Signaling network vulnerability attack simulation and prevention system based on AI

Abstract

The invention discloses an AI-based signaling network vulnerability attack simulation and prevention system, which comprises a signaling sensing layer, which is deployed in a core network element to carry out hardware-level decoding of an SS7 / Diameter / SIP / 5G-NR protocol, dynamically extracts protocol field-level metadata through a YARA rule base and generates a standardized signaling log; the threat modeling layer is used for receiving a signaling log output by the signaling sensing layer, generating mixed traffic after injecting a simulation attack, outputting the mixed traffic to the anomaly detection layer, constructing a multi-dimensional feature encoder by using a parameterized quantum gate based on a quantum generative adversarial network and a GFlowNet stream generation engine, generating a compliance attack vector in combination with the constraint of a 3GPP protocol state machine, and outputting the compliance attack vector to the anomaly detection layer; meanwhile, attack trajectory diversity sampling is completed through a Diameter protocol AVP nested state tree and a two-factor award function; the anomaly detection layer is used for receiving the mixed flow of the threat modeling layer and outputting a threat evaluation result to the response processing layer; and the response processing layer is used for receiving the threat assessment result of the anomaly detection layer and issuing a defense instruction to the signaling sensing layer.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of signaling network security, in particular to an AI-based signaling network vulnerability attack simulation and prevention system. BACKGROUND

[0002] Under the industrial background of 5G network evolution to service architecture and acceleration of 6G technology research, as the central nervous system of the communication network, the signaling network, on the one hand, the signaling protocols exposed by the 5G core network through N1 / N2 / N4 interfaces support network slicing, edge computing and other new features, and directly expose the originally closed signaling interaction to the Internet side, so that attackers can implement SBA service hijacking by forging AMF / SMF signaling or initiate hidden S6a / S9 interface penetration by using the Diameter protocol AVP nesting feature; on the other hand, the encrypted signaling and dynamic protocol extension mechanism cause the traditional deep packet inspection technology to fail, and the protection system based on feature matching is difficult to identify 0day exploit. To this end, the AI-based signaling network vulnerability attack simulation and prevention system is an intelligent defense platform specially designed for the security of the core signaling system of the communication network. By integrating deep learning, behavior analysis, fuzzy testing and other AI technologies, the system actively simulates the exploit attack scenarios for SS7, Diameter, SIP and other signaling protocols in the real world, and relies on the built-in AI module to continuously learn the tactics, techniques and procedures of attackers, dynamically updates the attack mode library to reproduce the APT organization-level attack chain, so as to perform full-life-cycle vulnerability mining, penetration testing and threat verification on the target signaling network in the virtualized environment. The generative adversarial network (GAN) plays a core role in key links such as attack scenario generation, vulnerability mining, defense strategy optimization and dynamic threat modeling through its unique adversarial training mechanism.

[0003] In the prior art, the generator and discriminator of the GAN adversarial training mechanism is prone to fall into the Nash equilibrium dilemma in the signaling network attack scenario generation. When the dimension of the attack vector exceeds 1024 (such as the NAS layer message of 5G signaling), the traditional stochastic gradient descent algorithm is difficult to converge, which may cause the generator to only reproduce historical attack modes such as SS7-MAP exploit, and cannot cover new attack vectors of the S6a / S9 interface of the Diameter protocol. Therefore, an AI-based signaling network vulnerability attack simulation and prevention system is proposed. SUMMARY

[0004] The purpose of the application is to solve the problems in the prior art and propose an AI-based signaling network vulnerability attack simulation and prevention system.

[0005] To achieve the above purpose, the application adopts the following technical solutions: An AI-based signaling network vulnerability attack simulation and prevention system comprises: Signaling-aware layer: Adopting distributed probe matrix and protocol reverse analysis engine, deployed in core network elements for hardware-level decoding of SS7 / Diameter / SIP / 5G-NR protocol, dynamically extracting protocol field-level metadata through YARA rule library and generating standardized signaling logs; Threat modeling layer: Receiving signaling logs output by the signaling-aware layer, generating mixed traffic after injecting simulated attacks, and outputting to the anomaly detection layer, relying on quantum generative adversarial networks and GFlowNet flow generation engines, using parameterized quantum gates to build multi-dimensional feature encoders, combining 3GPP protocol state machine constraints to generate compliant attack vectors, and simultaneously completing attack trajectory diversity sampling through Diameter protocol AVP nested state tree and double-factor reward function; Anomaly detection layer: Receiving mixed traffic from the threat modeling layer, outputting threat assessment results to the response handling layer, integrating topology invariant feature extraction, mask graph attention network, and dynamic graph reconstruction graph neural network, building anti-disturbance convolution kernels based on graph Laplacian matrix eigen decomposition, the mask graph attention network adopts a three-layer defense mechanism, anomaly score MLP→ binary mask generation→ GAT aggregation purification, the dynamic graph reconstruction graph neural network uses an entropy-driven reconstruction mechanism to compress signaling storm detection latency, that is, real-time calculation of node information entropy, exceeding the threshold triggers local graph reconstruction; Response handling layer: Receiving threat assessment results from the anomaly detection layer, issuing defense instructions to the signaling-aware layer, through the adaptive defense strategy engine and vulnerability repair suggestion system, dynamically adjusting ACL rules and traffic speed limit thresholds based on reinforcement learning, and combining CVE vulnerability library to generate hot patch deployment schemes, and outputting STIX 2.1 format compliance audit reports to interface with SOAR platform.

