1. A knowledge graph construction and query recommendation system in the field of radio signal attack and defense is characterized by comprising the following components: the system comprises an ontology graph design module, a data collection module, a knowledge graph construction module and an inquiry and recommendation module;

the ontology graph design module constructs a radio signal attack and defense field knowledge graph;

the data collection module collects data detected by a radio signal attacking and defending method and an algorithm relation in the radio signal attacking and defending field knowledge graph;

the knowledge graph building module stores the data collected in the data collecting module;

the inquiry and recommendation module inquires the data stored in the knowledge graph construction module and recommends a corresponding attack and defense strategy or algorithm.

2. The system for knowledge graph construction and query recommendation in the field of radio signal defense and attack according to claim 1,

the radio signal attack and defense field knowledge graph comprises model nodes, signal nodes, attack event nodes and attribute nodes;

the model nodes comprise similar models and defense models, and the signal nodes comprise a signal set, normal signals, similar signals, countermeasure signals and purifying signals, wherein the signal set consists of target signal data of attack event nodes; the attribute nodes comprise model attributes, attack event attributes, normal signal attributes, countermeasure signal attributes and purification signal attributes;

the model nodes are respectively connected with the signal nodes and the attack event nodes, and the signal nodes and the attack event nodes are connected with the attribute nodes.

3. The system for knowledge graph construction and query recommendation in the field of radio signal defense and attack according to claim 1,

the algorithm relations in the radio signal attack and defense field knowledge graph comprise an attack algorithm, a defense algorithm, a detection algorithm, a purification algorithm, an identification type, corresponding relations of all attribute nodes, a signal similarity relation and a model similarity relation.

4. The knowledge graph construction and query recommendation system in the radio signal defense and attack field according to claim 3,

the model similarity relation comprises model structure similarity and model data similarity;

the structural similarity of the model is as follows: for two models M₁，M₂Structure of whichSimilarity B (M)₁，M₂) As model M₁And M₂The same structures are the same type of layer and adjacent substructures;

the similarity of the model data is as follows:

wherein N is the number of signals in the knowledge graph, f_M(S) represents the output probability vector of the model M for the last layer of the input signal S, cos () represents the cosine similarity between the two vectors;

the model similarity is:

λ₁B(S₁,S₂)+λ₂D(S₁,S₂) (2)

wherein λ is₁、λ₂The model structure similarity and the model data similarity are weighted respectively.

5. The knowledge graph construction and query recommendation system in the radio signal defense and attack field according to claim 3,

the signal similarity relation comprises signal structure similarity and signal model similarity;

the signal structure similarity A (M)₁，M₂) Comprises the following steps:

wherein S (k, i) represents the amplitude at i of the kth path of the signal;

the signal model similarity is as follows:

wherein m represents a modelNumber of (a), f_j(S) represents the probability vector output by the last layer of the model j when the signal S is input, and cos () represents the cosine distance between the two vectors;

the signal similarity is:

λ₃A(S₁,S₂)+λ₄M(S₁,S₂) (5)

wherein λ is₃、λ₄The signal structure similarity and the signal model similarity are weighted respectively.

6. The system for knowledge graph construction and query recommendation in the field of radio signal defense and attack according to claim 1,

the knowledge graph construction module stores all knowledge graph data in the radio signal attack and defense field by adopting a neo4j graph database, and stores the information of the model nodes and the signal nodes by adopting a MySql database.

7. The system for knowledge graph construction and query recommendation in the field of radio signal defense and attack according to claim 1,

the query and recommendation module queries the information of the signal nodes or the model nodes from the MySql database, positions the signal nodes or the model node entities in the Neo4j graph database according to the information to obtain algorithm relationships and nodes connected by the entities, renders the nodes on a front-end page in real time by using a D3.js technology, and gives attack and defense strategy recommendation.

8. The system for knowledge graph construction and query recommendation in the field of radio signal defense and attack according to claim 7,

the attack and defense strategy recommendation comprises attack and defense strategy recommendation of a model and attack and defense strategy recommendation of a radio signal.

9. The system for knowledge graph construction and query recommendation in the field of radio signal defense and attack according to claim 8,

the attack and defense strategy recommendation of the model comprises the following steps:

s1, giving a model, a target radio signal data set and a target attack attribute;

s2, retrieving and acquiring a similar model in a knowledge graph spectrum;

s3, acquiring all attack event nodes and defense model nodes connected with the given model and the similar model, and sequencing the attack event nodes and the defense model nodes respectively according to the target attack attribute effect;

s4, selecting the attack methods and information of the first n attack event nodes as attack strategy recommendations, and selecting the defense methods and information corresponding to the first n defense models as defense strategy recommendations.

