstract
The invention discloses a kind of instruction SDC fragility prediction techniques that is novel, being based on shot and long term memory network (LSTM), when obtaining Silent Data Corruption (SDC) fragility correlated characteristic of code each instruction, we no longer need a large amount of direct fault location that can find those of most fragile instruction in program by prediction model.The SDC fragility of previous instruction, which needs to carry out a large amount of direct fault location operation, can just obtain, this process is extremely time-consuming, the present invention is analyzed by dependence characteristics of the inherent feature and instruction to instruction itself in propagation path, it therefrom finds out and is associated with maximum feature with the SDC fragility of instruction, and combine the shot and long term memory network model for being good at processing sequence data, realize the identification to instruction SDC fragility, and a large amount of direct fault location is no longer needed to operate, save a large amount of time and resource.
Description
Technical field
It is of the invention a kind of pre- to the SDC fragility progress of instruction in the general LLVM intermediate code of distinct program language
The method of survey.According to the inherent feature and dependence characteristics of the intermediate code instruction extracted, learnt by shot and long term memory network
How the SDC fragility of instruction is predicted.
Background technique
The performance of processor, diminution processing are continuously improved by new technology by nearest decades, the designer of processor
The size of device, but processor also becomes more and more fragile and unreliable, processor is than being easier that transient fault occurs in the past
(transient faults).Mistake in transient fault and processor design is different, and transient fault is interruption, occurs
It will restore after a period of time later, hardware circuit will not be damaged, this transient fault is referred to as soft error.Soft error will lead to
The program being currently running breaks down, the change of the value of transmitting and storage by influencing signal, cause such as satellite out of control etc.
The generation of accident.Soft error can influence the program being currently running in a manner of 3 kinds: (1) they will not be to the program being currently running
Impact (benign/mask), (2) they may make program crashing or hang up (crash or hang), (3) they
The output for leading to mistake be will lead to as a result, namely Silent Data Corruption (SDC) problem.Compared to collapsing for program
It bursts and hangs up, SDC problem is more hidden, and once may result in serious consequence.In order to solve these soft errors
Caused by SDC problem, designer has usually introduced to be made in the reinforcement means of hardware redundancy, such as memory (cache, memory)
Detect this failure with ECC and parity check bit, but the cost too expensive of method that this hardware is reinforced, cost is too
Height is not appropriate for desktop computer, notebook market.
Software-based redundancy reinforcement means provides one and consumes lower and more flexible selection, software-based
In redundancy reinforcement means, selective redundancy reinforcing mode is most advantageous, it can be in the feelings that program SDC fragility is effectively reduced
Reducing redundancy reinforcement means brings time and space consuming simultaneously under condition.Before the reinforcing for carrying out selectivity, in program
The prediction of the fragility of the SDC of each instruction makes most critical, and the instruction for being only correctly found most fragile in program could body
The advantage now selectively reinforced.Direct fault location is the simplest mode for finding the instruction of SDC fragility, although by intermediate code
The direct fault location of the certain number of instruction progress and the SDC fragility for the available instruction of frequency for counting its SDC problem, this
Kind mode very time-consuming, particularly with large program.Traditional machine learning (machine learning) is in such case
Under the detection that be used to instruct SDC fragility, obtained by extracting with the relevant feature of SDC fragility and direct fault location is instructed
Initial labels, classify using SDC fragility of the machine learning models such as support vector machines (SVM) and decision tree to instruction
Or regression forecasting.But deep learning model effect the problem of a large amount of predictions are classified far had surpassed traditional engineering in recent years
Learning method, wherein Recognition with Recurrent Neural Network (RNN) is a kind of neural network for processing sequence data and effect is very outstanding,
But the problem of due to gradient explosion, traditional RNN can not handle long-term Dependence Problem, and shot and long term memory network (LSTM) is then
It is the modified version of Recognition with Recurrent Neural Network (RNN), is born also for this problem is solved, and is due to instruction execution suitable
Sequence, the ability value in terms of instruction SDC fragility classification prediction must be explored.