[0006] The above technical solution further comprises: Further, the distributed probe matrix is an asymmetric duplex probe matrix for non-invasive protocol capture of idle core network elements, which uses FPGA hardware acceleration technology to perform physical layer traffic mirroring and protocol stack accelerated analysis on SS7 / Diameter / SIP / 5G-NR protocol, combined with protocol fingerprinting algorithm based on N-gram statistical model and dynamic state machine constructed by hidden Markov model, to complete reverse analysis of unknown protocol variants and hidden channels; Using the templating matching mechanism of the YARA rule engine, the original traffic is pattern matched through pre-defined protocol field encoding rules, and the binary data is mapped to structured metadata through field extraction formulas; The multi-protocol fields are mapped to a unified data model through structured metadata normalization processing, and the timestamps, protocol types, field value sets and hash check values are encapsulated using a structured log generation formula to form signaling logs conforming to the CEF standard; The YARA rule library is updated through a double buffering mechanism, and a rule conflict detection system is constructed using the Jaro-Winkler similarity algorithm. New attack features are injected into the analysis process in real time without causing rule redundancy.

[0007] Further, the threat modeling layer receives the signaling logs output by the signaling perception layer, and the specific steps for generating mixed traffic after injecting simulated attacks are as follows: Quantum feature encoder construction: The quantum feature encoder uses a quantum circuit constructed by a parameterized quantum gate to map multi-dimensional signaling features to a quantum superposition state; 3GPP protocol state machine constraint engine: The protocol state machine is modeled using Petri nets, and the compliance constraint formula is used to force the attack vector to follow the legal transition sequence of the SS7 / Diameter / SIP protocol, so that the generated attack operation conforms to the 3GPP specification; GFlowNet attack trajectory generation: The GFlowNet flow generation engine is introduced, and the state space modeling of the attack graph and the forward-backward flow matching algorithm are used to sample the diversity of attack trajectories while maintaining protocol compliance; Double-factor reward function: The weighted sum of attack effect score and protocol compliance score is combined, and the reinforcement learning mechanism is used to dynamically balance attack effectiveness and specification compliance; Mixed traffic output: The simulated attack traffic and the original signaling traffic are mixed in proportion using the traffic synthesis formula, and the IEEE 1588v2 protocol is used to ensure the time alignment of attack flow and normal flow, generating test traffic that conforms to the timing characteristics of real networks.

[0008] Further, the 3GPP protocol state machine constraint engine uses a cooperative mechanism of formal verification and dynamic constraints to maintain the innovation of attack vectors while making them conform to the 3GPP specification, including the following steps: Petri net formal modeling phase: First, build a five-tuple model of the protocol state machine where, represents the set of protocol state libraries, represents the set of protocol event transitions, defines the directed arc connection relationship, is the arc weight function, is the initial marking distribution, and the five-tuple model converts the 3GPP specification into a computable topology; State transition rule definition: Establish a transition trigger condition formula: θ is a threshold parameter related to the protocol version, if and only if the transition Former warehouse When the weighted sum of the tags exceeds a threshold, a state transition is triggered. Construction of compliance constraint formulas: Design a multi-dimensional constraint function family, including protocol version compatibility constraints. λ is the version matching coefficient; state reachability constraint: , For state Minimum number of live markers; message timing constraints: , For time-series sensitivity parameters, The protocol specifies a time delay; therefore, the comprehensive constraint formula is: ; Attack vector verification process: This involves verifying the generated attack sequences. Perform state projection, constraint checks, and path integrity verification. The state projection is represented as... make The constraint check is expressed as The path integrity check is represented as Does it constitute The legal path in; Dynamic adjustment mechanism: The model parameters are updated through a real-time feedback loop, and the weights are adaptively represented as follows. , The learning rate, threshold calibration is expressed as , is the damping coefficient.

[0009] Furthermore, the GFlowNet attack trajectory generation achieves a balance between protocol compliance and attack diversity by constructing a flow probability model, including the following steps; Attack graph state space expansion modeling: by constructing a weighted directed hypergraph The protocol state library With attack action The nodes are combined into composite state nodes, and a super-edge transition relationship and a protocol compliance weight function are defined, wherein the weight function integrates 3GPP constraints, state transition costs and path preference coefficients. A forward-reverse flow matching algorithm is adopted: By designing the probability flow conservation conditions for forward and reverse flows and the transition probability formula based on compliance weights and exploration coefficients, a bidirectional flow allocation mechanism is established. The forward flow F(s) represents the cumulative probability flow from the initial state to s, and the reverse flow R(s) represents the cumulative probability flow from s to the final state. The two are dynamically balanced through the flow conservation equation, i.e., for non-boundary states... , must meet (s)=R(s)+ Where Succ(s) is the set of successor states of s. (s→s′)=F(s)⋅P To allocate the forward flow, and with the design of adjustable transition probabilities, the randomness and goal orientation of path exploration are balanced while ensuring protocol compliance. Define a two-component training objective function: it includes a flow matching loss term based on state distribution sampling and an edge matching loss term based on KL divergence. The flow matching loss term based on state distribution sampling ensures the consistency of the flow probability model by minimizing the difference between the forward flow and the reverse flow plus the transition flow. The flow matching loss term is expressed as follows: FM =E s ~ dG ( s ) ,in, For state distribution, For the forward / backward network parameters, the edge matching loss term based on KL divergence improves the semantic rationality of the generated trajectory by approximating the distribution of real attack patterns. The edge matching loss term is expressed as... EM = DKL ( πθ (⋅∣ s )∥ πdata (⋅∣ s Generate trajectory distribution θ Distribution of attack patterns approaching reality data ; Diverse sampling strategy: Introduce a trajectory balancing sampler and construct a sampling weight formula by combining traffic flow ratio and diverse temperature parameters to achieve probability-driven diversified path selection within the compliance constraint framework.

[0010] Furthermore, the anomaly detection layer receives the mixed traffic from the threat modeling layer and outputs the threat assessment result, including the following steps: Topology-invariant feature extraction: Utilizing graph convolutional layers or graph attention layers in graph neural networks combined with topology metrics, deep structural features of the network topology are extracted through multi-scale graph pooling operations, constructing a time-series graph snapshot. Where V represents the set of network element nodes, and E represents the set of signaling links. It is represented as a node feature matrix, and the eigenvectors corresponding to the first k smallest eigenvalues ​​are taken as topologically invariant embeddings; Three-layer defense mask graph attention network: The mask graph attention network introduces a learnable attention mask mechanism, which weights and aggregates node features through a multi-head self-attention module. The mask generator dynamically calculates the attention weights of nodes / edges based on the deviation of current traffic features from historical behavior baselines, and combines a gating mechanism to filter noisy connections and strengthen the attention allocation of key paths. Dynamic graph reconstruction graph neural network: Real-time reconstruction of signaling interaction graph structure using reinforcement learning framework or variational autoencoder, optimizing graph connectivity through edge prediction task and community discovery algorithm, combining time decay factor to weight historical graph snapshots, triggering local graph structure pruning and reconstruction when detecting abnormal subgraph, and ensuring semantic consistency between reconstructed graph and normal topology pattern through contrastive learning module.