10. The system for knowledge graph construction and query recommendation in the field of radio signal defense and attack according to claim 8,

the attack and defense strategy recommendation of the radio signal comprises the following steps:

t1, giving a radio signal, a classification model and a target attribute;

t2, retrieving and acquiring similar radio signals in the knowledge graph spectrum;

t3, acquiring all antagonistic signal nodes connected with the given radio signal and the similar radio signal, and sequencing according to target attribute effects;

t4, selecting an attack algorithm corresponding to the first n confrontation samples as the attack strategy recommendation of the radio signal;

t5, taking the countermeasure signal node as a purification sample;

t6, acquiring a corresponding optimal countermeasure signal set and a corresponding purified signal node set according to an optimal attack algorithm, and sequencing according to the target attribute effect and whether defense succeeds or not;

and T7, selecting the cleaning method and information of the first n cleaning samples as the cleaning algorithm recommendation of the radio signal.

Background

In daily life, radio signals are used for information transmission, which is still the mainstream mode of modern information transmission, after the radio signals are received, specific modulation types of the radio signals need to be identified, then corresponding demodulation can be carried out according to the modulation types, information carried by the signals is extracted, people usually need to spend a great deal of time and energy to collect data samples with known types, and classification algorithms are adopted to train the data samples to form clear interfaces, so that modulation type identification is completed, and most deep learning classification algorithms are fragile and easy to attack, so when the signal classification models are attacked by a specific attack algorithm, the radio signals are slightly disturbed, so that the modulation type identification models cannot correctly identify the modulation types of given signals, and how to select a proper defense algorithm to defend, has extremely important practical value.

In the prior art, a plurality of classification models, attack algorithms, defense algorithms and detection algorithms exist in the radio signal modulation type classification field, the current radio signal field does not correlate the information, and a knowledge graph of the field is not formed, when the attack and defense strategies of the radio signal modulation type attack and defense field are selected, only classification precision after attack is considered, other attributes are ignored, the models and specific radio signal data are not analyzed, a uniform representation method is not formed for the plurality of attack and defense detection algorithms, classification models and radio signal data to correlate the models and the radio signal data, implicit knowledge in the models is difficult to mine, and for one classification model or radio signal, a better attack and defense and detection strategy cannot be recommended, and usually, only various algorithm attempts can be performed.

Disclosure of Invention

The invention aims to solve the technical problems in the prior art and provides a knowledge graph construction and query recommendation system in the field of radio signal attack and defense.

The method can form the knowledge map in the field of radio signal attack and defense, is beneficial to mining the association in the field of radio signal modulation type identification and proposing corresponding attack and defense strategy recommendation, and improves the artificial intelligence safety in the field of radio signal modulation type classification.

In order to achieve the purpose, the invention provides the following scheme: the invention provides a knowledge graph construction and query recommendation system in the field of radio signal attack and defense, which comprises the following steps: the system comprises an ontology graph design module, a data collection module, a knowledge graph construction module and an inquiry and recommendation module;

the ontology graph design module is used for constructing a radio signal attack and defense field knowledge graph;

the data collection module is used for collecting data detected by a radio signal attacking and defending method and an algorithm relation in the radio signal attacking and defending field knowledge graph;

the knowledge graph building module is used for storing the data collected in the data collecting module;

the inquiry and recommendation module is used for inquiring the data stored in the knowledge graph construction module and recommending corresponding attack and defense strategies or algorithms.

Preferably, the radio signal attack and defense domain knowledge graph comprises model nodes, signal nodes, attack event nodes and attribute nodes;

the model nodes comprise similar models and defense models, the signal nodes comprise signal sets, normal signals, similar signals, countermeasure signals and purifying signals, the signal sets are composed of target signal data of attack event nodes, and the attribute nodes comprise model attributes, attack event attributes, normal signal attributes, countermeasure signal attributes and purifying signal attributes;

the model node is respectively connected with the signal node and the attack event node, and the signal node and the attack event node are both connected with the attribute node.

Preferably, the algorithm relationship in the radio signal attack and defense domain knowledge graph comprises an attack algorithm, a defense algorithm, a detection algorithm, a purification algorithm, an identification type, a corresponding relationship of each attribute node, a signal similarity relationship and a model similarity relationship.