Summary of the invention
It is an object of the invention in the code write to distinct program language, by extracting consolidating for its LLVM intermediate command
There are feature and dependence characteristics, completes to predict the classification of instruction SDC fragility using shot and long term memory network model, mainly include
The following contents:
1) extraction of SDC Vulnerability Characteristics and fragility label obtain.LLFI direct fault location tool is to intermediate code command
Destination register carry out direct fault location, by destination register, each SDC fragility takes the average SDC fragility as instruction
Property, while the trace files of program execution are obtained, the Dynamic Execution number etc. of instruction therein is extracted by analyzing trace files
Dynamic inherent feature.Different language has different characteristic and grammer, so if from the source code of program to program middle finger
The cost that the feature of order extracts is very big, and LLVM compiler provides general intermediate generation to different program languages
Code, solves the problems, such as this, so carrying out the extraction of direct fault location and feature to the LLVM intermediate code of distinct program to construct
Data set.In the generation and communication process of SDC problem, instruct some features of itself that may have a certain impact, these
Feature includes type, operand type, data width, the loop nesting depth etc. of instruction.For example the instruction of arithmetic types is being sent out
The probability of raw SDC problem is apparently higher than address calculation instructions, so different types of instruction itself has different SDC fragilities
Property.In a cycle, the depth of circulation is higher, and instruction therein is often more crucial, so the nested depth of round of instruction
It has a certain impact to the SDC fragility of instruction.The spy that the propagation of SDC problem can be impacted in the propagation path of instruction
Sign includes mask instruction, address calculation instructions, operand type etc., these features are referred to as dependence characteristics.Mask instruction refers to
Logical operation and shift operation instruction, these instructions often have certain cover to make the instruction that mistake occurs or is propagating mistake
With.And address calculation instructions are more likely to that collapse occurs if error or hang up (crash/hang).So in order to every
The locating environment of item instruction is described in more detail, and the pass using LLVM is to instructing itself inherent feature and its local environment
In dependence characteristics extract.
2) feature selecting.For classifier, including shot and long term memory network (LSTM), the feature the more not to represent
More useful informations, at this moment the performance of classifier can decline instead with the increase of characteristic dimension.Cause under classifier performance
The reason of drop, is in those high-dimensional features to contain extraneous features and redundancy feature.Therefore, it is desirable to train one efficiently
Compact disaggregated model again also needs to select the feature of extraction, removes those before carrying out sorter model training
The information low to SDC problem importance retains those information high to SDC problem importance.The method of feature selecting is divided into
This 3 kinds of filter, wrapper, embedding, we select the single argument feature selecting (Univariate in filter
Feature selection) method screens feature.Single argument feature selection approach is by being independently of sorting algorithm
A kind of method based on statistics, it tests each feature, calculates some statistical indicator of each feature to measure spy
It seeks peace the relationship predicted between classification.For classification task, single argument feature selecting between the feature and label of data set into
The P-value value of each feature is calculated in row variance analysis (ANOVA), and according to Principle of Statistics, P-value is smaller then
Illustrate that the importance of feature is stronger, it is possible to the importance scores of each feature be obtained by log function, and heavy according to this
The property wanted score screens feature.
3) train classification models.The SDC fragility of instruction is substantially a classification problem, in order to obtain best classification
Effect needs to select most suitable model.After distinct program is converted to LLVM intermediate code, the feature of each instruction
Data are extracted, and obtain the data of static sequence instruction, and shot and long term memory network (LSTM) has on processing sequence data problem
Big advantage, so shot and long term memory network (LSTM) is selected as training pattern.But shot and long term memory network possesses
Different model parameters, these different model parameters all have a certain impact to final model prediction result, so we
According to the different model of the network number of plies of most critical, cell number, droupout rate, this 4 parameter settings of learning rate, model is formed
Data are used for each of these model sets model and are trained, finally select accuracy from these models by set
Optimal models are as last prediction model.
Detailed description of the invention
Fig. 1 is method overall framework figure proposed by the present invention；
Fig. 2 is characterized the score chart of different characteristic after selection；
Fig. 3 is dataset construction flow chart；
Fig. 4 is prediction model training flow chart；
Fig. 5 is shot and long term memory network internal structure chart.