[0011] Further, the mask graph attention network performs dynamic attention allocation and attack path suppression, including the following steps: Dynamic attention mask generation: input the current time signaling node feature matrix and the historical behavior baseline into the mask generator, calculate the feature deviation through dynamic time warping , input the deviation into the gating recurrent unit, combine the historical mask state to generate the dynamic mask . Multi-head self-attention aggregation: aggregate node features through multi-head attention mechanism, while using masks to strengthen key paths. For each attention head h, calculate the mask-adjusted attention coefficient, denoted as where is the value projection matrix, is the learnable attention vector, denotes the concatenation operation, the attention coefficient is weighted and normalized by Softmax, and the neighbor features are aggregated, denoted as , concatenate and linearly transform the outputs of H attention heads, and then aggregate the node features; Gating noise filtering and residual update: suppress sparse attack patterns through gating mechanism and retain normal topology information.

[0012] Further, the dynamic graph reconstruction graph neural network performs adaptive defense reconstruction of signaling interaction graph, including the following steps: Graph structure encoding based on temporal fusion: weight and aggregate the graph snapshots of the past T time points through an exponential decay function, input the aggregated adjacency matrix and node features H into the graph convolutional network to generate a dynamic context-aware graph representation; Reinforcement learning driven graph structure reconstruction: input the current graph structure The attack probability output by the anomaly detection layer is encoded as a state vector, the edge addition and deletion operation is defined as a discrete action space A, a multi-objective reward function is designed, including connectivity preservation, attack surface reduction and topology stability, the policy network is updated using the proximal policy optimization algorithm, the expected return is estimated by Monte Carlo sampling, and the edge connection is dynamically adjusted to block the attack path. Abnormal subgraph pruning and contrastive learning: traverse the subgraph by breadth-first search, if the loop length is detected (like ), mark it as an abnormal subgraph , calculate the reconstruction confidence of the edges in the abnormal subgraph, delete the edges with confidence lower than the threshold τ, and the reconstruction confidence calculation formula is σ( , the original graph G and the reconstructed graph are input into the contrastive learning framework, the feature similarity is calculated by the projection head . Community-aware connectivity optimization: use the Louvain algorithm to detect communities in the reconstructed graph, get the community label, and introduce a community penalty term for the cross-community edges , define the modularity loss function, so that the edge density within the community is higher than the random expectation.

[0013] Further, the response treatment layer receives the threat assessment result of the anomaly detection layer, and issues defense instructions to the signaling awareness layer, including the following steps Adaptive defense strategy engine: based on the real-time collected signaling network traffic data, system logs and historical attack cases, a hybrid model of spatio-temporal convolutional neural network and long short-term memory network (LSTM) constructed by deep learning framework is used to extract features and identify threat patterns from multi-source heterogeneous data, and a reinforcement learning algorithm is used to dynamically adjust the defense threshold and generate differentiated response strategies. Dynamic ACL rule optimization: based on the attack trajectory, a five-tuple mask rule is generated, and by combining the threat level and the business criticality coefficient, a logarithmic priority calculation formula is used to complete the dual goals of precise attack traffic blocking and business continuity protection, and the priority calculation formula is is the business criticality coefficient of the network element i. Entropy-driven traffic speed limit: the protocol distribution entropy is used to quantify the degree of traffic anomaly, and the calculation formula is , is the traffic of protocol p, when the entropy value is lower than the threshold θ, the dynamic threshold adjustment formula is used to implement speed limit. Vulnerability repair suggestion system: match the CVE vulnerability library and system components by Jaccard similarity algorithm, and the calculation formula is , combined with the risk assessment model based on exponential attenuation, the template engine is used to generate a hot patch scheme containing sysctl parameter adjustment and iptables rule; STIX 2.1 compliance report generation: mapping attack events, defense actions, and vulnerability information into standardized objects, interfacing with SOAR platforms such as Demisto through REST API, and realizing threat intelligence output verified by NIST 800-53 standards.

[0014] The present application has the following advantages: In the present application, in view of the Nash equilibrium dilemma of GAN in high-dimensional attack vector generation, the threat modeling layer adopts quantum generative adversarial network (QGAN) combined with parameterized quantum gates to construct a 2048-dimensional feature encoder, breaks through the dimension limitation of traditional GAN through the characteristics of quantum superposition state, and realizes probability flow optimization with GFlowNet flow generation engine, so that the generator can maintain convergence stability in high-dimensional scenarios such as 5G NAS layer messages; At the same time, the 3GPP protocol state machine constraint and the double-factor reward function (attack compliance + novelty) are introduced to ensure that the generated attack vector not only meets the protocol specification, but also covers new attack modes. BRIEF DESCRIPTION OF DRAWINGS