Preferably, the model similarity relationship comprises a model structure similarity and a model data similarity;

the structural similarity of the model is as follows: for two models M₁，M₂Structural similarity of B (M)₁，M₂) As model M₁And M₂The same structures are the same type of layer and adjacent substructures;

the similarity of the model data is as follows:

wherein N is the number of signals in the knowledge graph, f_M(S) represents the output probability vector of the model M for the last layer of the input signal S, cos () represents the cosine similarity between the two vectors;

the model similarity is:

λ₁B(S₁,S₂)+λ₂D(S₁,S₂) (2)

wherein λ is₁、λ₂The model structure similarity and the model data similarity are weighted respectively.

Preferably, the signal similarity relationship comprises a signal structure similarity and a signal model similarity;

the signal structure similarity A (M1, M2) is as follows:

wherein S (k, i) represents the amplitude at i of the kth path of the signal;

the signal model similarity is as follows:

wherein m represents the number of models, f_j(S) represents the probability vector output by the last layer of the model j when the signal S is input, and cos () represents the cosine distance between the two vectors;

the signal similarity is:

λ₃A(S₁,S₂)+λ₄M(S₁,S₂) (5)

wherein λ is₃、λ₄The signal structure similarity and the signal model similarity are weighted respectively.

Preferably, the knowledge graph construction module stores all the knowledge graph data in the radio signal defense and attack field by using a neo4j graph database, and stores the information of the model nodes and the signal nodes by using a MySql database.

Preferably, the query and recommendation module queries the information of the signal node or the model node from the MySql database, and then locates the signal node or the model node entity in the Neo4j graph database according to the information, so as to obtain an algorithm relationship and a node connected by the entity, render the algorithm and the node on a front-end page in real time by using a d3.js technology, and give an attack and defense strategy recommendation.

Preferably, the attack and defense strategy recommendation comprises an attack and defense strategy recommendation of a model and an attack and defense strategy recommendation of a radio signal.

Preferably, the attack and defense strategy recommendation of the model comprises:

s1, giving a model, a target radio signal data set and a target attack attribute;

s2, retrieving and acquiring a similar model in a knowledge graph spectrum;

s3, acquiring all attack event nodes and defense model nodes connected with the given model and the similar model, and sequencing the attack event nodes and the defense model nodes respectively according to the target attack attribute effect;

s4, selecting the attack methods and information of the first n attack event nodes as attack strategy recommendations, and selecting the defense methods and information corresponding to the first n defense models as defense strategy recommendations.

Preferably, the policy recommendation for the attack and defense of the radio signal comprises:

t1, giving a radio signal, a classification model and a target attribute;

t2, retrieving and acquiring similar radio signals in the knowledge graph spectrum;

t3, acquiring all antagonistic signal nodes connected with the given radio signal and the similar radio signal, and sequencing according to target attribute effects;

t4, selecting an attack algorithm corresponding to the first n confrontation samples as the attack strategy recommendation of the radio signal;

t5, taking the countermeasure signal node as a purification sample;

t6, acquiring a corresponding optimal countermeasure signal set and a corresponding purified signal node set according to an optimal attack algorithm, and sequencing according to the target attribute effect and whether defense succeeds or not;

and T7, selecting the cleaning method and information of the first n cleaning samples as the cleaning algorithm recommendation of the radio signal.

The invention discloses the following technical effects:

1. algorithms such as attack defense detection and the like in the field of radio signal modulation type identification are associated to form a knowledge graph, so that knowledge storage and query are facilitated.

2. The formed knowledge graph comprises models, radio signals, algorithms and some models and signals derived from the models and the signals, and is beneficial to mining the association of the modulation type identification field of the radio signals.

3. By calculating the similarity between the models and the similarity between the signals, the attack and defense strategy recommendation based on the knowledge graph is facilitated.

4. By recommending the attack and defense strategy, the recommended attack and defense strategy is given to the unknown radio signals or the classification model, which is beneficial to improving the artificial intelligence security in the field of radio signal modulation type classification.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed to be used in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without inventive exercise.

FIG. 1 is a block diagram of the system of the present invention;

fig. 2 is an ontology diagram constructed by the knowledge graph in the field of radio signal attack and defense.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in further detail below.

Referring to fig. 1-2, the present embodiment provides a knowledge graph constructing and querying recommendation system in the field of radio signal attack and defense, including: the system comprises an ontology graph design module, a data collection module, a knowledge graph construction module and an inquiry and recommendation module.

The ontology graph design module is used for constructing a radio signal attack and defense field knowledge graph; the data collection module performs experiments according to different classification signals, radio models, attack algorithms and detection algorithms according to the knowledge map ontology diagram completed in the ontology diagram design module to obtain data; the knowledge graph building module is used for storing the data collected in the data collecting module; the inquiry and recommendation module is used for inquiring the data stored in the knowledge graph construction module and recommending corresponding attack and defense strategies or algorithms.