Specific embodiment
Specific introduction is done to the present invention below in conjunction with drawings and concrete examples.
The benchmark collection Mibench that the present invention will acquire is as test program, and therefrom selection a part represents
Property strong program, be related to automobile and industry manufacture, consumer electronics, office automation, network, safety communicates six classes.From selecting
Program LLVM intermediate code in extract and instruct the relevant inherent feature of SDC fragility and dependence characteristics.LLFI injection is used
In the SDC fragility value of acquisition instruction, and according to injection result of the LLFI implantation tool on test program to the SDC of instruction
Fragility is classified, and the building and division of data acquisition system are completed.Finally, crisp to the SDC of instruction by shot and long term memory network
Weak property is learnt, and the instruction SDC fragility prediction model based on LSTM is obtained.The general frame of method of the invention such as Fig. 1
Shown, specific implementation process is as follows:
The extraction of step 1:SDC Vulnerability Characteristics and fragility label obtain
The acquisition of instruction SDC fragility label is obtained by LLFI direct fault location tool, program operation after injection
As a result when different with the program operation result before injection, then it is assumed that SDC mistake occurs for program.To the destination register of instruction
Each is injected, and the number of injection is Ti, SiFor the number that SDC mistake occurs in the number of injection, WiIt is posted for instruction purpose
The data width of storage, using purpose calculator each SDC fragility mean value as the SDC fragility label of instruction, it is public
Formula is as follows.
The feature extraction of the SDC fragility of instruction includes inherent feature and dependence characteristics.Traversal program intermediate code it is every
One instruction carries out the information such as the basic block message of instruction, place function information, instruction type by LLVM compiler frame
Obtain, obtain the static inherent feature of each instruction, such as instruction Dynamic Execution number of dynamic inherent feature in addition and
The call number of function then is analyzed to obtain by the trace files for generating LLFI direct fault location tool where instruction.Equally
Each instruction in traversal program, according to the characteristic of the Static Single Assignment of LLVM intermediate code, passes through the def- in LLVM
Use chain obtains other instruction sets for using present instruction result, constantly iteration this operation, until reaching end instruction
(store, br, call, because store is instructed and br instruction is all without destination register, call instruction can then generate new stack for this
Frame, these instructions can all terminate the propagation of data), the instruction encountered in iterative process is all added in set, is ultimately formed
The propagation path of every instruction, and pass through the dependence characteristics that LLVM compiler frame extracts instruction from path, as mask is instructed
Number, be related to the number of instructions of address calculation, the dependence characteristics such as type of operand.Construction process such as Fig. 3 institute of data set
Show.
Step 2: feature selecting
Feature selecting, after carrying out feature extraction, the redundancy or nothing of extraction are carried out to the SDC Vulnerability Characteristics of instruction
It excessively will lead to the performance decline of classifier with feature.Single argument feature selection approach is by being to be based on independently of sorting algorithm
A kind of method of statistics, for classification task, single argument feature selecting carries out variance point between the feature and label of data set
It analyses (ANOVA), initially sets up hypothesis, it is believed that it is not related between characteristic variable and label, inspection level is set, and default value is general
It is 0.05.Then, the sum of sguares of deviation from mean SS that always makes a variation calculated according to sample data set and capable from mean square and MS, further according to
In group between group from mean square and F-value value being calculated.Finally, being distributed according to obtained F-value value and corresponding F
Probability density function find corresponding P-value value, when P-value value be less than setting inspection level when, then can refuse
It is not related between the characteristic variable done and label it is assumed that i.e. this feature is important label before absolutely.Meanwhile it obtaining
P-value value it is smaller, illustrate that the importance of feature is stronger, it is possible to pass through log function in following formula and mapping letter
Number will obtain the importance scores of each feature and map that [0,1], and be carried out according to this importance scores to feature
Screening.