[0015] FIG. 1 A system block diagram of an AI-based signaling network vulnerability attack simulation prevention system is provided. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0017] Please refer to FIG. 1 As shown in the figure, the present application is an AI-based signaling network vulnerability attack simulation prevention system, which comprises: Signaling perception layer: adopting distributed probe matrix and protocol reverse analysis engine, deployed in core network elements for hardware-level decoding of SS7 / Diameter / SIP / 5G-NR protocol, dynamically extracting protocol field-level metadata through YARA rule library and generating standardized signaling logs; The threat modeling layer receives the signaling log output by the signaling awareness layer, generates mixed traffic after injecting simulated attacks, and outputs to the anomaly detection layer. The threat modeling layer relies on a quantum generative adversarial network (QGAN) and a GFlowNet flow generation engine, uses a parameterized quantum gate to construct a multi-dimensional (2048-dimensional) feature encoder to break through the dimension disaster of traditional GANs, combines 3GPP protocol state machine constraints to generate compliant attack vectors, and simultaneously completes attack trajectory diversity sampling through a Diameter protocol AVP nested state tree and a double-factor reward function; The anomaly detection layer receives the mixed traffic from the threat modeling layer, outputs threat assessment results to the response handling layer, integrates topology invariant feature extraction (TIFE), a masked graph attention network (GAT), and a dynamic graph reconstruction graph neural network (DGR-GNN), constructs an anti-disturbance convolution kernel based on graph Laplacian matrix feature decomposition, the masked graph attention network (GAT) adopts a three-layer defense mechanism, abnormal score MLP→binary mask generation→GAT aggregation purification, and the dynamic graph reconstruction graph neural network (DGR-GNN) uses an entropy value driven reconstruction mechanism to compress signaling storm detection delay, that is, real-time calculation of node information entropy, and triggers local graph reconstruction when the threshold is exceeded; The response handling layer receives the threat assessment results from the anomaly detection layer, issues defense instructions to the signaling awareness layer, and through an adaptive defense strategy engine and a vulnerability repair suggestion system, dynamically adjusts ACL rules and traffic speed limit thresholds based on reinforcement learning, simultaneously generates a hot patch deployment scheme in combination with a CVE vulnerability library, and outputs an STIX 2.1 format compliant audit report to interface with a SOAR platform.

[0018] In one embodiment, the distributed probe matrix is an asymmetric duplex probe matrix for non-intrusive protocol capture of idle core network elements. The asymmetric duplex probe matrix uses FPGA hardware acceleration technology to perform physical layer traffic mirroring and protocol stack accelerated analysis on SS7 / Diameter / SIP / 5G-NR protocols, combines a protocol fingerprint recognition algorithm based on an N-gram statistical model and a dynamic state machine constructed by a hidden Markov model (HMM) to complete reverse analysis of unknown protocol variants and covert channels; Using the templated matching mechanism of the YARA rule engine, the original traffic is pattern matched through pre-defined protocol field encoding rules (such as the 6C 04 0B 81 feature sequence of SS7 TCAP messages), and the binary data is mapped to structured metadata (such as IMSI, Session-Id, etc. standardized fields) through field extraction formulas; Through structured metadata normalization processing, the multi-protocol fields are mapped to a unified data model, and structured log generation formulas are used to encapsulate timestamps, protocol types, field value sets, and hash check values to form CEF-compliant signaling logs; The hot loading update of the YARA rule library is completed through the double buffering mechanism, the rule conflict detection system is constructed by combining the Jaro-Winkler similarity algorithm, the new attack characteristics are injected into the analysis process in real time, and rule redundancy is not caused.

[0019] In one embodiment, the threat modeling layer receives signaling logs output by the signaling awareness layer, and specific steps of generating mixed traffic after injecting simulated attacks: Quantum feature encoder (QFE) construction: a quantum circuit constructed by a quantum feature encoder (QFE) using a parameterized quantum gate (PQC) maps multi-dimensional (2048-dimensional) signaling features to a quantum superposition state; 3GPP protocol state machine constraint engine: the protocol state machine is modeled in the form of Petri net, and the attack vector is forced to follow the legal transition sequence of the SS7 / Diameter / SIP protocol through a compliance constraint formula, so that the generated attack operation conforms to the 3GPP specification; GFlowNet attack trajectory generation: a GFlowNet flow generation engine is introduced, and through the extension of the state space modeling of the attack graph and the forward-backward flow matching algorithm, the diversity of the attack trajectory is sampled while maintaining the protocol compliance, effectively avoiding the mode collapse problem of traditional GAN; Double-factor reward function: combine the weighted sum of attack effect score and protocol compliance score, and dynamically balance attack effectiveness and specification compliance through reinforcement learning mechanism; Mixed traffic output: the simulated attack traffic and the original signaling traffic are mixed in proportion by using the traffic synthesis formula, and the time alignment of the attack flow and the normal flow is guaranteed by using the IEEE 1588v2 protocol, to generate test traffic that conforms to the timing characteristics of the real network.

[0020] In one embodiment, the 3GPP protocol state machine constraint engine uses a cooperative mechanism of formal verification and dynamic constraints to maintain the innovation of the attack vector while making the attack vector comply with the 3GPP specification, including the following steps: Petri net formal modeling phase: first, build a five-tuple model of the protocol state machine , wherein represents a set of protocol state libraries (such as IDLE, CALL_INIT, ACTIVE, etc. States of the SS7 protocol), represents a set of protocol event transitions (such as IAM, ACM, REL, etc. Signaling messages), defines the connection relationship of the directed arc, is the arc weight function, is the initial marking distribution, and the 3GPP specification is converted into a computable topology structure through the five-tuple model; State transition rule definition: Establish transition trigger condition formula , θ is a threshold parameter related to protocol version, and the state transition is triggered only when the weighted mark of the pre-set library of the transition exceeds the threshold value, for example, in the Diameter protocol, the sending of CEA message needs to meet the weight value of the pre-set library (connection establishment stage) and the current mark ; ; Compliance constraint formula construction: Design multi-dimensional constraint function cluster, including protocol version compatibility constraint: , λ is the version matching coefficient; state reachability constraint: , is the minimum survival mark number of state ; message timing constraint: , is the timing sensitivity parameter, is the protocol specified delay; and then the comprehensive constraint formula is ; Attack vector verification process: Perform state projection, constraint check and path integrity check on the generated attack sequence , the state projection is represented as , , the constraint check is represented as , and the path integrity check is represented as whether it constitutes a legal path in ; Dynamic adjustment mechanism: Update model parameters through real-time feedback loop, weight adaptive expression is , is the learning rate, and threshold calibration is represented as , is the damping coefficient.

