Analyzable anti-attack network security method and system based on AI unified model

Abstract

The invention provides an analyzable anti-attack network security method and system based on an AI unified model. The method comprises the following steps: S1, carrying out attack source tracing and attack mode identification on an input data stream; s2, performing protocol structure analysis and grammar element extraction on the input data stream in a grammar verification layer, and starting a grammar rule matching process to obtain a grammar exception perception set; s3, fusing attack vector information on the basis of a grammar anomaly perception set in a semantic analysis layer, constructing a semantic relation graph, and outputting a semantic risk vector; s4, taking the semantic risk vector as input, combining a business scene, resource constraint and strategy preference, modeling a defense target, and outputting an optimal response path; and S5, forming a model evolution path based on local feedback and global collaboration. Through a three-layer full-information analysis mechanism and behavior feedback driving, interpretable recognition of attack intentions and collaborative optimization of defense paths are realized, attack recognition is comprehensive, response decision is accurate, and strategy evolution is controllable.

Description

Technical Field

[0001] The present invention relates to a network security method, in particular to a resolvable anti-attack network security method based on an AI unified model, and further to a resolvable anti-attack network security system that adopts the resolvable anti-attack network security method based on an AI unified model. Background Art

[0002] The revolution in the artificial intelligence (AI) paradigm has brought new technological potential to cyberspace security. Mechanism-based AI theory, by reconstructing the three-dimensional information synergy of "syntax, semantics, and pragmatics," has significantly improved the interpretability and inherent security of intelligent systems. Traditional AI research has long been constrained by the traditional material science paradigm, leaving security defense systems in a difficult, reactive state. With the widespread application of AI technology in code generation, traditional code security assurance systems are facing the daunting challenge of a paradigm shift. Currently, the core contradiction in cybersecurity lies in how to establish an interpretable, verifiable, and traceable synergy between intelligent generation efficiency and inherent security.

[0003] Current network security methods generally rely on rule matching and static model judgment, separating the deployment of traffic analysis, anomaly detection and policy generation. There are three main technical shortcomings: First, grammatical rule recognition is limited to protocol templates and field matching, making it difficult to understand the attack behavior chain behind the data, resulting in a significant decrease in recognition accuracy in highly complex attack scenarios; second, semantic layer modeling lacks structural traceability, making it impossible to restore the context of attack paths and annotate behavioral intentions, resulting in a blind spot in the path of security response; third, policy deployment relies heavily on manual configuration and preset policy libraries, making it difficult to adapt to the rapid evolution of attack methods, resulting in low response timeliness and policy adaptation rate.

[0004] In addition, most existing systems use black box models or heuristic rule sets, lacking a structured cognitive loop. This results in the inability of policy deployment effects to reversely influence model judgments, and makes it difficult to accumulate and evolve security knowledge, causing the entire defense system to be in a non-adaptive state of "response island" for a long time.

[0005] Therefore, in the existing cyberspace defense system, there are still key technical bottlenecks in network security, such as incomplete attack identification, inaccurate response decisions, and uncontrollable strategy evolution. Summary of the Invention

[0006] The technical problem to be solved by this invention is the need to provide a resolvable anti-attack network security method based on a unified AI model. This method aims to achieve interpretable identification of attack intent and collaborative optimization of defense paths through the structured integration of three layers of information: syntax verification, semantic parsing, and pragmatic optimization, as well as behavioral feedback drive. This solves the technical problems of the prior art, such as incomplete attack identification, inaccurate response decisions, and uncontrollable strategy evolution. Furthermore, a resolvable anti-attack network security system that utilizes this resolvable anti-attack network security method based on a unified AI model is further provided.

[0007] To this end, the present invention provides a resolvable anti-attack network security method based on an AI unified model, comprising the following steps: Step S1: tracing the attack source and identifying the attack method of the input data stream; Step S2: Perform protocol structure analysis and grammatical element extraction on the input data stream at the grammar verification layer, start the grammar rule matching process, and obtain a grammar anomaly perception set; Step S3: At the semantic parsing layer, based on the grammatical anomaly perception set, the attack vector information is integrated, the semantic entities in the behavior chain are extracted to construct a semantic relationship graph, and the constructed semantic relationship graph is iteratively calculated to output a semantic risk vector; Step S4: In the pragmatic optimization layer, the semantic risk vector output by the semantic analysis layer is used as input. The defense target is modeled based on the current business scenario, resource constraints, and policy preferences. Through the policy planning and scheduling mechanism, an optimal and executable response path is output. In step S5, after detecting and responding to an attack, each edge node first encapsulates it into a structured response knowledge unit. Then, a structured desensitized knowledge vector is constructed through encoding and compression and encrypted and uploaded to the coordination center. Finally, after completing knowledge aggregation, the coordination center transmits feedback information back to each edge node to achieve synchronous updating of model parameters and policy rules, and adjusts the construction logic and scoring parameters of the semantic relationship graph to form a model evolution path based on local feedback and global collaboration.

[0008] A further improvement of the present invention is that step S1 includes the following sub-steps: Step S101, identifying and analyzing the attack source; Step S102: identifying and classifying attack methods; Step S103: Encode the attack source analysis results and the attack method labels to generate a multi-dimensional feature vector of the attack behavior. The multi-dimensional feature vector includes source features, behavior features, communication structure features, and potential target orientation. The feature modeling process of the multi-dimensional feature vector adopts an adaptive vector projection algorithm. Step S104: encapsulate the multidimensional attack feature set corresponding to the multidimensional feature vector into an attack vector in a unified format as the standard input of the syntax verification layer; wherein the attack vector includes the attack source address, method category, feature dimension code, attack behavior label and current connection context; while encapsulating and generating the attack vector, a behavior identifier is assigned to the attack vector.

[0009] A further improvement of the present invention is that the process of identifying and analyzing the attack source in step S101 is as follows: first, real-time communication behavior is captured and preprocessed by a traffic monitoring module deployed at the edge, and metadata is extracted, wherein the metadata includes the source address, destination address, protocol type, port information, packet frequency, and connection duration; then, a connection trajectory table with a five-tuple as the session identifier is constructed, and a sliding window and timestamp association mechanism is used to reconstruct the access path structure of a single source and multiple targets to form a trajectory tree, wherein the five-tuple includes the source address, destination address, protocol, port, and timestamp; then, a behavior deviation detection algorithm is used to compare and analyze the indicators of connection frequency, request direction change, and abnormal retry, and path branches that may contain malicious behavior are marked from the trajectory tree; finally, for each source node, its trust risk index is calculated based on its access pattern, historical label, communication path complexity, and similarity with known attack IP addresses. When the trust risk index exceeds a preset dynamic threshold, the node is determined to be a potential attack source, and its communication context, path map, and identification information are extracted as input for subsequent classification and modeling; The process of identifying and classifying attack methods in step S102 is as follows: based on the identification and analysis of the attack source, a pre-trained attack behavior classifier is called to identify the attack method of suspicious traffic; the process of identifying the attack method of suspicious traffic is represented by a fusion feature vector. First, the grammatical anomaly features, time series features, and context features are used as input, and the attack methods are matched and classified through a semantic weighting mechanism. The grammatical anomaly features include field out-of-bounds and protocol nesting anomalies, and the time series features include inter-packet delay and connection frequency; then, the attack patterns are matched and classified through a semantic weighting mechanism, and the identification results of each attack method are output as a structured annotation, which is used as an attack behavior label; the attack behavior label includes an attack label, feature confidence, and an impact path.

[0010] A further improvement of the present invention is that step S2 includes the following sub-steps:

[0011] Step S201 automatically identifies the communication protocol stack used by the current traffic, performs hierarchical decoding on the data packet, and systematically decomposes the header field, payload content, parameter structure, and method call sequence in the message to form a structured grammatical parse tree. When forming the structured grammatical parse tree, by integrating the protocol specification knowledge base and attack behavior labels, the key fields in the structured grammatical parse tree are marked with type identifiers as pre-labels for anomaly annotation and risk weight calculation. Each node of the structured grammatical parse tree corresponds to a protocol field or grammatical fragment, which is used to annotate its field name, field type, occurrence position, and nesting level. The nodes are connected by edges to represent the dependency order and combination structure to construct a structured grammatical representation, which includes field sequence, structural hierarchy, protocol template ID, and label information. In step S202, based on the constructed grammatical representation, the grammatical rule matching process is started. The rule library corresponding to the grammatical rule matching process adopts a three-layer architecture design, which includes a basic protocol specification layer, a business scenario customization layer, and a dynamic abnormal behavior strategy layer. In the grammatical rule matching process, the fields of each node in the structured grammatical parsing tree are matched and retrieved to determine whether they comply with the corresponding grammatical template and calling specifications, and a field-level exception list and its corresponding risk indicators are output to form a grammatical anomaly perception set.

[0012] A further improvement of the present invention is that step S3 includes the following sub-steps: Step S301: Based on the grammatical anomaly perception set, attack vector information is integrated to extract semantic entities in the behavior chain and map the semantic entities into a semantic relationship graph. The semantic entities include the request initiator, action type, target object, and scenario annotation. The semantic relationship graph uses the nodes of the structured grammatical parse tree as boundary nodes and the semantic entities as semantic center nodes. Edge weights are used to represent the association strength and operation sequence between semantic entities. Step S302: Input the semantic relationship graph into a graph neural network model, use a multi-layer graph convolution mechanism to identify the attack intent of the semantic relationship graph, and logically annotate its behavior path to obtain the intent path; Step S303: Score the intention paths marked in the semantic relationship graph. The scoring dimensions include the atypicality of the operation sequence, the sensitivity level of the target involved, and the abnormal weight of the upstream field. The scoring dimensions are normalized, and combined with the attack type label and the behavior sequence confidence, a vector splicing method is used to align the score of each scoring dimension with the label vector to form a fixed-length feature vector, which is then encapsulated into the final semantic risk vector. The semantic risk vector is represented by a fixed-length structure. The content of the semantic risk vector includes the attack type, intention weight, impact range level, and reasoning confidence.