The importance scores of each feature are as shown in Figure 2 after feature selecting.
score′i=-log10(P-valuei)
Step 3: train classification models
After distinct program is converted to LLVM intermediate code, the characteristic of each instruction is extracted, and is obtained quiet
The data of state sequence instruction, shot and long term memory network (LSTM) have big advantage on processing sequence data problem,.Shot and long term
Memory network possesses different model parameters, these different parameter settings can obtain different models, and can be to training result
Certain influence is generated, so we are according to the network number of plies of most critical, cell number, droupout rate, learning rate this 4 parameters
Different models is set, forms candidate family set, obtained data set is used for each of these candidate family set
Model and the accuracy rate on test set for recording each model, finally according to the accuracy rate of model from these candidate families
Optimal models are selected as last prediction model.
For data set D={ X of the feature selecting after processed1, X2..., Xi..., Xd, each of them data are Xi
={ x1, x2..., xn, y }, training set and test set are constructed by 5: 1, before training by data set according to time series
Timestep is divided into k equal part, and timestep*k=d is trained as follows:
Step 1. is according to this 4 parameters of different model parameters such as the network number of plies, cell number, droupout rate, learning rate
Different model set M={ m is set1, m2..., mg}。
Step 2. takes out a model from model set, carries out the initialization of network parameter.
Timestep data input network in training set is trained by step 3., takes the defeated of the last one time step
Out as the output of hidden layer, and final classification results are exported after softmax function as the input of full articulamentum.Most
Afterwards, it gives output result to cross entropy loss function and calculates penalty values.
If step 4. penalty values are not converged, repeatedly step 3, and according to the net of the continuous iteration of learning rate more new model
Network parameter reaches convergence until penalty values.
Obtained convergent model is used for training set by step 5., the accuracy rate being recorded on training set, if Models Sets
Closing M, there are also remaining models not to use, then goes to step 2.
Step 6. picks out the highest model of accuracy rate as optimal models from model set M.
Training prediction model flow chart is as shown in Figure 4.
Claims (6)
1. the instruction SDC fragility prediction technique based on shot and long term memory network, it is characterised in that:
1) instruction that this method is directed to is the general LLVM intermediate code instruction of program；
2) the SDC fragility of instruction is defined, and passes through LLVM based Fault Injection (LLFI) direct fault location work
Tool obtains the SDC fragility of every instruction in program；
3) feature relevant to instruction SDC fragility is extracted from program intermediate code instruction itself and propagation path；
4) after extracting instruction features, feature selecting is carried out to the SDC Vulnerability Characteristics of instruction；
5) this method is shot and long term memory network applying in instruction SDC fragility classification prediction for the first time.
2. the instruction SDC fragility prediction technique based on shot and long term memory network as described in claim 1, it is characterised in that this
The instruction that method is directed to is that the general LLVM intermediate code instruction of program has in the detection research of SDC fragility in face of program
Source code, also have in face of program assembly code, however, being program source code or assembly code, their type is too many
And it is complicated, very big trouble is brought to the detection of SDC, with the help of LLVM compiler, any language can be converted to
General LLVM intermediate code, this brings great convenience to the detection of SDC.Compared to the SDC of basic block and function fragility
Property prediction, the SDC fragility prediction of instruction-level more can obtain to fine granularity the fragile information of program.
3. the instruction SDC fragility prediction technique based on shot and long term memory network as described in claim 1, it is characterised in that fixed
The SDC fragility of justice instruction, and every SDC fragility instructed in program is obtained by LLFI direct fault location, for program
In each LLVM intermediate code instruct Ii, it SDC fragility instruction destination register in each injection obtain
The SDC fragility of the average value of SDC fragility, instruction is defined as follows.
Wherein, WiRepresent instruction IiDestination register bit wide, SiIt represents and carries out direct fault location using LLFI direct fault location tool
The number of SDC mistake, T occur afterwardsiIndicate the number of progress direct fault location.
4. the instruction SDC fragility prediction technique based on shot and long term memory network as described in claim 1, it is characterised in that from
Feature relevant to instruction SDC fragility, the intrinsic spy of instruction are extracted in program intermediate code instruction itself and propagation path
Size, instruction type, the instruction Dynamic Execution number inherent feature of sign such as instruction place basic block can be to a certain extent
Reflect the SDC fragility of instruction, instructs the feature on propagation path to be known as dependence characteristics, they have on the path of propagation can
More serious mistake can be covered or caused to the SDC mistake that instruction occurs, number, behaviour's institute's number including mask instruction
Type, address calculation instructions number, address calculation instructions operand type etc..