The radio signal attack and defense field knowledge graph comprises model nodes, signal nodes, attack event nodes and attribute nodes; the model nodes are represented by circles in the knowledge-graph, the signal nodes are represented by squares in the knowledge-graph, the attack event nodes are represented by ellipses in the knowledge-graph, and the attribute nodes are represented by hexagons in the knowledge-graph; the model nodes comprise similar models and defense models, the signal nodes comprise signal sets, normal signals, similar signals, countermeasure signals and purifying signals, the signal sets are composed of target signal data of attack event nodes, and the attribute nodes comprise model attributes, attack event attributes, normal signal attributes, countermeasure signal attributes and purifying signal attributes; the model node is respectively connected with the signal node and the attack event node, and the signal node and the attack event node are both connected with the attribute node.

The model attributes include, but are not limited to: training time of the model, recognition accuracy of the model on modulation types of different data sets and model structure; the attack event attributes include, but are not limited to: attack parameters, attack time consumption, average time consumption, attack success rate, average disturbance quantity (l0, l1), average amplitude change rate, average phase signal difference and average disturbance signal ratio; the signal attributes include, but are not limited to: average amplitude, average phase, signal-to-noise ratio, modulation type, large phase, maximum amplitude, minimum phase, amplitude of each dimension of the radio signal and output probability of the radio signal under each classification model; the countermeasure signal attribute includes, but is not limited to: attack time consumption, disturbance amount (l0, l1), amplitude change rate, phase signal difference and disturbance signal ratio; the decontamination signal properties include, but are not limited to: time consuming to clean, signal variance after clean.

The algorithm relationship in the radio signal defense and attack field knowledge graph comprises an attack algorithm (specific attack algorithm), a defense algorithm (specific defense algorithm), a detection algorithm (specific detection algorithm), a purification algorithm (specific radio signal defense algorithm), an identification type (specific modulation type obtained by identifying the modulation type of a signal by a model), the corresponding relationship of each attribute node, the similar relationship between signals and the similar relationship between models.

The model similarity relation comprises model structure similarity and model data similarity.

For most deep learning models, the similarity is difficult to define, the invention provides a heuristic similarity evaluation method, and the similarity between models is determined according to the structure of the models and the prediction condition of data, because most attack algorithm principles are attacked according to the gradient of the loss of the models, and the models with similar model structures have approximate gradients in the reverse conduction process of the loss, the similarity of the model structures is as follows: for two models M₁，M₂Structural similarity of B (M)₁，M₂) As model M₁And M₂The same structures are the same type of layer and adjacent substructures; i.e., if M₁，M₂There are adjacent convolutional layers and fully-connected layers, then there are 3 identical structure numbers because it has both convolutional layers, fully-connected layers, and adjacent convolutional layers and fully-connected layers.

The similarity of the model data is as follows:

wherein N is the number of signals in the knowledge graph, f_M(S) represents the output probability vector of the model M for the last layer of the input signal S, cos () represents the cosine similarity between the two vectors; the similarity of the model data measures two classification models in the same batchDegree of similarity of prediction under input of number data.

The model similarity is:

λ₁B(S₁,S₂)+λ₂D(S₁,S₂) (2)

wherein λ is₁、λ₂Respectively weighting the model structure similarity and the model data similarity; and setting a threshold, and if the structural similarity between the two models is greater than the threshold, determining that a similarity relationship exists between the two models.

The signal similarity relationship includes signal structure similarity (similarity of signal amplitude structure) and signal model similarity.

For two k-way m-dimensional signals S₁，S₂The signal structure similarity A (M1, M2) is:

wherein S (k, i) represents the amplitude at i of the k path of the signal, and the smaller the value, the closer the two signal amplitude structures are.

The signal model similarity M (S)₁，S₂) Comprises the following steps:

wherein m represents the number of models, f_j(S) represents the probability vector output by the last layer of the model j when the signal S is input, and cos () represents the cosine distance between the two vectors; the similarity of the signal models can measure the similarity of the prediction probabilities of two signals facing the same model, and the smaller the value, the closer the prediction results of the two signals under the same model are.

The signal similarity is:

λ₃A(S₁,S₂)+λ₄M(S₁,S₂) (5)

wherein the content of the first and second substances,λ₃、λ₄respectively weighting the signal structure similarity and the signal model similarity; and setting a threshold, wherein if the similarity between the two signals is smaller than the threshold, the two signals have a similarity relation.

The knowledge graph building module stores knowledge graph data of all the radio signal defense and attack fields by adopting a neo4j graph database, stores a classification model and detailed information of radio signals by combining a MySql database, is convenient and rapid to query, and when the knowledge graph is finally queried, the classification model or the signals in the graph are rapidly retrieved according to the MySql, and then the information of peripheral connection entities and relationships is queried and obtained according to the graph data stored in the neo4 j.