[0013] A further improvement of the present invention is that in step S302, the semantic relationship graph is iteratively calculated for multiple rounds through the graph neural network model. In the convolution process of each round of iterative calculation, the feature vector of each node is calculated using the formula: Update, where Indicates the Layer Time Node The eigenvector of is a node The neighbor set of For the Layer trainable weight matrices, From the adjacent node To Node The attention weight, is the activation function

[0014] A further improvement of the present invention is that step S4 includes the following sub-steps: Step S401: First, according to the attack type, intention weight, impact range level and reasoning confidence contained in the semantic risk vector, the corresponding defense target and control boundary are derived; then, for the current business scenario, the formula Construct the corresponding policy objective function ,in, Indicates blocking capability, Indicates resource consumption, Indicates response delay, 、 and Preset weight coefficients representing blocking capability, resource consumption, and response delay respectively; Step S402: First, according to the attack type and system constraints, a matching candidate policy set is retrieved from the policy knowledge base. The candidate policy set is represented as ,in, Represents the nth candidate strategy, where n is a natural number. The attributes of each candidate strategy include the applicable attack type label, target module or interface application scope, operation cost measurement value, strategy conflict matrix, and historical response effect score and negative impact record; then, through the formula Construct the subset of all legally combinable policies , obtain strategy subset combinations through search and filtering , and for each strategy subset combination Perform utility scoring. The comprehensive utility scoring function used for utility scoring is: ,in, represents the comprehensive utility score, Representation Strategy Scoring the coverage of key nodes in the attack path, Representation Strategy The average resource consumption of Representation Strategy Estimated response time; In step S403, the response instruction set is sent to the security execution engine, which calls the corresponding network control interface, access control module or identity management subsystem according to the policy content to implement the actual deployment of the policy. During the policy execution process, the execution status, response delay, target blocking status and system feedback information are recorded in real time to form a complete response behavior log.

[0015] A further improvement of the present invention is that in step S402, a combination of greedy search and constraint pruning is used in the process of searching and screening to realize the combination of strategy subsets. The filter follows the following conditions: Condition A1, candidate strategy set The covered attack nodes must include at least one high-confidence intention node in the semantic relationship graph; Condition A2: the same target module is not allowed to deploy strategies with mutually exclusive attributes; Condition A3: Response delay must not exceed the maximum acceptable value set by the system for the current business scenario ; Condition A4: If there are multiple strategy subset combinations If conditions A1 to A3 are met, select the comprehensive utility score The largest group is used as the current optimal strategy path , and generate the corresponding response instruction set.

[0016] A further improvement of the present invention is that step S5 includes the following sub-steps: Step S501: The listener module records the feedback status of the policy node in real time and determines the defense effectiveness based on the change in the attack chain status. The feedback status includes whether the defense instruction is successfully triggered, whether the attack chain has been intervened, and whether the security status of key resources in the system has changed. The observation information corresponding to the feedback status is encapsulated as a response knowledge unit and injected back into the semantic parsing layer and policy scoring module. The fields of the response knowledge unit include the policy identifier, target path node, execution status, attack behavior change indicator, system side effect indicator, and confidence weight adjustment suggestion. Step S502: Implementing cross-domain collaborative defense. Edge nodes periodically extract representative samples from the response knowledge units, encode and compress the attack types, policy intervention results, and key behavioral characteristics of the representative samples, construct structured desensitized knowledge vectors, and upload them to the coordination center via an encrypted channel as shared information. The coordination center clusters and statistically analyzes feedback from multiple nodes, extracting summary evaluation results of policy effects under different attack labels. After receiving the feedback evaluation results, each edge node initiates an update. The update process includes: the semantic parsing layer expands the local semantic relationship graph and updates the node type dictionary based on the feedback of new attack intentions or path structures; the policy scoring module adjusts the weight coefficients and priority sorting rules in the comprehensive utility scoring function. Step S503: When a certain strategy subset combination If a strategy shows a stable interruption effect in a preset number of deployments, the adaptation weight between the strategy and its corresponding semantic node will be automatically strengthened; if a strategy continuously shows intervention failure or large side effects on different paths, its comprehensive utility score will be reduced and it will be removed from the candidate strategy set within the preset time window.

[0017] The present invention also provides a resolvable anti-attack network security system based on an AI unified model, which adopts the resolvable anti-attack network security method based on an AI unified model as described above, and includes: Attack identification module, used to trace the attack source and identify the attack method of the input data stream; The grammar verification module parses the protocol structure and extracts grammatical elements from the input data stream at the grammar verification layer, starts the grammar rule matching process, and obtains the grammar anomaly perception set; The semantic parsing module, based on the grammatical anomaly perception set at the semantic parsing layer, integrates attack vector information, extracts semantic entities in the behavior chain to construct a semantic relationship graph, and iteratively calculates the constructed semantic relationship graph to output a semantic risk vector. The pragmatic optimization module, in the pragmatic optimization layer, uses the semantic risk vector output by the semantic analysis layer as input. It combines the current business scenario, resource constraints, and policy preferences to model the defense target and output an executable optimal response path through a policy planning and scheduling mechanism. In the knowledge evolution and closed-loop feedback module, after detecting and responding to attacks, each edge node is first encapsulated into a structured response knowledge unit. Then, a structured desensitized knowledge vector is constructed through encoding and compression and encrypted and uploaded to the coordination center. Finally, after completing knowledge aggregation, the coordination center transmits feedback information back to each edge node to achieve synchronous updating of model parameters and policy rules, and adjust the construction logic and scoring parameters of the semantic relationship graph to form a model evolution path based on local feedback and global collaboration.

[0018] Compared with the existing technology, the beneficial effects of the present invention are: through the structured connection of the three layers of information of grammar verification, semantic analysis and pragmatic optimization, as well as behavioral feedback drive, the "grammar-semantics-pragmatics" three-layer full information analysis mechanism is reconstructed, and the integrated fusion of attack intention identification, response strategy generation and feedback-driven evolution is realized; the present invention can fundamentally improve the accuracy of attack identification and the resilience of defense response, realize the enhanced interpretability of strategy paths and the adaptive evolution of scoring mechanisms, and realize the distributed expansion of the system's collaborative defense capabilities and the continuous enhancement of the knowledge update mechanism, solving the technical problems of the existing technology such as incomplete attack identification, inaccurate response decisions and uncontrollable strategy evolution. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a schematic diagram of a system logic framework of an embodiment of the present invention; Figure 2 This is a schematic diagram of the security issues of the three major schools in the existing technology; Figure 3 It is a three-dimensional perspective diagram of total information theory; Figure 4 It is a schematic diagram of a system closed-loop feedback optimization process according to an embodiment of the present invention. DETAILED DESCRIPTION

[0020] Before introducing the preferred embodiments of the present invention in detail, the relevant technologies are first described and illustrated.

[0021] Regarding the definitions of abbreviations and key terms, PH refers to Package Hallucination; COT refers to Chain of Though; AST refers to Abstract Syntax Tree; m-GTAI refers to mechanism-based theory of AI.

[0022] On the dilemma of artificial intelligence safety research.

[0023] Because current AI research adheres to the traditional disciplinary paradigm of divide-and-conquer and pure formalism, the overall field of AI research has been fragmented, resulting in three major schools of thought: structuralism, functionalism, and behaviorism. These three schools of thought all have significant limitations in the areas of cybersecurity and code security. More importantly, these limitations are not due to accidental errors in AI technology, but rather to systemic flaws in the underlying research paradigm.

[0024] Structuralism simplifies intelligence to a simple imitation of the human brain's organizational structure, employing black-box models that rely heavily on statistical learning. These models overly focus on the syntactic features of the data during training, ignoring semantic and pragmatic information. This results in a lack of interpretability, and their divide-and-conquer training approach undermines the organic unity of program semantics, pragmatics, and grammar.

[0025] The functionalist school attempts to construct formal security rules through expert systems, which requires abstracting security logic into static symbolic reasoning. Many dynamic semantic information are fixed as conditional judgment statements, resulting in the researched security strategies often being divorced from real security scenarios and only performing static symbolic reasoning.

[0026] The behaviorist school focuses on the immediate feedback of perception and action, and regards the human thinking process as a conditioned reflex and stimulus response process. In the security field, it has evolved into a threat response mechanism centered on behavioral feature matching, blocking malicious behavior by monitoring shallow data indicators such as API call frequency and network traffic peaks. However, this school is short-sighted in local optimization issues, which can easily lead to global imbalance and make it difficult to balance security and utility. In particular, when privacy and efficiency conflict, decisions are often made that are either this or that. The current status of security issues of the three major schools is as follows: Figure 2 shown.

[0027] While the three schools of thought have distinct research paths, they face similar challenges in the security field. They all view security as an external attribute of intelligent systems, rather than an inherent element, leading to a lack of inherent security. When attackers exploit semantic information to construct logic bombs or rely on pragmatic information to infiltrate permission systems, the limitations of traditional AI become weak points in security defenses.

[0028] A breakthrough in total information theory.

[0029] During the paradigm shift in artificial intelligence, the emergence of "Total Information Theory" marked a fundamental overturn in traditional views of information. Traditional intelligence research, long constrained by the material science paradigm, simplified information into symbolic manipulation at the "syntax" level, ignoring the deeper connections between "semantics" and "pragmatics" in human intelligence. This disconnect directly contributed to AI's inability to understand complex scenarios. The key point of Total Information Theory is to reconstruct the information environment between syntax, semantics, and pragmatics, providing a theoretical foundation for the explainability of intelligent systems.

[0030] Syntactic information is the most superficial dimension of information, presented as a data structure, such as the sequence of code characters or the structure of network protocol messages. Human perception systems can directly grasp the formal characteristics of external stimuli, and sensors and statistical models in the field of artificial intelligence can extract syntactic information. However, such intelligent systems alone become "blind signal processors." For example, firewalls can only detect abnormal byte distributions in data packets but cannot identify malicious payloads, and code review tools can only identify syntactic errors but cannot understand the attack intent of logical vulnerabilities.

[0031] The introduction of pragmatic information imbues grammatical data with a value dimension, revealing the "meaning of information to the subject's goals" and forming the core basis for intelligent system decision-making. The human attention mechanism is essentially pragmatic screening, with the brain sorting information based on its relevance to the target. When monitoring network traffic, operations and maintenance personnel focus on anomalous conversations based on their accumulated pragmatic judgments. In an AI environment, pragmatic information is reflected in the logic for generating security policies. When an intrusion detection system identifies signs of high-frequency port scanning, it must determine the response level based on the value weight of the network assets, rather than mechanically blocking based solely on grammatical features.