[0021] In one embodiment, the GFlowNet attack trajectory generation realizes the unification of protocol compliance and attack diversity by constructing a flow probability model, including the following steps: Attack graph state space expansion modeling: Attack graph state space expansion modeling is realized by constructing a weighted directed hypergraph , combining the protocol state machine library and attack action into a composite state node, and defining hyperedge transition relationship and protocol compliance weight function, wherein the weight function integrates 3GPP constraint conditions, state transition cost and path preference coefficient; Adopting forward-backward flow matching algorithm: by designing the probability flow conservation conditions of forward flow and reverse flow and the transition probability formula based on compliance weight and exploration coefficient, a bidirectional flow allocation mechanism is established, wherein the forward flow F(s) represents the cumulative probability flow from the initial state to s, the reverse flow R(s) represents the cumulative probability flow from s to the termination state, and the two are dynamically balanced through the flow conservation equation, i.e. , needs to meet (s)=R(s)+ , wherein Succ(s) is the successor state set of s, (s→s′)=F(s)⋅P is the forward flow allocation, and the transition probability is designed with adjustable parameters, balancing the randomness and goal orientation of path exploration under the premise of ensuring protocol compliance; Defining a double-component training objective function: including a flow matching loss term based on state distribution sampling and an edge matching loss term based on KL divergence, the flow matching loss term based on state distribution sampling ensures the consistency of the flow probability model by minimizing the difference between the forward flow and the reverse flow plus the transition flow, and the flow matching loss term is represented as FM =E s ∼ dG ( s ) , wherein is the state distribution, is the forward / backward network parameter, and the edge matching loss term based on KL divergence improves the semantic rationality of the generated trajectory by approximating the real attack mode distribution, and the edge matching loss term is represented as EM = DKL ( πθ (⋅∣ s )∥ πdata (⋅∣ s )), the generated trajectory distribution θ approximates the real attack mode distribution data ; Diversity sampling strategy: introducing a trajectory balance sampler, combining flow proportion and diversity temperature parameters to construct a sampling weight formula, and realizing probability-driven diversified path selection within the compliance constraint framework.

[0022] Example description: SS7 protocol location service attack trajectory generation.

[0023] When an attacker tries to obtain user location information by forging SRI messages.

[0024] Initial state: s0=IDLE GFlowNet generates trajectories: Trajectory 1: IDLE → IAM → ACM → CON → SRI (legitimate parameters) → success (compliant path).

[0025] Trajectory 2: IDLE → IAM → REL → RLC → SRI (tampered parameters) → failure (non-compliant path filtered out).

[0026] Trajectory 3: IDLE → IAM → ACM → CPG → SRI (timing anomaly) → failure (timing constraint violation).

[0027] Flow matching process: Forward flow F(s0→IAM)=0.8 (dominated by compliance weight).

[0028] Reverse flow R(SRI)=0.3 (affected by historical attack data).

[0029] Flow conservation adjustment: reduce illegal transition edge weight, strengthen compliant path traffic.

[0030] Output results: Among the 100 generated trajectories, 92 pass the protocol constraint check.

[0031] Contains 8 different attack variants (parameter tampering, timing jump, nested attack, etc.).

[0032] Diversity index Diversity@10=0.73 (significantly better than GAN's 0.42).

[0033] In an embodiment, the anomaly detection layer receives the mixed traffic of the threat modeling layer and outputs a threat assessment result, including the following steps: Topology invariant feature extraction: use graph convolution layer (GCN) or graph attention layer (GAT) in graph neural network (GNN) combined with node degree centrality, clustering coefficient and other topological metric indicators, extract deep structure features of network topology through multi-scale graph pooling operation, construct time series graph snapshots where V represents a set of network element nodes (such as MME, SGSN), E: represents a set of signaling links (SS7 / Diameter / SIP protocol connections), represents a node feature matrix (traffic load, protocol type, etc.), and the first k smallest eigenvalues are taken as the feature vectors of the topology invariant embedding; Triple defense mask graph attention network: Mask graph attention network introduces a learnable attention mask mechanism, which weights and aggregates node features through multi-head self-attention modules. The mask generator dynamically calculates the attention weights of nodes / edges based on the deviation of current traffic features from historical behavior baselines, combined with a gating mechanism (such as GRU) to filter noise connections and strengthen the attention allocation of key paths. Meanwhile, residual connections and layer normalization are used to enhance the model's generalization ability to sparse attack patterns. Dynamic graph reconstruction graph neural network: Real-time reconstruction of signaling interaction graph structure using reinforcement learning framework or variational autoencoder (VAE), optimizing graph connectivity through edge prediction tasks and community discovery algorithms, combining time decay factor to weight historical graph snapshots. When detecting abnormal subgraphs (such as loop attacks or illegal jump paths), trigger local graph structure pruning and reconstruction, and ensure semantic consistency between reconstructed graphs and normal topology patterns through contrastive learning modules (such as SimCLR).

[0034] In one embodiment, the mask graph attention network performs dynamic attention allocation and attack path suppression, including the following steps: Dynamic attention mask generation: input the current time signaling node feature matrix (N is the number of nodes, d is the feature dimension) and the historical behavior baseline (input by sliding window statistics) into the mask generator, calculate the feature deviation through dynamic time warping, input the deviation into the gated recurrent unit (GRU), combine with the historical mask state to generate the dynamic mask ; Multi-head self-attention aggregation: aggregate node features through multi-head attention mechanism, and strengthen key paths through mask. For each attention head h, calculate the mask-adjusted attention coefficient, denoted as where is the value projection matrix, is the learnable attention vector, denotes the concatenation operation, the attention coefficient is normalized through mask weighted Softmax, and the neighbor features are aggregated, denoted as , concatenate and linearly transform the outputs of H attention heads, and then aggregate the node features;