5. the instruction SDC fragility prediction technique described in claim 1 based on shot and long term memory network, it is characterised in that extract
After instruction features, feature selecting, after carrying out feature extraction, the redundancy of extraction are carried out to the SDC Vulnerability Characteristics of instruction
Or useless feature excessively will lead to the performance decline of classifier.Monotropic measure feature choosing is carried out to the data set that feature extraction obtains
It selects, variance analysis (ANOVA) is carried out to data set, the relationship between feature and class label is measured according to statistic, is calculated
The F-value of each feature simultaneously obtains P-value according to it, then calculates the importance point of each feature according to the following formula
It counts and is mapped to [0,1].
score′i=-log10(P-valuei)
Wherein, F-value is the ratio of Mean squares between groups and Mean squares within group, and F-value obeys F distribution, and P-value is for determining spy
The parameter of relevance between sign and label is obtained by inquiring F distribution table, when it is lower than inspection level, then it is assumed that feature
It is important to label, and P-value is smaller then more important.
6. the instruction SDC fragility prediction technique described in claim 1 based on shot and long term memory network, it is characterised in that we
Method is for the first time application of the shot and long term memory network in instruction SDC fragility classification prediction, and the instruction SDC proposed in recent years is fragile
Property prediction technique is based primarily upon traditional machine learning method, including support vector machines, support vector regression, post-class processing
Deng, but based on being put forward for the first time when the instruction SDC fragility prediction mode of shot and long term memory network in deep learning.Feature is selected
Select the data set D={ X after processing1, X2..., Xi..., Xd, each of them data are Xi={ x1, x2..., xn, y },
Training set and test set are constructed by 5: 1, data set is divided into k equal part according to time series timestep before training,
Timestep*k=d is trained as follows:
Step 1. is according to this 4 parameter settings of different model parameters such as the network number of plies, cell number, droupout rate, learning rate
Different model set M={ m1, m2..., mg}。
Step 2. takes out a model from model set, carries out the initialization of network parameter.
Timestep data input network in training set is trained by step 3., and the output of the last one time step is taken to make
For the output of hidden layer, and final classification results are exported after softmax function as the input of full articulamentum.Finally, will
Output result gives cross entropy loss function and calculates penalty values.
If step 4. penalty values are not converged, repeatedly step 3, and are joined according to the network of the continuous iteration of learning rate more new model
Number reaches convergence until penalty values.
Obtained convergent model is used for training set by step 5., the accuracy rate being recorded on training set, if model set M is also
There is remaining model not use, then goes to step 2.
Step 6. picks out the highest model of accuracy rate as optimal models from model set M.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810893739.8A CN109063775A (en) | 2018-08-03 | 2018-08-03 | Instruction SDC fragility prediction technique based on shot and long term memory network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810893739.8A CN109063775A (en) | 2018-08-03 | 2018-08-03 | Instruction SDC fragility prediction technique based on shot and long term memory network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109063775A true CN109063775A (en) | 2018-12-21 |
Family
ID=64678709
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810893739.8A Pending CN109063775A (en) | 2018-08-03 | 2018-08-03 | Instruction SDC fragility prediction technique based on shot and long term memory network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109063775A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111159011A (en) * | 2019-12-09 | 2020-05-15 | 南京航空航天大学 | Instruction vulnerability prediction method and system based on deep random forest |
| CN112765609A (en) * | 2020-12-31 | 2021-05-07 | 南京航空航天大学 | Multi-bit SDC fragile instruction identification method based on single-class support vector machine |
| CN112965854A (en) * | 2021-04-16 | 2021-06-15 | 吉林大学 | Method, system and equipment for improving reliability of convolutional neural network |