[0032] Semantic information connects form and meaning, focusing on the essence of information content. Human comprehension of language relies on semantic networks, combining concepts into knowledge systems. The semantics of the word "encryption" are linked to deeper concepts such as "key management" and "algorithm strength." A lack of semantics in the security field can lead to logical discontinuities.

[0033] The major breakthrough of full information theory lies in revealing the dynamic transformation mechanisms of grammar, pragmatics, and semantics. The human brain, through the "information-knowledge-intelligence" transformation chain, integrates grammatical information and prior knowledge to achieve semantic parsing and target value decision-making. This is reflected in cybersecurity as a multi-layered defense coordination system. The core flaw of the three major schools of traditional artificial intelligence is that they fragment the integrity of full information. Only by using full information as a link can an intelligent security architecture with "verifiable form, explainable content, and traceable value" be established.

[0034] This theoretical innovation provides a new "three-dimensional perspective" for the paradigm shift in artificial intelligence. Syntactic information enables precise interaction between machines and the physical world, semantic information gives them a deep understanding of behavioral logic, and pragmatic information plays a crucial role in enabling value rationality in complex environments. Academics once viewed semantics, grammar, and pragmatics as parallel concepts at the same level. However, grammar and pragmatics are intuitive concepts, while semantics requires abstraction from their collective understanding and is a higher-level concept.

[0035] About the core advantages of mechanism-based artificial intelligence.

[0036] Traditional artificial intelligence research has long been constrained by the "divide and rule" and "pure formalization" paradigms of traditional material sciences, resulting in the intelligent generation mechanism falling into an unexplainable black box state and a fragmented cognitive dilemma. The mechanismist artificial intelligence theory forms a universal intelligent generation mechanism by constructing a "full information-knowledge-intelligence" conversion chain. Its theoretical breakthroughs are concentrated in the two dimensions of explainability and universality.

[0037] Interpretability stems from the transparent transformation of semantic information. Mechanism-based artificial intelligence theory emphasizes that the essence of intelligence lies not in the mechanical permutation and combination of symbols or the statistical correlation of data, but rather in the co-evolution of the three-dimensional information ecosystem of "syntax, semantics, and pragmatics." Taking secure code generation as an example, traditional models focus on surface features such as variable naming and function calls. Mechanism-based models map the grammatical structure of code to semantic logic and pragmatic dimensions. For example, they transform the grammatical features of "file read and write operations" into semantic representations of "data leakage risk." Dynamic permission constraints prevent vulnerabilities, allowing security policies to be based on a deep semantic understanding of code behavior, freeing them from reliance on empirical rules.

[0038] Universality stems from the inclusiveness of the theoretical framework of the intelligent generation mechanism. The four types of information conversion principles proposed by mechanismism form a closed-loop feedback system that can flexibly adapt to security needs in multiple fields. In network behavior authentication scenarios, traditional models rely on superficial indicators such as traffic peaks and API call frequencies. The mechanismist model transforms network behavior data into semantic representations of user behavior intentions, combines them with business value weights to generate dynamic authentication strategies. The intelligent agent generates multi-path solutions through semantic mapping and pragmatic trade-offs, breaking through the rigid rule-based limitations of traditional expert systems.

[0039] About the current status of network security defense field.

[0040] Traditional cybersecurity defense technologies are currently facing severe challenges. Research shows that existing defense systems have systemic flaws in responding to new cyberattacks. The root causes lie in three major contradictions: fragmented attack detection dimensions, security policies relying on static rules, and unexplainable security logic.

[0041] At the attack detection mechanism level, existing technologies face the dual dilemma of shallow feature reliance and a lack of dynamic adaptability. While detection models based on feature engineering perform well against known attack patterns, they have a high underreporting rate for dynamic threats such as zero-day attacks and polymorphic obfuscation. More seriously, while existing deep learning models can improve detection rates, their black-box nature renders defense decisions untraceable, severely limiting the accuracy and auditability of defense strategies.

[0042] Existing security enhancement solutions often exhibit fragmented "local optimization" characteristics, making it difficult to build a global collaborative defense ecosystem. Threat intelligence platforms based on rule engines rely on manually defined attack patterns, resulting in high false positive rates for APT attacks and delayed rule base updates. Knowledge sharing in emerging technologies such as federated learning lacks semantic alignment, and local node threat signatures lack cross-domain semantic mapping, resulting in reduced global model generalization capabilities. The collaborative defense framework's lack of a unified utility evaluation model leads to conflicting defense actions, exposing contradictions between security, performance, and business objectives.

[0043] A more fundamental security crisis stems from a cognitive divide in technological approaches. Cybersecurity defense has diverged into two approaches: "feature-driven" and "behavior-driven." The former relies on statistical features at the syntactical level, making it difficult to parse attack semantics, while the latter ignores protocol syntax compliance and can easily misjudge legitimate operations. Currently, the cybersecurity defense field is at a critical stage of transition from "passive response" to "active immunity." While some progress has been made, the fragmented detection dimensions, static policy generation, and black-box security logic remain systemic threats. Future breakthroughs will require transcending the boundaries of traditional tools, embedding full information theory to construct a three-dimensional collaborative defense system, and achieving a paradigm shift from "feature-based interception" to "cognitive immunity."

[0044] An existing technology related to the present invention is CrossAlert, which is a multi-stage network attack detection method based on semantic embedding and prototype network, aiming to solve the problems of high false alarm rate and insufficient domain adaptability of traditional intrusion detection systems in complex multi-stage attack (MSA) scenarios.

[0045] In terms of technical implementation, CrossAlert first performs multi-stage feature modeling on the raw alert data, including basic features (source IP address, destination IP address frequency encoding, and attack phase labels) and sequence features (contextual alert sequences based on sliding windows). By manually annotating 20% ​​of the alert data, the BERT model is fine-tuned to generate semantic embeddings, leveraging the BERT attention mechanism to capture temporal dependencies. To address domain transfer, the algorithm incorporates a prototypical network for few-shot learning. In scenarios where target domain data is scarce, the prototypical network dynamically adjusts classification boundaries by calculating the Euclidean distance between the alert feature vector and the source domain attack phase prototype.

[0046] This technology, validated through datasets, significantly improves detection performance compared to traditional rule-matching methods. However, this existing technology suffers from the following drawbacks: While CrossAlert technology improves multi-stage attack detection capabilities through semantic embedding and prototype networks, its technical architecture still suffers from systemic flaws and cannot meet the network security requirements of dynamic adversarial environments. This technology relies on manually predefined attack stage labels, which rigidifies the semantic parsing process into a static rule mapping. Attack stage classification requires manual labeling of 20% of alert data to fine-tune the BERT model. This predefined labeling system cannot be dynamically expanded when faced with variant attacks or new attack methods, leading to misclassifications. For example, when an attacker delivers encrypted malicious payloads via the legitimate HTTPS protocol, the model, lacking the ability to deeply reason about semantic ambiguity, may classify high-risk attacks as normal traffic, while high-frequency visits from legitimate users may be mislabeled as attacks due to triggering static feature thresholds. Furthermore, the attack stage classification is based on empirical summaries of historical attack patterns and is not dynamically linked to contextual pragmatic information, resulting in security decisions being detached from actual protection needs.

[0047] This technology's domain adaptation mechanism uses a prototype network to achieve cross-domain knowledge transfer, but its essence remains shallow feature alignment, failing to address the fundamental impact of semantic drift on detection performance. The prototype network matches source and target domain feature vectors using Euclidean distance, but does not model the logical association of cross-domain semantics. Furthermore, the high-dimensional concatenation of BERT embeddings and sequence features results in high computational complexity, making it difficult to support high-concurrency, real-time detection requirements when deployed on edge nodes, severely restricting the technology's large-scale application.

[0048] From a defense paradigm perspective, CrossAlert continues the passive response logic of traditional intrusion detection systems. Its anomaly scoring module focuses on statistical features at the grammatical layer and fails to establish a semantic risk map of attack intent. Logic bombs or lateral penetration behaviors constructed by attackers through legitimate protocols bypass detection because they conform to grammatical rules, exposing blind spots in semantic layer defense. At the same time, rule base updates rely on the manual import of Snort feature sets. When faced with zero-day attacks or APT attack chains, the generation of defense strategies lags behind the attack evolution cycle, creating a vicious cycle of "rule update-attack breakthrough." More critically, the final classifier uses a black-box neural network decision-making mechanism, which cannot trace the reasoning path associated with the alert, making it difficult for security personnel to verify the rationality of defense actions and easily burying audit risks in high-compliance scenarios.

[0049] The root cause of these flaws lies in the cognitive fragmentation of the technical approach: semantic parsing, anomaly scoring, and domain adaptation are mechanically separated into independent modules, without achieving the coordinated verification of grammatical, semantic, and pragmatic information through a unified theoretical framework. This "divide and conquer" strategy is essentially a continuation of the traditional material science paradigm, and is fundamentally inconsistent with the cross-dimensional penetration characteristics of high-level attacks. The static labeling system, shallow migration mechanism, and passive detection logic of existing technologies are unable to build an intrinsic security architecture that is "formally verifiable, semantically parsable, and strategically traceable." A paradigm shift is needed to upgrade defensive capabilities.

[0050] Another related prior art is AutoSafeCoder, a multi-agent collaborative framework based on a large language model. It aims to improve the security of generated code through an iterative verification mechanism combining static analysis and dynamic testing. The framework's core design concept is to decompose the code generation process into three steps: function implementation, vulnerability detection, and dynamic verification. These steps are collaboratively completed by a coding agent, a static analyzer agent, and a fuzzing agent.

[0051] In the field of cyberspace security, AutoSafeCoder breaks down the network defense process into three key steps: policy generation, static verification, and dynamic adversarial testing. These steps are collaboratively completed by a policy generation agent (Policy Agent), a protocol analysis agent (Protocol Analyzer Agent), and a fuzzing agent (Fuzzing Agent). The policy generation agent generates an initial defense strategy based on the GPT-4 model, while the protocol analysis agent performs a formal security audit of the strategy by injecting the MITRE ATT&CK attack pattern knowledge base.