[0035] In one embodiment, the dynamic graph reconstruction graph neural network performs adaptive defense reconstruction of the signaling interaction graph, including the following steps: Temporal fusion based graph structure encoding: past T time snapshots are aggregated by an exponential decay function, and the aggregated adjacency matrix and node features H are input into a graph convolutional network (GCN) to generate a dynamic context-aware graph representation. Reinforcement learning driven graph structure reconstruction: the current graph structure is encoded into a state vector s by the attack probability output by the anomaly detection layer, edge addition and deletion operations are defined as a discrete action space A, a multi-objective reward function is designed, including connectivity preservation, attack surface reduction, and topology stability, the policy network is updated using the proximal policy optimization algorithm, the expected return is estimated by Monte Carlo sampling, and the edge connection is dynamically adjusted to block the attack path. Abnormal subgraph pruning and contrastive learning: traverse the subgraph by breadth-first search (BFS), if the loop length is detected =5, mark it as an abnormal subgraph , calculate the reconstruction confidence of the edges in the abnormal subgraph , delete the edges with confidence lower than threshold τ, and the reconstruction confidence calculation formula is σ( , input the original graph G and the reconstructed graph into the contrastive learning framework, calculate the feature similarity by the projection head . Community-aware connectivity optimization: use the Louvain algorithm to detect communities in the reconstructed graph, get community labels, and introduce a community penalty term for cross-community edges , define a modularity loss function to make the edge density within the community higher than the random expectation.

[0036] Example explanation: Suppose the attacker in the signaling network forms a loop attack (D→E→F→D) by forging nodes D, E, and F, and the DGR-GNN performs the following operations: Temporal encoding: fuse the past 3 time snapshots to generate an adjacency matrix containing historical protocol patterns.

[0037] Reinforcement learning reconstruction: after the policy network detects the D-E-F loop, it chooses to delete the edge E→F, and the reward function gives positive feedback due to the increase.

[0038] Abnormal pruning: BFS detects a 3-node loop (L=3<5), calculates the reconstruction confidence of edge E→F , and performs pruning.

[0039] Contrastive learning: input the original graph and the reconstructed graph into SimCLR, calculate the feature similarity sim=0.85, and ensure the semantic consistency of the reconstructed graph and the normal topology.

[0040] Community optimization: Louvain algorithm divides D, E, and F into different communities, and the cross-community edge weight is suppressed to prevent attack path diffusion.

[0041] In an embodiment, the response treatment layer receives the threat assessment result of the anomaly detection layer, and issues a defense instruction to the signaling awareness layer, including the following steps Adaptive defense strategy engine: based on real-time collected signaling network traffic data, system logs and historical attack cases, a spatio-temporal convolutional neural network (TCN) and long short-term memory network (LSTM) hybrid model constructed by a deep learning framework (such as TensorFlow or PyTorch) is used to extract features and identify threat patterns from multi-source heterogeneous data, and a reinforcement learning algorithm (such as DQN or PPO) is used to dynamically adjust the defense threshold and generate differentiated response strategies; Dynamic ACL rule optimization: based on attack trajectory, five-tuple mask rules are generated, and through the combination of threat level and business criticality coefficient, a logarithmic priority calculation formula is used to complete the dual goals of precise attack traffic blocking and business continuity protection, the priority calculation formula is expressed as Business criticality coefficient (0-1) of network element i; Entropy-driven traffic speed limit: the entropy value is used to quantify the degree of traffic anomaly, the calculation formula is expressed as , Traffic of protocol p, when the entropy value is lower than the threshold value θ, the dynamic threshold adjustment formula is used Implement speed limit; Vulnerability repair suggestion system: match CVE vulnerability library and system components through Jaccard similarity algorithm, the calculation formula is expressed as , combined with an exponential decay-based risk assessment model, a hot patch solution containing sysctl parameter adjustment and iptables rule is generated using a template engine; STIX 2.1 compliance report generation: map attack events, defense actions, and vulnerability information into standardized objects, and through REST API, interface with SOAR platforms such as Demisto to realize threat intelligence output for NIST 800-53 standard verification.

[0042] Although embodiments of the present application have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and alterations can be made without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. An AI-based signaling network vulnerability attack simulation and prevention system, characterized in that, include: Signaling Awareness Layer: Employs a distributed probe matrix and protocol reverse analysis engine, deployed on core network elements to perform hardware-level decoding of SS7 / Diameter / SIP / 5G-NR protocols, and dynamically extracts protocol field-level metadata through the YARA rule base to generate standardized signaling logs; Threat Modeling Layer: Receives signaling logs output from the signaling perception layer, injects simulated attacks to generate mixed traffic, and outputs it to the anomaly detection layer. Relying on the quantum generative adversarial network and GFlowNet flow generation engine, it uses parameterized quantum gates to construct a multi-dimensional feature encoder, combines 3GPP protocol state machine constraints to generate compliant attack vectors, and completes attack trajectory diversity sampling through Diameter protocol AVP nested state tree and two-factor reward function. Anomaly Detection Layer: Receives mixed traffic from the Threat Modeling Layer, outputs threat assessment results to the Response and Handling Layer, integrates topology-invariant feature extraction, masked graph attention network, and dynamic graph reconstruction neural network, constructs perturbation-resistant convolutional kernels based on graph Laplacian matrix eigenvalue decomposition, the masked graph attention network adopts a three-layer defense mechanism: anomaly scoring MLP → binary mask generation → GAT aggregation and purification, the dynamic graph reconstruction neural network uses an entropy-driven reconstruction mechanism to compress signaling storm detection latency, i.e., real-time calculation of node information entropy, triggering local graph reconstruction when the threshold is exceeded; Response and Handling Layer: Receives threat assessment results from the anomaly detection layer, issues defense instructions to the signaling awareness layer, dynamically adjusts ACL rules and traffic rate limiting thresholds based on reinforcement learning through the adaptive defense strategy engine and vulnerability remediation suggestion system, generates hot patch deployment schemes by combining the CVE vulnerability database, and outputs a STIX 2.1 format compliance audit report for integration with the SOAR platform.