The query and recommendation module queries information of the signal nodes or the model nodes from the MySql database according to a signal or a model expected to be queried, positions the signal nodes or the model node entities in the Neo4j graph database according to the information so as to obtain algorithm relationships and nodes connected by the entities, renders the nodes on a front-end page in real time by using a D3.js technology, and gives corresponding attack and defense strategy recommendations.

The attack and defense strategy recommendation comprises attack and defense strategy recommendation of a model and attack and defense strategy recommendation of a radio signal.

The attack and defense strategy recommendation of the model comprises the following steps:

s1, a model, a target radio signal data set and target attack attributes (expected attack effect, such as tendency to high attack success rate or low attack time consumption, low disturbance amount and the like) are given; and searching the model entity which is most similar to the attribute of the model in the knowledge graph according to the attribute of the model. The step is realized by a graph data query sentence Cypher language.

And S2, retrieving and acquiring a similar model in the knowledge graph.

S3, acquiring all attack event nodes connected with the given model and the similar model, setting the attack event nodes as an attack event node set T, and acquiring all defense model nodes as a defense model node set D; and sequencing the attack nodes in the node set according to the target attributes, if the target attributes are high attack success rates, sequencing the attack nodes in the node set in a descending order according to the attack success rates, otherwise, sequencing the defense model nodes in the defense model node set according to the target attributes, if the target attributes are high attack success rates, sequencing the attack nodes in an ascending order according to the attack success rates, and otherwise, sequencing the attack nodes in the defense model node set in the ascending order.

S4, selecting the attack methods and information of the first n ordered attack event nodes as attack strategy recommendation, and selecting the defense methods and information corresponding to the first n ordered defense models as defense strategy recommendation.

The attack and defense strategy recommendation of the radio signal comprises the following steps:

t1, giving a radio signal, a classification model and a target attribute; the properties of the radio signals are calculated from the given radio signals, and the radio signal entities most similar to their properties are retrieved from the knowledge graph according to the properties of the radio signals. The step is realized by a graph data query sentence Cypher language.

And T2, retrieving and acquiring similar radio signals in the knowledge graph spectrum.

And T3, all the countermeasure signal nodes connected with the given radio signal and the similar radio signal are acquired and set as a countermeasure signal node set P, and corresponding sorting is carried out according to target attribute effects of the countermeasure signal nodes (the effect is better and the effect is higher). And if the target attribute is low disturbance quantity, performing descending sorting according to the disturbance quantity attribute of the countermeasure signal node, and vice versa.

And T4, selecting the attack algorithm corresponding to the first n ordered confrontation samples as the attack strategy recommendation of the radio signal.

T5, obtaining clean samples of all antagonistic signal sample connections of said radio signal connection.

And T6, obtaining corresponding countermeasure signal nodes according to an optimal attack algorithm, recording the corresponding countermeasure signal nodes as an optimal countermeasure signal set B, obtaining purification signal nodes connected with all countermeasure signal receiving units in the set B, recording the purification signal nodes as a purification signal node set C, and correspondingly sorting all the purification nodes in the purification node set C according to the effect of the target attribute and whether defense is successful (whether the identification type of the purification nodes is equal to the type of the normal signals or not) (for example, when the target attribute is low in purification time consumption, sorting is performed in a descending order according to purification time consumption and defense success).

And T7, selecting the cleaning method and information of the first n cleaning samples as the cleaning algorithm recommendation of the radio signal.

The above target attributes (or target attack attributes) may be combined to achieve a customized attack effect or defense (decontamination) effect.

The system of the invention associates the algorithms such as attack defense detection and the like in the field of radio signal modulation type identification to form a knowledge map, which is convenient for knowledge storage and query, the formed knowledge map comprises models, radio signals, algorithms and derived models and signals, which is beneficial to mining the association of the field of radio signal modulation type identification, and is beneficial to carrying out attack defense strategy recommendation based on the knowledge map by calculating the similarity among the models and the similarity among the signals, and the system gives a recommended attack defense strategy to unknown radio signals or classification models by the attack defense strategy recommendation, thereby being beneficial to improving the artificial intelligence security in the field of radio signal modulation type classification.

The above-described embodiments are merely illustrative of the preferred embodiments of the present invention, and do not limit the scope of the present invention, and various modifications and improvements of the technical solutions of the present invention can be made by those skilled in the art without departing from the spirit of the present invention, and the technical solutions of the present invention are within the scope of the present invention defined by the claims.