[0052] The collaborative mechanism design employs a phased iterative optimization strategy. The static analysis phase performs a maximum of four rounds of interaction, while the dynamic testing phase exhaustively explores attack path boundary conditions through 150 rounds of mutation cycles. Agent interactions rely on structured threat intelligence templates, and the strategy generation agent reconstructs defense logic based on remediation suggestion templates. A sandbox environment ensures the security of dynamic testing and isolates potentially malicious traffic.

[0053] The core innovation of this technology lies in its deep integration of LLM's semantic reasoning capabilities with network security attack and defense features. Through multi-agent collaborative knowledge sharing, it overcomes the bottleneck of delayed rule base updates. However, its defensive effectiveness is still limited by the static feature matching of attack patterns and the local optimization characteristics of policy generation. This still leaves blind spots in response and the risk of misjudgment when facing semantic-level APT attacks or zero-day vulnerability exploits.

[0054] This existing technology has the following shortcomings: Although the AutoSafeCoder framework reduces the vulnerability rate of defense strategies through a multi-agent collaboration mechanism, its technical architecture still has systemic flaws, making it difficult to cope with the dynamic and hidden challenges of high-level network attacks. This technology relies on a pre-defined MITRE ATT&CK attack pattern library for strategy generation and verification, resulting in the defense logic being solidified into static feature matching of known threats. For example, the protocol analysis agent uses rule templates to identify traditional attack features such as SQL injection or buffer overflow. However, when faced with attackers using semantic obfuscation techniques, the model misses detection due to its lack of deep semantic parsing capabilities for protocol payloads. More seriously, the variation and combination of attack features may bypass the syntax detection threshold of the rule library. For example, when a malicious payload is segmented into multiple compliant protocol fragments for transmission, the static analysis module cannot reconstruct the complete semantic context and misjudges it as legitimate traffic, exposing the inherent vulnerability of rule-driven defense.

[0055] The serial collaboration model among multiple agents leads to significant efficiency bottlenecks and policy fragmentation. The interaction between the policy generation agent and the protocol analysis agent is limited to four iterations. When faced with cross-protocol and cross-session APT attack chains, locally optimized defense rules cannot build a globally coordinated protection system. Furthermore, the dynamic verification capabilities of the fuzz testing agent are limited by the quality of the generated initial attack seeds. While its type-aware mutation strategy can simulate known attack patterns, it lacks coverage for zero-day vulnerability exploits or semantic-level logic flaws.

[0056] A deeper flaw stems from the disconnect between security attributes and defense decisions. The framework separates syntax verification, behavior simulation, and policy tuning into independent steps, failing to build a semantic risk map of attack intent and link it with the pragmatic value of business scenarios. The policy generation agent mechanically relies on historical attack data to generate rules, and fails to incorporate pragmatic elements such as asset value weights and real-time threat intelligence into the decision-making model, resulting in a conflict between defense actions and business continuity goals. For example, in highly sensitive scenarios such as financial transactions, the model may cause service degradation due to excessive interception of suspected attack traffic, while some low-frequency penetration behaviors can bypass detection because they conform to the grammatical features of the preset rule base. In addition, the black-box policy generation mechanism makes the defense logic untraceable. Security personnel cannot parse whether a specific interception action is based on protocol anomalies, semantic risks, or historical statistics, which seriously hinders the traceability analysis of attack links and the closed-loop tuning of defense strategies.

[0057] The root of these issues lies in the cognitive limitations of its technical approach. AutoSafeCoder views security as an external constraint on feature matching and rule combination, rather than as an inherent cognitive capability of the system. Its static knowledge base updates rely on manual intervention, and its defense strategies lack cross-domain semantic alignment and autonomous evolution mechanisms. Faced with rapidly evolving attack techniques, it remains stuck in a passive cycle of "rule lag—attack breakthrough." This traditional "divide and conquer" paradigm is fundamentally opposed to the multi-dimensional penetration characteristics of advanced attacks, making it difficult to support the dynamic defense requirements of high-security scenarios.

[0058] Therefore, based on the mechanism-based AI unified model, the present invention proposes a parsable anti-attack network security method and system to address the key technical bottlenecks existing in the cyberspace defense system, such as "incomplete attack identification, inaccurate response decisions, and uncontrollable strategy evolution". The method aims to achieve interpretable identification of attack intentions and collaborative optimization of defense paths through the structured integration of three layers of information: grammar, semantics, and pragmatics, and behavioral feedback-driven.

[0059] In response to the above-mentioned problems of the existing technology, the technical solution proposed in the present invention is a security response cognitive system with semantic drive as the core, which breaks through the fragmentation and static limitations of traditional defense architecture in information modeling and strategy evolution. Through the mapping of the entire information structure and the construction of the mechanismist interpretation path, the system of the present invention realizes the linkage identification of grammatical anomalies, semantic behaviors and pragmatic risks in the attack chain, and uses the response effect as a feedback signal to drive the continuous tuning of the strategy module. The system can form structured knowledge units based on historical response behaviors, and complete defense experience migration and model evolution in multi-node collaboration, promoting the transformation of the cyberspace security system from passive defense to cognitive-enhanced active security.

[0060] The preferred embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0061] like Figure 1 As shown, this embodiment provides a resolvable anti-attack network security method based on an AI unified model, including the following steps: Step S1: tracing the attack source and identifying the attack method of the input data stream; Step S2: Perform protocol structure analysis and grammatical element extraction on the input data stream at the grammar verification layer, start the grammar rule matching process, and obtain a grammar anomaly perception set; Step S3: At the semantic parsing layer, based on the grammatical anomaly perception set, the attack vector information is integrated, the semantic entities in the behavior chain are extracted to construct a semantic relationship graph, and the constructed semantic relationship graph is iteratively calculated to output a semantic risk vector; Step S4: In the pragmatic optimization layer, the semantic risk vector output by the semantic analysis layer is used as input. The defense target is modeled based on the current business scenario, resource constraints, and policy preferences. Through the policy planning and scheduling mechanism, an optimal and executable response path is output. In step S5, after detecting and responding to an attack, each edge node first encapsulates it into a structured response knowledge unit. Then, a structured desensitized knowledge vector is constructed through encoding and compression and encrypted and uploaded to the coordination center. Finally, after completing knowledge aggregation, the coordination center transmits feedback information back to each edge node to achieve synchronous updating of model parameters and policy rules, and adjusts the construction logic and scoring parameters of the semantic relationship graph to form a model evolution path based on local feedback and global collaboration.

[0062] The logical structure framework of this embodiment is as follows Figure 1 As shown in Figure 1, the entire system consists of an attack identification module, a three-layer full-information defense mechanism, a comprehensive decision-making module, and a knowledge closed-loop evolution mechanism. The three-layer full-information defense mechanism includes a grammatical verification layer, a semantic parsing layer, and a pragmatic optimization layer.

[0063] Before describing the detailed specific implementation methods, the meaning of each node is explained first: the source node refers to the node of the trajectory tree in step S101. The edge node refers to the edge defense node deployed at the network entrance, edge server or resource scheduling link in step S502. The boundary node refers to the boundary node of the semantic relationship graph in step S301, which is composed of the nodes of the structured grammatical parsing tree. The high-confidence intention node refers to the screening node in the semantic relationship graph that reaches the preset threshold. The preset threshold can be customized and adjusted according to actual needs. The semantic center node refers to the center node of the semantic relationship graph in step S301, which is composed of semantic entities.

[0064] Step S1 in this embodiment implements the attack identification phase, namely the attack identification module. As the pre-analysis module of the cyberspace defense system, the attack identification module consists of two submodules: attack source identification and analysis, and attack method identification and classification. Leveraging the cognitive drive capabilities of the mechanism-based artificial intelligence unified model, it undertakes the tasks of locating the source of cyberattacks and modeling their behavior, providing the foundation for subsequent processing.

[0065] Step S1 in this embodiment preferably includes the following sub-steps: Step S101, identifying and analyzing the attack source; Step S102: identifying and classifying attack methods; Step S103 is used to implement multidimensional feature modeling and output. The process is as follows: the attack source analysis results and attack method labels are encoded to generate a multidimensional feature vector of the attack behavior. The multidimensional feature vector includes source features, behavior features, communication structure features, and potential target directions. The feature modeling process of the multidimensional feature vector preferably adopts an adaptive vector projection algorithm. The algorithm first embeds and encodes the results extracted in steps S101 and S102, wherein discrete features are encoded using the word vector model FastText, and continuous features are mapped to the interval [0,1] using the Min-Max Normalization algorithm; Step S104 is used to generate an attack vector, and its process is as follows: encapsulating the result of the multidimensional feature vector encoding into an attack vector in a unified data format, and using it as the standard input of the syntax verification layer; the encapsulation process is to organize the various fields and corresponding encoded values ​​in the multidimensional feature vector according to the key-value pair structure of the JSON format to form a text string that conforms to the JSON syntax rules, so as to ensure that the data structure is clear and the semantics are clear, which facilitates efficient transmission and parsing between various modules of the system; wherein, the attack vector includes key fields such as the attack source address, method category, feature dimension code, attack behavior label and current connection context, wherein the current connection context specifically covers information such as the communication session identifier and the message structure summary, and has the characteristics of clear structure, clear semantics and universal interface; while encapsulating and generating the attack vector, a behavior identifier is assigned to the attack vector.

[0066] The output of the attack identification module plays a fundamental role in the entire system. It not only undertakes the initial analysis of the attack source and method, but also undertakes the responsibility of constructing the leading signal of the syntax-semantics-pragmatics three-dimensional link, ensuring information coupling and semantic consistency between modules.

[0067] The process of identifying and analyzing the attack source in step S101 of this embodiment is as follows: first, the real-time communication behavior is captured and preprocessed by the edge-deployed traffic monitoring module to extract metadata, which includes the source address, destination address, protocol type, port information, packet frequency, and connection duration. Then, a connection trajectory table is constructed with a five-tuple as the session identifier to achieve traceability. The sliding window and timestamp association mechanism is used to reconstruct the access path structure of a single source and multiple targets to form a trajectory tree. The five-tuple includes the source address, destination address, protocol, port, and timestamp. Next, a behavior deviation detection algorithm is used to compare and analyze the connection frequency, request direction change, and abnormal retry indicators, and path branches that may contain malicious behavior are marked from the trajectory tree. Finally, for each source node, its trust risk index is calculated based on its access pattern, historical tags, communication path complexity, and similarity with known attack IP addresses. When the trust risk index exceeds a preset dynamic threshold, the node is determined to be a potential attack source, and its communication context, path map, and identification information are extracted as input for subsequent classification and modeling.