2. The AI-based signaling network vulnerability attack simulation and prevention system according to claim 1, characterized in that, The distributed probe matrix is ​​an asymmetric duplex probe matrix used for non-aware protocol capture of idle core network elements. The asymmetric duplex probe matrix uses FPGA hardware acceleration technology to perform physical layer traffic mirroring and protocol stack acceleration parsing of SS7 / Diameter / SIP / 5G-NR protocols. Combined with the protocol fingerprinting algorithm based on the N-gram statistical model and the dynamic state machine constructed by the hidden Markov model, it completes the reverse parsing of unknown protocol variants and hidden channels. By leveraging the templated matching mechanism of the YARA rule engine, pattern matching is performed on the raw traffic through predefined protocol field encoding rules, and binary data is mapped into structured metadata by combining field extraction formulas. By normalizing structured metadata, multi-protocol fields are mapped to a unified data model, and a structured log generation formula is used to encapsulate timestamps, protocol types, field value sets, and hash check values ​​to form signaling logs that conform to the CEF standard. The YARA rule base is hot-loaded and updated through a double buffering mechanism. Combined with the rule conflict detection system built by the Jaro-Winkler similarity algorithm, new attack features are injected into the parsing process in real time without causing rule redundancy.

3. The AI-based signaling network vulnerability attack simulation and prevention system according to claim 1, characterized in that, The specific steps for the threat modeling layer to receive signaling logs output by the signaling awareness layer and inject simulated attacks to generate hybrid traffic are as follows: Quantum Feature Encoder Construction: A quantum circuit constructed using parameterized quantum gates is used to map multidimensional signaling features to quantum superposition states through a quantum feature encoder. 3GPP Protocol State Machine Constraint Engine: The protocol state machine is formally modeled using Petri nets, and the attack vector is forced to follow the legal transition sequence of the SS7 / Diameter / SIP protocol through compliance constraint formulas, so that the generated attack operations comply with 3GPP specifications. GFlowNet Attack Trajectory Generation: Introducing the GFlowNet flow generation engine, by extending the state space modeling of the attack graph and the forward-backward flow matching algorithm, the diversity sampling of attack trajectories is achieved while maintaining protocol compliance; Two-factor reward function: Combining the weighted sum of attack effectiveness score and protocol compliance score, the attack effectiveness and compliance are dynamically balanced through reinforcement learning mechanism; Hybrid traffic output: The simulated attack traffic is mixed with the original signaling traffic in a proportional manner using a traffic synthesis formula, and the IEEE 1588v2 protocol is used to ensure that the attack traffic and the normal traffic are time-aligned, generating test traffic that conforms to the timing characteristics of real network.

4. The AI-based signaling network vulnerability attack simulation and prevention system according to claim 3, characterized in that, The 3GPP protocol state machine constraint engine, through a collaborative mechanism of formal verification and dynamic constraints, ensures that attack vectors conform to 3GPP specifications while maintaining their innovativeness. This includes the following steps: Formal modeling phase of Petri nets: First, construct the quintuple model of the protocol state machine. ,in, This represents the collection of protocol state libraries. Represents the set of protocol event transitions. Define directed arc connections. For the arc weight function, As the initial label distribution, the 3GPP specification is transformed into a computable topology using a quintuple model; State transition rule definition: Establish transition trigger condition formula: θ is a threshold parameter related to the protocol version, if and only if the transition Former warehouse When the weighted sum of the tags exceeds a threshold, a state transition is triggered. Construction of compliance constraint formulas: Design a multi-dimensional constraint function family, including protocol version compatibility constraints. λ is the version matching coefficient; state reachability constraint: , For state Minimum number of live markers; message timing constraints: , For time-series sensitivity parameters, The protocol specifies a time delay; therefore, the comprehensive constraint formula is: ; Attack vector verification process: This involves verifying the generated attack sequences. Perform state projection, constraint checks, and path integrity verification. The state projection is represented as... make The constraint check is expressed as The path integrity check is represented as Does it constitute The legal path in; Dynamic adjustment mechanism: The model parameters are updated through a real-time feedback loop, and the weights are adaptively represented as follows. , The learning rate, threshold calibration is expressed as , is the damping coefficient.

5. The AI-based signaling network vulnerability attack simulation and prevention system according to claim 3, characterized in that, The GFlowNet attack trajectory generation achieves a balance between protocol compliance and attack diversity by constructing a flow probability model, including the following steps; Attack graph state space expansion modeling: by constructing a weighted directed hypergraph The protocol state library With attack action The nodes are combined into composite state nodes, and a super-edge transition relationship and a protocol compliance weight function are defined, wherein the weight function integrates 3GPP constraints, state transition costs and path preference coefficients. A forward-reverse flow matching algorithm is adopted: By designing the probability flow conservation conditions for forward and reverse flows and the transition probability formula based on compliance weights and exploration coefficients, a bidirectional flow allocation mechanism is established. The forward flow F(s) represents the cumulative probability flow from the initial state to s, and the reverse flow R(s) represents the cumulative probability flow from s to the final state. The two are dynamically balanced through the flow conservation equation, i.e., for non-boundary states... , must meet Where Succ(s) is the set of successor states of s. To allocate the forward flow, and with the design of adjustable transition probabilities, the randomness and goal orientation of path exploration are balanced while ensuring protocol compliance. Define a two-component training objective function: it includes a flow matching loss term based on state distribution sampling and an edge matching loss term based on KL divergence. The flow matching loss term based on state distribution sampling ensures the consistency of the flow probability model by minimizing the difference between the forward flow and the reverse flow plus the transition flow. The flow matching loss term is expressed as follows: ,in, For state distribution, For the forward / backward network parameters, the edge matching loss term based on KL divergence improves the semantic rationality of the generated trajectory by approximating the distribution of real attack patterns. The edge matching loss term is expressed as... Generate trajectory distribution θ Distribution of attack patterns approaching reality data ; Diverse sampling strategy: Introduce a trajectory balancing sampler and construct a sampling weight formula by combining traffic flow ratio and diverse temperature parameters to achieve probability-driven diversified path selection within the compliance constraint framework.