| CN113610154A (en) * | 2021-08-06 | 2021-11-05 | 吉林大学 | GPGPU program SDC error detection method and device |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103984632A (en) * | 2014-05-29 | 2014-08-13 | 东南大学 | SDC vulnerable instruction recognition method based on error propagation analysis |
-
2018
- 2018-08-03 CN CN201810893739.8A patent/CN109063775A/en active Pending
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103984632A (en) * | 2014-05-29 | 2014-08-13 | 东南大学 | SDC vulnerable instruction recognition method based on error propagation analysis |
Non-Patent Citations (2)
| Title |
|---|
| FANG WU等: "Vulnerability Detection with Deep Learning", 《2017 3RD IEEE INTERNATIONAL CONFERENCE ON COMPUTER AND COMMUNICATIONS》 * |
| 张倩雯等: "基于机器学习的指令SDC脆弱性分析方法", 《小型微型计算机系统》 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111159011A (en) * | 2019-12-09 | 2020-05-15 | 南京航空航天大学 | Instruction vulnerability prediction method and system based on deep random forest |
| CN111159011B (en) * | 2019-12-09 | 2022-05-20 | 南京航空航天大学 | Instruction vulnerability prediction method and system based on deep random forest |
| CN112765609A (en) * | 2020-12-31 | 2021-05-07 | 南京航空航天大学 | Multi-bit SDC fragile instruction identification method based on single-class support vector machine |
| CN112765609B (en) * | 2020-12-31 | 2022-06-07 | 南京航空航天大学 | Multi-bit SDC fragile instruction identification method based on single-class support vector machine |
| CN112965854A (en) * | 2021-04-16 | 2021-06-15 | 吉林大学 | Method, system and equipment for improving reliability of convolutional neural network |
| CN112965854B (en) * | 2021-04-16 | 2022-04-29 | 吉林大学 | Method, system and equipment for improving reliability of convolutional neural network |
| CN113610154A (en) * | 2021-08-06 | 2021-11-05 | 吉林大学 | GPGPU program SDC error detection method and device |
| CN113610154B (en) * | 2021-08-06 | 2023-12-29 | 吉林大学 | GPGPU program SDC error detection method and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Li et al. | Incremental learning imbalanced data streams with concept drift: The dynamic updated ensemble algorithm | |
| CN109063775A (en) | Instruction SDC fragility prediction technique based on shot and long term memory network | |
| CN112579477A (en) | Defect detection method, device and storage medium | |
| US10698697B2 (en) | Adaptive routing to avoid non-repairable memory and logic defects on automata processor | |
| US10521210B1 (en) | Programming language conversion | |
| CN114238100A (en) | Java vulnerability detection and positioning method based on GGNN and layered attention network | |
| CN101404033A (en) | Automatic generation method and system for noumenon hierarchical structure | |
| CN108804558A (en) | A kind of defect report automatic classification method based on semantic model | |
| CN105893876A (en) | Chip hardware Trojan horse detection method and system | |
| CN105095756A (en) | Method and device for detecting portable document format document | |
| US20170270424A1 (en) | Method of Estimating Program Speed-Up in Highly Parallel Architectures Using Static Analysis | |
| CN111400713B (en) | Malicious software population classification method based on operation code adjacency graph characteristics | |
| CN110955892B (en) | Hardware Trojan horse detection method based on machine learning and circuit behavior level characteristics | |
| Yu et al. | Deep learning-based hardware Trojan detection with block-based netlist information extraction | |
| CN116361788A (en) | Binary software vulnerability prediction method based on machine learning | |
| Chaudhuri et al. | Functional criticality analysis of structural faults in AI accelerators | |
| CN112035345A (en) | Mixed depth defect prediction method based on code segment analysis | |
| Hamerly et al. | Using machine learning to guide architecture simulation. | |
| CN101487876B (en) | Optimization method and apparatus for verification vectors | |
| CN111738290B (en) | Image detection method, model construction and training method, device, equipment and medium | |
| CN113626034A (en) | Defect prediction method based on combination of traditional features and semantic features | |
| CN115599698B (en) | Software defect prediction method and system based on class association rule | |
| Gunasekara et al. | On natural language processing applications for military dialect classification | |
| Hashemi et al. | Graph centrality algorithms for hardware trojan detection at gate-level netlists | |
| Brzezinski et al. | Structural XML classification in concept drifting data streams |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20181221 |