[0068] The process of identifying and classifying attack methods in step S102 of this embodiment is as follows: based on the identification and analysis of the attack source, a pre-trained attack behavior classifier is called to identify the attack method of suspicious traffic; the process of identifying the attack method of suspicious traffic is represented by a fusion feature vector. First, grammatical anomaly features, time series features, and contextual features are used as input, and typical attack methods are matched and classified through a semantic weighting mechanism. The grammatical anomaly features include field out-of-bounds and protocol nesting anomalies, and the time series features include inter-packet delay and connection frequency; then, the attack pattern is matched and classified through a semantic weighting mechanism. The system supports accurate identification of common attack types such as DDoS, SQL injection, cross-site scripting, remote command execution, and port scanning. At the same time, it can continuously expand attack categories through a self-learning mechanism and improve the ability to identify unknown attack variants. The identification result of each attack method outputs a structured annotation, which is used as an attack behavior label; the attack behavior label includes information such as the attack method label, feature confidence, and impact path.

[0069] Contextual features refer to the environmental and behavioral background information of attack traffic at the time of the attack. This information can help more accurately identify the attack method. Specific features include, but are not limited to, historical interaction records of the current network session, the authentication status of both communicating parties, the correlation behavior patterns of previous and subsequent packets, protocol status information, and the network topology at the time of the attack.

[0070] It should be noted that the attack method label refers to a label used to describe the attack behavior at the operational level (technical implementation method), focusing on the attack's technical type or implementation method. The attack behavior label is a label that summarizes the overall intent or security goal of the attack behavior at the semantic level. It is used in the semantic risk assessment and policy generation stages, emphasizing the consequences and security risks of the behavior. The relationship between the two is as follows: the attack behavior label includes the attack method label. In step S102, the attack behavior label = attack method label + feature confidence + impact path. In other words, the attack behavior label is actually the output structured annotation. In this embodiment, step S103 uses the attack method label, while steps S104 and S201 use the attack behavior label.

[0071] Step S2 in this embodiment implements the syntax verification phase. The syntax verification layer is primarily responsible for comprehensively checking the protocol compliance, field structure integrity, and communication behavior formatting of the input data stream. This layer not only captures explicit violations at the rule level but also provides a structured risk focus for the semantic layer, thereby supporting the precise location of attack intent inference.

[0072] In this embodiment, step S2 preferably includes step S201 and step S202. Step S201 is used to implement protocol structure analysis and grammatical element extraction. Step S2 is used to implement multi-layer rule matching and abnormal grammar identification.

[0073] In step S201, the syntax verification layer first receives the attack vector output from the attack identification phase and extracts the connection context information and protocol structure fields contained therein as the raw input for syntax parsing. The attack vector encapsulates the communication session identifier, message structure summary, and attack behavior label, which constitute the initial representation of the communication behavior at the syntax verification layer. On this basis, the communication protocol stack used by the current traffic is automatically identified, and the data packets are layered decoded (Layered Decoding). The header fields, payload content, parameter structure and method call sequence in the message are systematically disassembled to form a structured grammatical parse tree. When forming the structured grammatical parse tree, by integrating the protocol specification knowledge base and attack behavior labels, key fields in the structured grammatical parse tree are marked with type identifiers, such as "identity authentication field", "URL path parameter", "command injection sensitive field", etc., which serve as pre-labels for anomaly annotation and risk weight calculation. Each node of the structured grammatical parse tree corresponds to a protocol field or grammatical fragment, which is used to annotate its field name, field type, occurrence position and nesting level. The nodes are connected by edges to represent the dependency order and combination structure, so as to construct and output a structured grammatical representation. The grammatical representation contains field sequence, structural hierarchy, protocol template ID and label information, which serves as direct input for subsequent grammatical rule matching and risk assessment. The structured grammar parse tree refers to SyntaxTree, a standard tree-like data structure used to express the grammatical hierarchy and composition of protocol messages or codes. By mapping protocol fields and their nested structures into tree nodes, the tree edges represent the dependencies and combinations between fields, providing an intuitive and structured foundation for grammar rule matching and anomaly detection. The construction process of the structured grammar representation is based on the preset protocol specification. By treating protocol fields as nodes, following their hierarchical relationships and parsing order, and using edges to represent the dependencies and combinations between nodes, a complete structured graphical representation is formed.

[0074] In step S202, after completing the protocol structure parsing, the grammar rule matching process is initiated based on the constructed grammar representation. The corresponding rule library adopts a three-layer architecture design, consisting of a basic protocol specification layer, a business scenario customization layer, and a dynamic abnormal behavior strategy layer. During the grammar rule matching process, the fields of each node in the structured grammar parse tree are matched and searched to determine whether they conform to the corresponding grammar template and call specifications. For example, in HTTP sessions, the system detects abnormal features such as illegal header splicing, excessive field nesting depth, or out-of-bounds parameter combinations. If a grammar field deviates from the rule template, the system calculates its anomaly index based on the degree of deviation, field sensitivity, and historical co-occurrence patterns. The anomaly index is calculated using a gradient boosting tree (GBT) model. This model, through training and learning of multi-dimensional features, can effectively capture complex nonlinear relationships and output an anomaly index. Finally, a list of field-level anomalies and their corresponding anomaly indices is output, forming a preliminary set of grammar anomaly perceptions to guide the subsequent focus selection of the semantic graph. Historical co-occurrence patterns refer to the association patterns between commonly observed abnormal fields or grammatical fragments discovered by the system through statistical analysis of large amounts of historical network traffic and attack events. Specifically, certain abnormal fields frequently co-occur in specific attack types or scenarios, forming a stable pattern. Leveraging this pattern, this embodiment can infer the degree of abnormality and potential risk based on the currently detected abnormal field combination. The introduction of historical co-occurrence patterns helps improve the accuracy of anomaly detection, reduce false positives, and enhance the ability to identify complex attack chains.

[0075] Step S3 described in this embodiment is used to implement the semantic parsing stage, and preferably includes steps S301 to S303.

[0076] Step S301 is used to extract semantic entities and construct an attack graph. First, the grammatical anomaly perception set from the grammatical validation layer is received. The attention feature fusion algorithm (AFF) is used to fuse the attack vector information from step S1 with the grammatical anomaly perception set to obtain a high-dimensional vector representation. Subsequently, the semantic entity extraction phase begins. This phase employs a multi-label classification model based on a multi-layer perceptron (MLP) and a sigmoid activation function to predict label probabilities for the high-dimensional vector representation, enabling automatic identification of semantic entities. These semantic entities include request initiators, action types, target objects, and scene annotations, organized into a "subject-action-object" structure. Finally, these semantic entities are mapped into a semantic relationship graph. This graph uses the nodes of the structured grammatical parse tree as boundary nodes and the semantic entities as semantic center nodes. Edge weights represent the strength of associations and the order of operations between semantic entities. This graph preserves the topological dependencies of the original grammatical structure while embedding the causal chain of the attack context. This semantic relationship graph serves as the primary input for subsequent graph neural network intent inference.

[0077] Step S302 is used to implement graph neural network model reasoning and attack intent identification. The semantic relationship graph is input into the Graph Neural Network (GNN) model. Using a multi-layer graph convolution mechanism, this model captures the high-level semantic associations hidden in the graph structure while preserving the original node attributes. The GNN model focuses on the following two types of semantic paths:

[0078] (a) For a serial operation path, such as the "login → upload → execute → delete" operation chain, the system checks whether it conforms to the normal business logic sequence and whether there are semantic anomalies such as sequence jumps or permission violations;

[0079] (b) Target-focused paths, where multiple independent operations are directed to the same highly sensitive resource, are used to determine whether there is structured intrusion behavior.

[0080] In step S302 of this embodiment, the semantic relationship graph is subjected to multiple rounds of iterative calculations through the graph neural network model. In the convolution process of each round of iterative calculation, the feature vector of each node is calculated using the formula: Update, where Indicates the Layer Time Node The eigenvector of is a node The neighbor set of For the Layer trainable weight matrices, From the adjacent node To Node The attention weight, is the activation function.

[0081] Through the aforementioned graph reasoning process (i.e., iterative computation), the system uses a multi-layer perceptron (MLP) to automatically identify high-confidence intent nodes in the semantic relationship graph, logically annotate their behavior paths, and output attack type labels, intent confidence scores, and intent paths, providing a semantic basis for subsequent risk quantification and strategy generation. The attack type label refers to the attack type prediction output from this step.

[0082] Step S303 is used to quantify and structure semantic risk. After identifying an attack intent, the system quantitatively assesses its potential security impact. The intent paths marked in the semantic relationship graph are scored. Scoring dimensions include the atypicality of the operation sequence (e.g., unconventional interface chain calls), the sensitivity level of the target involved (e.g., user authentication interface, operation and maintenance services), and the weight of upstream field anomalies (e.g., behavior chains triggered by severe grammatical anomalies). These scoring dimensions are normalized and, combined with the attack type label and intent confidence score, a vector concatenation method is used to align the scores of each scoring dimension with the label vector to form a fixed-length feature vector, which is then encapsulated into a semantic risk vector. The semantic risk vector is represented as a fixed-length structure. Its content includes the attack type, intent weight, impact range level, and reasoning confidence, and serves as the primary input to the pragmatic optimization layer decision model. The attack type label here refers to the attack type prediction result outputted in step S302. The label vector refers to the vector structure composed of the attack type prediction result outputted in step S302 and the intent confidence score.

[0083] Step S4 described in this embodiment is used to implement the pragmatic optimization stage. The pragmatic optimization layer, as the response decision module in the defense system of the present invention, undertakes the reasoning and execution tasks from "semantic risk" to "defense strategy". This stage uses the semantic risk vector output by the semantic parsing layer as input, combines the current business scenario, resource constraints and policy preferences, models the defense target, and outputs the optimal executable response path through the policy planning and scheduling mechanism. Step S4 (i.e., the pragmatic optimization stage) includes steps S401 to S403, which are respectively used for response target modeling and scenario adaptation, dynamic policy generation and optimization, and policy execution and defense deployment.