6. The AI-based signaling network vulnerability attack simulation and prevention system according to claim 1, characterized in that, The anomaly detection layer receives the mixed traffic from the threat modeling layer and outputs the threat assessment results, including the following steps: Topology-invariant feature extraction: Utilizing graph convolutional layers or graph attention layers in graph neural networks combined with topology metrics, deep structural features of the network topology are extracted through multi-scale graph pooling operations, constructing a time-series graph snapshot. Where V represents the set of network element nodes, and E represents the set of signaling links. It is represented as a node feature matrix, and the eigenvectors corresponding to the first k smallest eigenvalues ​​are taken as topologically invariant embeddings; Three-layer defense mask graph attention network: The mask graph attention network introduces a learnable attention mask mechanism, which uses a multi-head self-attention module to weight and aggregate node features. The mask generator dynamically calculates the attention weight of nodes / edges based on the deviation between the current traffic features and the historical behavior baseline. Combined with a gating mechanism, it filters noisy connections and strengthens the attention allocation of critical paths. Dynamic Graph Reconstruction Graph Neural Network: It utilizes reinforcement learning frameworks or variational autoencoders to reconstruct the signaling interaction graph structure in real time, optimizes graph connectivity through edge prediction tasks and community detection algorithms, and performs weighted fusion of historical graph snapshots by combining time decay factors. When abnormal subgraphs are detected, local graph structure pruning and reconstruction are triggered, and semantic consistency between the reconstructed graph and the normal topology pattern is ensured through a contrastive learning module.

7. The AI-based signaling network vulnerability attack simulation and prevention system according to claim 6, characterized in that, The mask graph attention network performs dynamic attention allocation and attack path suppression, including the following steps: Dynamic attention mask generation: This involves generating the feature matrix of the signaling node at the current time step. Compared with historical behavioral baseline Input mask generator, calculate feature bias through dynamic time warping , will deviation Input gated loop unit, combined with historical mask state Generate dynamic mask ; Multi-head self-attention aggregation: Node features are aggregated through a multi-head attention mechanism, while critical paths are reinforced using masks. For each attention head h, the mask-adjusted attention coefficients are calculated, denoted as... ,in, The projection matrix is ​​the value. For learnable attention vectors, This represents the concatenation operation, which uses mask-weighted Softmax normalized attention coefficients and aggregates neighbor features, as shown below. The outputs of H attention heads are concatenated and linearly transformed to aggregate node features. Gated noise filtering and residual update: Suppress sparse attack patterns through gating mechanisms while preserving normal topology information.

8. The AI-based signaling network vulnerability attack simulation and prevention system according to claim 6, characterized in that, The dynamic graph reconstruction graph neural network performs adaptive defense reconstruction of the signaling interaction graph, including the following steps: Graph structure encoding based on temporal fusion: The graph snapshots of the past T time steps are weighted and aggregated by an exponential decay function. The aggregated adjacency matrix and node features H are then input into a graph convolutional network to generate a dynamic context-aware graph representation. Reinforcement learning-driven graph structure reconstruction: reconstructing the current graph structure The attack probability output by the anomaly detection layer is encoded as a state vector. Edge addition and deletion operations are defined as discrete action space A. A multi-objective reward function is designed, which includes a connectivity preservation term, an attack surface reduction term, and a topology stability term. The policy network is updated using a near-end policy optimization algorithm. The expected reward is estimated through Monte Carlo sampling, and then the edge connections are dynamically adjusted to block the attack path. Anomaly Subgraph Pruning and Contrastive Learning: Subgraphs are traversed using breadth-first search; if a loop is detected... (like If the value is 5, it is marked as an abnormal subgraph. ,right In the process of calculating the reconstruction confidence score of edges, edges with a confidence score lower than a threshold τ are deleted. The reconstruction confidence score calculation formula is expressed as follows: Compare the original graph G with the reconstructed graph Input contrastive learning framework, through projection head Calculate feature similarity ; Community-aware connectivity optimization: The Louvain algorithm is used to detect communities in the reconstructed graph, obtain community labels, and optimize cross-community edges. We introduce a community penalty term and define a modularity loss function to ensure that the edge density within the community is higher than the random expectation.

9. The AI-based signaling network vulnerability attack simulation and prevention system according to claim 1, characterized in that, The response and handling layer receives the threat assessment results from the anomaly detection layer and issues defense instructions to the signaling awareness layer, including the following steps: Adaptive defense strategy engine: Based on real-time collected signaling network traffic data, system logs and historical attack cases, it uses a hybrid model of spatiotemporal convolutional neural network and long short-term memory network (LSTM) built with a deep learning framework to extract features and identify threat patterns from multi-source heterogeneous data, and combines reinforcement learning algorithms to dynamically adjust defense thresholds and generate differentiated response strategies. Dynamic ACL rule optimization: Five-tuple mask rules are generated based on attack trajectories. By combining a logarithmic priority calculation formula that integrates threat level and business criticality coefficients, the dual objectives of precise attack traffic blocking and business continuity assurance are achieved. The priority calculation formula is expressed as follows: For network element i, the service criticality coefficient is... Entropy-driven rate limiting: This method quantifies the degree of traffic anomalies by calculating the protocol distribution entropy. The calculation formula is expressed as follows: , For the traffic of protocol p, when the entropy value is lower than the threshold θ, a dynamic threshold adjustment formula is used. Speed ​​limits will be implemented. Vulnerability remediation suggestion system: This system uses the Jaccard similarity algorithm to match CVE vulnerability databases with system components. The calculation formula is expressed as follows: Combining a risk assessment model based on exponential decay, a hot patching scheme containing sysctl parameter adjustments and iptables rules is generated using a template engine. STIX 2.1 Compliance Report Generation: Maps attack events, defense actions, and vulnerability information into standardized objects, and connects with SOAR platforms such as Demisto via REST API to achieve threat intelligence output verified by the NIST 800-53 standard.

Images