[0084] In step S401, the system first uses a pre-trained decision tree model to derive the most appropriate defense objectives and control boundaries based on the attack type, intent weight, impact level, and reasoning confidence contained in the semantic risk vector. During the reasoning process, the system feeds the semantic risk vector layer by layer into the decision tree structure, selecting a path based on each feature value and splitting condition, ultimately reaching a leaf node and outputting a result. This result includes the defense objective type (e.g., blocking, demotion, isolation), response level (e.g., strict, mild), and control boundaries (e.g., scope, duration, target component). The system maps risk events to specific business scenario labels, such as "high concurrent user login," "core data writing," or "low-risk background task," by accessing scenario configuration parameters in the operating environment.

[0085] Then, for the current business scenario, through the formula Construct the corresponding policy objective function ,in, Represents blocking capability, which is used to measure the efficiency of the strategy in cutting off key behavior points in the attack path; Indicates resource consumption, which is used to evaluate the degree of utilization of CPU, memory, IO and other resources during policy execution; Represents response latency, which measures the length of time it takes for a policy to be generated and deployed. 、 and The preset weight coefficients represent blocking capability, resource consumption, and response delay respectively; these preset weight coefficients follow the business priority strategy and can be set and adjusted according to actual needs. For example, in the financial payment scenario, blocking capability has a higher weight; while in the content service platform, response delay may become the main optimization target, and the system supports real-time adjustment of the weight configuration. For different business scenarios, by automatically building the corresponding strategy objective function , clarify the combination of indicators that need to be optimized during this response process, such as prioritizing availability, minimizing response time, or maximizing blocking efficiency, etc., which can comprehensively evaluate the actual value of the response strategy.

[0086] Step S402: First, according to the attack type and system constraints, a matching candidate policy set is retrieved from the policy knowledge base. The candidate policy set is represented as , including but not limited to: IP blocking, session isolation, permission downgrade, traffic speed limit, interface rollback and other operations. Represents the nth candidate policy, where n is a natural number. The attributes of each candidate policy include the policy score, applicable attack type tags (such as "SQL injection", "DDoS", and "command injection"), target module or interface scope (such as "authentication interface", "API entry", and "database write end"), operation cost metrics (including average execution time, CPU utilization, and I / O write frequency), policy conflict matrix (used to record whether the policy can be deployed in parallel with other policies), and historical response effect scores and negative impact records.

[0087] Then, by the formula Construct the subset of all legally combinable policies , using a strategy combination algorithm based on greedy search and conflict filtering to select candidate strategy sets The strategy items in the system are sorted based on the strategy score, and all strategy combinations that meet the system constraints are generated by gradually building feasible subsets. , and for each strategy subset combination Perform utility scoring. The comprehensive utility scoring function used for utility scoring is: ,in, represents the comprehensive utility score, Representation Strategy Scoring the coverage of key nodes in the attack path, Representation Strategy The average resource consumption of Representation Strategy Estimated response time.

[0088] In step S402 of this embodiment, greedy search and constraint pruning are combined to achieve the strategy subset combination. The filter follows the following conditions: Condition A1, candidate strategy set The covered attack nodes must include at least one high-confidence intention node in the semantic relationship graph; Condition A2: the same target module is not allowed to deploy strategies with mutually exclusive attributes; Condition A3: Response delay must not exceed the maximum acceptable value set by the system for the current business scenario ; Condition A4: If there are multiple strategy subset combinations If conditions A1 to A3 are met, select the comprehensive utility score The largest group is used as the current optimal strategy path , and generate the corresponding response instruction set.

[0089] In step S403, the response instruction set is sent to the security execution engine, which calls the corresponding network control interface, access control module or identity management subsystem according to the policy content to implement the actual deployment of the policy; for example, the blocking policy will call the firewall module to update the blocking rules, the speed limit policy will notify the network scheduling module to adjust the bandwidth allocation, and the permission rollback policy will synchronize the access control table and revoke the high-sensitivity interface call permission. During the policy execution process, the execution status, response delay, target blocking status and system feedback information are recorded in real time to form a complete response behavior log. This log is not only used for post-audit and policy effectiveness evaluation, but also serves as an important input for the subsequent knowledge feedback mechanism, supporting the system to form a positive closed loop in the effectiveness analysis and dynamic fine-tuning of the defense strategy.

[0090] Figure 4 This demonstrates the overall process of this embodiment during the knowledge evolution and feedback closed-loop phase. After detecting and responding to an attack, each edge node extracts and compresses a structured response knowledge unit (RKU summary). Using a feature distillation mechanism, it extracts multi-dimensional information, including syntax, semantics, and pragmatics, to generate a desensitized knowledge vector. This vector is encrypted and uploaded to the federated coordination center (also known as the coordinator). After completing knowledge aggregation, the coordination center (also known as the coordinator) transmits feedback information back to each node, enabling the synchronous updating of model parameters and policy rules, thus establishing a federated cognitive enhancement closed-loop.

[0091] Preferably, step S5 in this embodiment includes the following steps S501 to S503.

[0092] Step S501 is used for response result perception and knowledge injection. After the defense strategy is deployed, the system automatically enters the response behavior perception phase through the network security response and defense module, observing behavioral changes during the actual implementation of the strategy. The core task of this phase is to extract structured response behavior features based on the attack state changes after strategy execution and inject them into the system knowledge structure to optimize the subsequent intent recognition and strategy generation processes, thereby achieving the continuous cognitive capabilities of "interpretable behavior + evolvable strategy" emphasized by the mechanism-based AI unified model.

[0093] Specifically, after the policy is issued, the feedback status of the policy node is recorded in real time through the listener module, and the defense effect is judged based on the change of the attack chain status. The feedback status includes whether the defense instruction is successfully triggered, whether the attack chain has an intervention effect, and whether the security status of key resources in the system has changed; the observation information corresponding to the feedback status is encapsulated as a response knowledge unit (RKU) and injected back into the semantic parsing layer and the policy scoring module. Among them, the comprehensive utility scoring function Stored in the strategy scoring module.

[0094] The system judges whether the strategy is effective not only by relying on the prompts of the execution results, but also by judging the defense effect based on the changes in the state of the attack chain. For example, if there was originally a complete "login → upload → execution" path in the semantic relationship graph, and after the strategy is triggered, the "execution" behavior no longer occurs, or its triggering frequency decreases significantly, the system will determine that the strategy has substantially interfered with the attack chain. On the contrary, if the attack path is not interrupted and only a small disturbance occurs near the strategy node, the system will mark the strategy as "partial failure" or "invalid". In order to make this feedback available for system updates, this embodiment encapsulates the above-mentioned behavior observation information into a Response Knowledge Unit (RKU). RKU is a structured feedback record carrier, and its fields include: (a) Policy identifier (indicating the number and type of policy being executed); (b) Target path node (target position in the attack graph corresponding to the marking strategy); (c) Execution status (success / failure / partial success); (d) indicators of changes in attack behavior (e.g., whether specific actions have been terminated or whether the frequency of calls has decreased); (e) system side effect indicators (e.g., service response time delay, connection number fluctuation); (f) Confidence weight adjustment suggestions (instructions to increase or decrease the risk weight of relevant nodes in the semantic graph).

[0095] These response knowledge units do not directly affect the strategy selection decision, but are injected back into two key positions in the system model: first, the semantic parsing layer, the nodes marked as "intervention success" in RKU will be used as training samples for the next round of intent recognition, improving the model's recognition sensitivity to similar attack paths; and the nodes corresponding to "invalid strategy" will enhance their risk scores to avoid being underestimated; second, the strategy scoring module, the "strategy-behavior chain intervention relationship" recorded in RKU is converted into a sample of strategy coverage score, which is used to optimize the parameter estimation logic in the strategy scoring function and eliminate strategy paths with poor historical performance.

[0096] Step S502 is used to implement cross-domain collaborative defense. In this embodiment, the node refers to an edge defense node deployed at the network entrance, edge server, or resource scheduling link. Each edge node has a basic semantic parsing layer and policy scoring module for local traffic intent identification, risk quantification, and response strategy selection. It can be flexibly configured based on computing resources and supports module-level model fine-tuning and parameter updates.

[0097] During cross-domain knowledge collaboration, edge nodes periodically extract representative samples from the response knowledge units as federated shared information. To prevent data leakage, the key information of the representative samples (attack type, policy intervention results, and key behavioral characteristics) is first encoded using the Sentence-BERT model. Principal component analysis (PCA) is then used to compress the information and construct a structured, desensitized knowledge vector. For example, a triple representing information such as "a command execution attack was successfully interrupted by an IP blocking policy at this node" is converted into a low-dimensional semantic vector and uploaded to the coordination center via an encrypted channel as shared information.

[0098] The coordination center does not perform unified modeling. Instead, it clusters and statistically analyzes feedback from multiple nodes, extracting aggregated evaluation results of the effectiveness of strategies under different attack labels. Outputs include: a certain strategy has a high success rate in most nodes, a certain type of attack frequently escapes a certain type of service node, and so on. This aggregated feedback is then distributed to each edge node, driving lightweight, adaptive updates to local models.

[0099] After receiving the feedback evaluation results, each edge node starts the update process. The update process includes: the semantic parsing layer expands the local semantic relationship graph and updates the node type dictionary based on the feedback of new attack intentions or path structures, thereby improving the pre-identification capability of new attacks; the strategy scoring module adjusts the weight coefficients and priority sorting rules in the comprehensive utility scoring function to make it closer to the strategy combination that performs well in the collaborative network, thereby improving the response effect and strategy hit rate.

[0100] Step S503 implements federated knowledge evolution. With the continuous iteration of edge node feedback and local model updates, the system gradually establishes a knowledge evolution mechanism driven by behavioral observation (also known as behavioral feedback). In this mechanism, each strategy generation and deployment is no longer just a single defensive response to the current attack, but rather an input into the system's cognitive structure. The system continuously refines its understanding of attack intent and strategy selection preferences through real-world feedback on strategy effectiveness.

[0101] Specifically, when a strategy subset combination If a strategy demonstrates stable disruption effectiveness over a preset number of deployments, its fit weight with its corresponding attack intent will be automatically strengthened, allowing similar strategy paths to be more quickly identified in subsequent judgments. Conversely, if a strategy consistently fails to intervene or has significant side effects on different paths, its overall utility score will be lowered and it will be removed from the candidate strategy set within a preset time window. The preset number of times refers to a pre-set threshold, and the preset time window refers to a fixed, pre-set time window, both of which can be adjusted based on actual conditions and needs.

[0102] Therefore, this embodiment does not rely on external training sets or offline learning processes, but is completely based on the internal feedback chain of the system's autonomous observation, autonomous judgment, and autonomous tuning.

[0103] This embodiment further provides a resolvable anti-attack network security system based on an AI unified model, which adopts the resolvable anti-attack network security method based on an AI unified model as described above, and includes: Attack identification module, used to trace the attack source and identify the attack method of the input data stream; The grammar verification module parses the protocol structure and extracts grammatical elements from the input data stream at the grammar verification layer, starts the grammar rule matching process, and obtains the grammar anomaly perception set; The semantic parsing module, based on the grammatical anomaly perception set at the semantic parsing layer, integrates attack vector information, extracts semantic entities in the behavior chain to construct a semantic relationship graph, and iteratively calculates the constructed semantic relationship graph to output a semantic risk vector. The pragmatic optimization module, in the pragmatic optimization layer, uses the semantic risk vector output by the semantic analysis layer as input. It combines the current business scenario, resource constraints, and policy preferences to model the defense target and output an executable optimal response path through a policy planning and scheduling mechanism. In the knowledge evolution and closed-loop feedback module, after detecting and responding to attacks, each edge node is first encapsulated into a structured response knowledge unit. Then, a structured desensitized knowledge vector is constructed through encoding and compression and encrypted and uploaded to the coordination center. Finally, after completing knowledge aggregation, the coordination center transmits feedback information back to each edge node to achieve synchronous updating of model parameters and policy rules, and adjust the construction logic and scoring parameters of the semantic relationship graph to form a model evolution path based on local feedback and global collaboration.

[0104] In summary, this embodiment proposes an active cognitive network security method system that can parse and resist attacks based on the AI ​​unified model. By reconstructing the three-layer full information parsing mechanism of "syntax-semantics-pragmatics" and behavioral feedback drive, it realizes the integrated fusion of attack intention identification, response strategy generation and feedback-driven evolution; compared with the static configuration, rule separation and response isolation of existing security policies, this embodiment transforms the "defense logic" from external configuration to the system's endogenous cognitive ability, and constructs an explainable and evolvable defense closed loop in the entire process of attack identification, path parsing and policy deployment; attack identification is comprehensive, response decisions are accurate and policy evolution is controllable.

[0105] The three-layer full-information analysis mechanism refers to a three-layer structural analysis process driven by attack vectors. Specifically, it includes: 1. The syntax verification layer explicitly identifies field anomalies and format deviations through protocol structure analysis and rule matching verification; 2. The semantic analysis layer uses a semantic relationship graph as a structure to identify attack intent and path patterns through a graph neural network; 3. The pragmatic optimization layer outputs a dynamic optimal strategy path based on a multi-objective response objective function. This three-dimensional collaborative structure significantly differs from traditional passive detection systems that rely solely on rule or template recognition and constitutes the core analysis framework of this embodiment.

[0106] More specifically, first, this embodiment can fundamentally improve the accuracy of attack identification and the resilience of defense responses. Existing network defense mechanisms rely on static rule matching and behavioral template recognition, which makes it difficult to cope with complex threats such as dynamic evolution of attack paths, semantic disguise, and cross-layer bypassing. This embodiment extracts abnormal features of field structure and protocol behavior through the syntax verification layer, and the semantic parsing layer builds a semantic graph based on the attack vector and identifies high-risk intent paths. The pragmatic optimization layer generates differentiated response strategies based on dimensions such as target sensitivity, resource status, and attack confidence. Taking cross-domain command injection as an example, it not only identifies field combination anomalies, but also traces the semantic causal relationship of key nodes in the behavioral path, and comprehensively evaluates the potential impact range of the behavioral chain, thereby achieving precise interception with visible structure and adjustable strategy, and possessing a complete closed-loop capability from anomaly perception to path blocking, significantly improving response timeliness and interception success rate.

[0107] This embodiment generates standardized attack vectors as a unified input interface for subsequent processing modules, integrating grammatical structure analysis, semantic behavior modeling, and pragmatic strategy generation, achieving full decoupling of input format and intermediate information. Attack vectors encapsulate key fields such as attack source characteristics, behavior labels, context descriptions, and communication structures, providing a universal interface for cross-module linkage and cross-domain sharing.

[0108] Second, it achieves enhanced explainability of policy paths and adaptive evolution of scoring mechanisms. Traditional defense systems typically make decisions based on black-box models, resulting in opaque defense paths and difficult-to-audit policy rationales, severely limiting their usability in high-trust scenarios. This embodiment, based on a unified AI model, annotates behavioral nodes and edge weight strengths in a semantic graph. It then uses the utility function of the pragmatic optimization layer to perform a multi-objective comprehensive scoring of policy combinations based on blocking capabilities, resource consumption, and response latency. The entire policy selection process is traceable and reproducible. The system supports the visualization of grammatical anomaly graphs, intent recognition results, and policy scoring factors, assisting security experts in policy review and parameter adjustment. Response execution results are encapsulated as response knowledge units (RKUs) for back-injection into the semantic parsing module and policy scoring module, improving the adaptability of the model structure and enabling progressive optimization of policy logic.

[0109] This embodiment uses a structured representation of response results—the Response Knowledge Unit (RKU)—to record the actual impact path, behavioral changes, and system side effect indicators of policy execution. The RKU enables reverse feedback from policy effects on the semantic graph construction logic and policy scoring functions, establishing an injectable cognitive update chain. This is a key mechanism supporting the system's evolutionary closed-loop capabilities.

[0110] Third, the system achieves distributed expansion of its collaborative defense capabilities and continuous enhancement of its knowledge update mechanism, supporting the construction of a coordinated defense system across multiple edge nodes. Edge nodes make autonomous defense decisions based on local semantic analysis and policy scoring, and periodically upload desensitized and compressed RKU summaries to the federated coordination node (i.e., the coordination center). The coordination center performs cluster analysis on multi-source feedback and distributes policy effectiveness reports back to each node. Based on this information, each edge node fine-tunes its semantic graph construction logic and scoring parameters, forming a model evolution path based on local feedback and global collaboration. This mechanism, while ensuring node autonomy, enables the horizontal transfer of defense knowledge and the dynamic unification of policy evaluation criteria, significantly enhancing the system's overall resilience to unknown attacks and evolving paths.

[0111] In a multi-node deployment scenario, this embodiment supports edge nodes uploading desensitized and compressed RKU summaries, which are aggregated and generated as feedback reports by the coordination node. Each node autonomously fine-tunes the semantic recognition template and strategy utility scoring factor in the local model based on the feedback, realizing the horizontal migration of knowledge and the dynamic update of local models, and building a collaborative cognitive network with continuous evolution capabilities.

[0112] Fourth, it has strong versatility and a solid foundation for building a cognitive security ecosystem. It has good scalability and universal adaptability, does not rely on specific protocol templates or attack labels, and can be widely used in high-security scenarios such as government and enterprise boundary protection, industrial control system intrusion detection, and IoT device access verification. It supports the configuration of high-level defense targets through natural language, automatically mapping them to grammatical rules, semantic graph structures, and pragmatic weight distribution, and generating executable and interpretable response instruction sets. At the same time, new attack paths that appear in multiple rounds of responses can be injected back to the expert configuration end, supporting the knowledge evolution path jointly built by humans and machines, and promoting the continuous evolution of the cognitive security ecosystem of "local perception-global optimization".

[0113] Unlike traditional security solutions that rely on empirical templates or black-box algorithms, this embodiment employs a mechanism-based approach to construct a structurally interpretable reasoning path. This includes explicit structural modeling of semantic graphs, a utility scoring function for policy combinations, and a feedback injection mechanism for model updates. These mechanisms all possess a clear decision-making foundation and behavioral logic, forming a complete closed loop of "identification-reasoning-response-feedback," ensuring verifiability and auditability in high-security scenarios.

[0114] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A parseable anti-attack network security method based on an AI unified model, characterized in that: The following steps are involved: Step S1: tracing the attack source and identifying the attack method of the input data stream; Step S2: Perform protocol structure analysis and grammatical element extraction on the input data stream at the grammar verification layer, start the grammar rule matching process, and obtain a grammar anomaly perception set; Step S3: At the semantic parsing layer, based on the grammatical anomaly perception set, the attack vector information is integrated, the semantic entities in the behavior chain are extracted to construct a semantic relationship graph, and the constructed semantic relationship graph is iteratively calculated to output a semantic risk vector; Step S4: In the pragmatic optimization layer, the semantic risk vector output by the semantic analysis layer is used as input. The defense target is modeled based on the current business scenario, resource constraints, and policy preferences. Through the policy planning and scheduling mechanism, an optimal and executable response path is output. In step S5, after detecting and responding to an attack, each edge node first encapsulates it into a structured response knowledge unit. Then, a structured desensitized knowledge vector is constructed through encoding and compression and encrypted and uploaded to the coordination center. Finally, after completing knowledge aggregation, the coordination center transmits feedback information back to each edge node to achieve synchronous updating of model parameters and policy rules, and adjusts the construction logic and scoring parameters of the semantic relationship graph to form a model evolution path based on local feedback and global collaboration.

2. The analyzable anti-attack network security method based on the AI ​​unified model according to claim 1 is characterized in that: The step S1 includes the following sub-steps: Step S101, identifying and analyzing the attack source; Step S102: identifying and classifying attack methods; Step S103: Encode the attack source analysis results and the attack method labels to generate a multi-dimensional feature vector of the attack behavior. The multi-dimensional feature vector includes source features, behavior features, communication structure features, and potential target orientation. The feature modeling process of the multi-dimensional feature vector adopts an adaptive vector projection algorithm. Step S104: encapsulate the multidimensional attack feature set corresponding to the multidimensional feature vector into an attack vector in a unified format as the standard input of the syntax verification layer; wherein the attack vector includes the attack source address, method category, feature dimension code, attack behavior label and current connection context; while encapsulating and generating the attack vector, a behavior identifier is assigned to the attack vector.

3. The analyzable anti-attack network security method based on the AI ​​unified model according to claim 2 is characterized in that: The process of identifying and analyzing the attack source in step S101 is as follows: first, the traffic monitoring module deployed at the edge captures and pre-processes real-time communication behavior to extract metadata, including source address, destination address, protocol type, port information, packet frequency, and connection duration; Then, a connection trajectory table with a five-tuple as the session identifier is constructed, and a sliding window and timestamp association mechanism is used to reconstruct the access path structure of a single source and multiple targets to form a trajectory tree. The five-tuple includes the source address, destination address, protocol, port, and timestamp. Next, a behavior deviation detection algorithm is used to compare and analyze the connection frequency, request direction change, and abnormal retry indicators, marking the path branches that may contain malicious behavior from the trajectory tree. Finally, for each source node, its trust risk index is calculated based on its access pattern, historical tags, communication path complexity, and similarity with known attack IP addresses. When the trust risk index exceeds a preset dynamic threshold, the node is identified as a potential attack source, and its communication context, path map, and identification information are extracted as input for subsequent classification and modeling. The process of identifying and classifying attack methods in step S102 is as follows: based on the identification and analysis of the attack source, a pre-trained attack behavior classifier is called to identify the attack method of suspicious traffic; the process of identifying the attack method of suspicious traffic is represented by a fusion feature vector. First, grammatical anomaly features, time series features, and contextual features are used as input, and the attack methods are matched and classified through a semantic weighting mechanism. The grammatical anomaly features include field out-of-bounds and protocol nesting anomalies, and the time series features include inter-packet delay and connection frequency; then, the attack patterns are matched and classified through a semantic weighting mechanism, and the identification results of each attack method are output as a structured annotation, which is used as an attack behavior label; the attack behavior label includes an attack method label, feature confidence, and impact path.

4. The analyzable anti-attack network security method based on an AI unified model according to any one of claims 1 to 3, characterized in that: The step S2 includes the following sub-steps: Step S201 automatically identifies the communication protocol stack used by the current traffic, performs layered decoding on the data packet, and systematically decomposes the header field, payload content, parameter structure, and method call sequence in the message to form a structured grammar parse tree. When forming the structured grammar parse tree, by integrating the protocol specification knowledge base and attack behavior labels, key fields in the structured grammar parse tree are marked with type identifiers, which serve as pre-labels for anomaly annotation and risk weight calculation. Each node of the structured grammar parse tree corresponds to a protocol field or grammar fragment, and is used to annotate its field name, field type, occurrence position, and nesting level. The nodes are connected by edges to represent the dependency order and combination structure, so as to construct a structured grammar representation body, which includes field sequence, structure level, protocol template ID, and label information. In step S202, based on the constructed grammatical representation, the grammatical rule matching process is started. The rule library corresponding to the grammatical rule matching process adopts a three-layer architecture design, which includes a basic protocol specification layer, a business scenario customization layer, and a dynamic abnormal behavior strategy layer. In the grammatical rule matching process, the fields of each node in the structured grammatical parsing tree are matched and retrieved to determine whether they comply with the corresponding grammatical template and calling specifications, and a field-level exception list and its corresponding risk indicators are output to form a grammatical anomaly perception set.

5. The analyzable anti-attack network security method based on an AI unified model according to any one of claims 1 to 3, characterized in that: The step S3 includes the following sub-steps: Step S301: Based on the grammatical anomaly perception set, attack vector information is integrated to extract semantic entities in the behavior chain and map the semantic entities into a semantic relationship graph. The semantic entities include the request initiator, action type, target object, and scenario annotation. The semantic relationship graph uses the nodes of the structured grammatical parse tree as boundary nodes and the semantic entities as semantic center nodes. Edge weights are used to represent the association strength and operation sequence between semantic entities. Step S302: Input the semantic relationship graph into a graph neural network model, use a multi-layer graph convolution mechanism to identify the attack intent of the semantic relationship graph, and logically annotate its behavior path to obtain the intent path; Step S303: Score the intention paths marked in the semantic relationship graph. The scoring dimensions include the atypicality of the operation sequence, the sensitivity level of the target involved, and the abnormal weight of the upstream field. The scoring dimensions are normalized, and combined with the attack type label and the behavior sequence confidence, a vector splicing method is used to align the score of each scoring dimension with the label vector to form a fixed-length feature vector, which is then encapsulated into the final semantic risk vector. The semantic risk vector is represented by a fixed-length structure. The content of the semantic risk vector includes the attack type, intention weight, impact range level, and reasoning confidence.

6. The analyzable anti-attack network security method based on the AI ​​unified model according to claim 5 is characterized in that: In step S302, the semantic relationship graph is iterated multiple times by using the graph neural network model. In the convolution process of each round of iterative calculation, the feature vector of each node is calculated using the formula: Update, where Indicates the Layer Time Node The eigenvector of is a node The neighbor set of For the Layer trainable weight matrices, From the adjacent node To Node The attention weight, is the activation function.

7. The analyzable anti-attack network security method based on an AI unified model according to any one of claims 1 to 3, characterized in that: The step S4 includes the following sub-steps: Step S401: First, according to the attack type, intention weight, impact range level and reasoning confidence contained in the semantic risk vector, the corresponding defense target and control boundary are derived; then, for the current business scenario, the formula Construct the corresponding policy objective function ,in, Indicates blocking capability, Indicates resource consumption, Indicates response delay, 、 and Preset weight coefficients representing blocking capability, resource consumption, and response delay respectively; Step S402: First, according to the attack type and system constraints, a matching candidate policy set is retrieved from the policy knowledge base. The candidate policy set is represented as ,in, Represents the nth candidate strategy, where n is a natural number. The attributes of each candidate strategy include the applicable attack type label, target module or interface application scope, operation cost measurement value, strategy conflict matrix, and historical response effect score and negative impact record; then, through the formula Construct the subset of all legally combinable policies , obtain strategy subset combinations through search and filtering , and for each strategy subset combination Perform utility scoring. The comprehensive utility scoring function used for utility scoring is: ,in, represents the comprehensive utility score, Representation Strategy Scoring the coverage of key nodes in the attack path, Representation Strategy The average resource consumption of Representation Strategy Estimated response time; In step S403, the response instruction set is sent to the security execution engine, which calls the corresponding network control interface, access control module or identity management subsystem according to the policy content to implement the actual deployment of the policy. During the policy execution process, the execution status, response delay, target blocking status and system feedback information are recorded in real time to form a complete response behavior log.

8. The analyzable anti-attack network security method based on the AI ​​unified model according to claim 7 is characterized in that: In step S402, greedy search and constraint pruning are combined to achieve the combination of strategy subsets. The filter follows the following conditions: Condition A1, candidate strategy set The covered attack nodes must include at least one high-confidence intention node in the semantic relationship graph; Condition A2: the same target module is not allowed to deploy strategies with mutually exclusive attributes; Condition A3: Response delay must not exceed the maximum acceptable value set by the system for the current business scenario ; Condition A4: If there are multiple strategy subset combinations If conditions A1 to A3 are met, select the comprehensive utility score The largest group is used as the current optimal strategy path , and generate the corresponding response instruction set.

9. The analyzable anti-attack network security method based on the AI ​​unified model according to claim 7 is characterized in that: The step S5 includes the following sub-steps: Step S501: The listener module records the feedback status of the policy node in real time and determines the defense effectiveness based on the changes in the attack chain status. The feedback status includes whether the defense instruction is successfully triggered, whether the attack chain is intervened, and whether the security status of key resources in the system has changed. The observation information corresponding to the feedback status is encapsulated into a response knowledge unit and injected back into the semantic parsing layer and the strategy scoring module. The fields of the response knowledge unit include the strategy identifier, target path node, execution status, attack behavior change indicator, system side effect indicator, and confidence weight adjustment suggestion; Step S502: Implementing cross-domain collaborative defense: Edge nodes periodically extract representative samples from the response knowledge units, encode and compress the attack types, policy intervention results, and key behavioral characteristics of the representative samples, construct structured desensitized knowledge vectors, and upload them to the coordination center via an encrypted channel as shared information. The coordination center clusters and statistically analyzes feedback from multiple nodes, extracting summary evaluation results of policy effectiveness under different attack labels. After receiving the feedback evaluation results, each edge node starts updating. The update process includes: the semantic parsing layer expands the local semantic relationship graph and updates the node type dictionary based on the feedback new attack intent or path structure; The strategy scoring module adjusts the weight coefficients and priority sorting rules in the comprehensive utility scoring function; Step S503: When a certain strategy subset combination If a strategy shows a stable interruption effect in a preset number of deployments, the adaptation weight between the strategy and its corresponding semantic node will be automatically strengthened; if a strategy continuously shows intervention failure or large side effects on different paths, its comprehensive utility score will be reduced and it will be removed from the candidate strategy set within the preset time window.

10. A parseable anti-attack network security system based on an AI unified model, characterized by: The method for resolvable anti-attack network security based on an AI unified model according to any one of claims 1 to 9 is adopted, and includes: Attack identification module, used to trace the attack source and identify the attack method of the input data stream; The grammar verification module parses the protocol structure and extracts grammatical elements from the input data stream at the grammar verification layer, starts the grammar rule matching process, and obtains the grammar anomaly perception set; The semantic parsing module, based on the grammatical anomaly perception set at the semantic parsing layer, integrates attack vector information, extracts semantic entities in the behavior chain to construct a semantic relationship graph, and iteratively calculates the constructed semantic relationship graph to output a semantic risk vector. The pragmatic optimization module, in the pragmatic optimization layer, uses the semantic risk vector output by the semantic analysis layer as input. It combines the current business scenario, resource constraints, and policy preferences to model the defense target and output an executable optimal response path through a policy planning and scheduling mechanism. In the knowledge evolution and closed-loop feedback module, after detecting and responding to attacks, each edge node is first encapsulated into a structured response knowledge unit. Then, a structured desensitized knowledge vector is constructed through encoding and compression and encrypted and uploaded to the coordination center. Finally, after completing knowledge aggregation, the coordination center transmits feedback information back to each edge node to achieve synchronous updating of model parameters and policy rules, and adjust the construction logic and scoring parameters of the semantic relationship graph to form a model evolution path based on local feedback and global collaboration.

Images