Computer security protection system based on artificial intelligence

Abstract

The invention relates to a computer security protection system based on artificial intelligence, which comprises a data acquisition layer, a data processing and feature engineering layer, an AI model layer, a real-time detection and response layer, a feedback and self-adaption layer and a management visualization layer. According to the method, multiple learning paradigms are fused, the problems of model outdated, attack bypassing and alarm fatigue can be solved, the robustness and the optimization accuracy can be improved, a decision-making basis is generated by adopting tools such as SHAP, the interpretability is improved, man-machine cooperation is achieved, key decisions are reserved for manual auditing, and the safety is improved.

Description

Technical Field

[0001] This invention relates to a computer security protection system based on artificial intelligence. Background Technology

[0002] While existing AI-based computer security systems excel in threat detection, response speed, and adaptability, supervised learning relies on large amounts of high-quality labeled data, which is costly. Furthermore, novel attacks lack contextual data, training data is inconsistent with the actual environment, model robustness is flawed, interpretability and credibility are insufficient, system coordination is flawed, and single points of failure pose a risk. Summary of the Invention

[0003] The technical problem to be solved by the present invention is: in order to overcome the above-mentioned technical problems, the present invention provides a computer security protection system based on artificial intelligence.

[0004] The technical solution adopted by the present invention to solve its technical problem is: a computer security protection system based on artificial intelligence, including a data acquisition layer, a data processing and feature engineering layer, an AI model layer, a real-time detection and response layer, a feedback and adaptation layer, and a management visualization layer; In the data acquisition layer, data sources include network traffic, terminal behavior, system logs, external intelligence, and user behavior; The network traffic data includes firewall logs, NetFlow, and packet captures; The terminal behavior data includes process monitoring, file operations, and registry changes; The system log data includes operating system logs and application logs; The external intelligence data includes threat intelligence platforms and building databases; The user behavior data includes login mode, permission usage, and abnormal operations; In the data processing and feature engineering layer, the data processing steps include data normalization, feature extraction, and dimensionality reduction and encoding; Feature extraction targets include network layer features, host layer features, and user layer features; Network layer characteristics include traffic frequency, protocol distribution, and connection duration; Host-level characteristics include abnormal memory usage and changes in file hashes; User-level characteristics include login geographic location and operation time sequence; The AI ​​model layer includes the core model and model training; The core models include supervised learning, unsupervised learning, reinforcement learning, and graph neural networks; Supervised learning is used for malware detection and intrusion detection. Unsupervised learning functions include anomaly detection and cluster analysis. The function of reinforcement learning is to dynamically adjust defense strategies; The function of graph neural networks is to analyze the attacker's lateral movement path; The real-time monitoring and response layer includes a dynamic detection engine, a response module, and interpretive tools; The feedback and adaptive layer includes online learning mechanisms, adversarial training mechanisms, and false alarm analysis mechanisms; The management and visualization layer includes dashboards, alarm systems, and audit logs.

[0005] As a preferred approach, supervised learning includes classification algorithms and regression algorithms. Classification algorithms include logistic regression, decision trees, random forests, support vector machines, and neural networks, while regression algorithms include linear regression and time series models.

[0006] As a preferred approach, the unsupervised learning process includes input data, model training, and output results. Unsupervised learning algorithms include clustering algorithms, anomaly detection algorithms, and dimensionality reduction algorithms.

[0007] As preferred methods, reinforcement learning algorithms include value-based, policy-based, actor-critic, and multi-agent RL.

[0008] Preferably, graph neural networks include graph convolutional network models, graph attention network models, graph autoencoder models, spatiotemporal graph network models, and heterogeneous graph neural network models.

[0009] Preferably, the dynamic detection engine includes a data acquisition layer, a behavior analysis layer, a threat determination layer, a response layer, and a feedback optimization layer. The dynamic analysis methods used by the dynamic detection engine include sandbox execution, sequence modeling, graph behavior analysis, and differential analysis.

[0010] As a preferred embodiment, the response module includes a decision-making layer, an execution layer, an isolation layer, a repair layer, and a source tracing layer.

[0011] As a preferred approach, the core explanatory methods of explanatory tools include model-inherent explanatory and ex-post explanatory techniques. Inherent models include decision trees, linear regression, and rule engines, while ex-post explanatory techniques include LIME, SHAP, attention mechanisms, and adversarial example analysis.

[0012] Preferred methods for online learning mechanisms include incremental learning algorithms, streaming data processing, and model version control; techniques for adversarial training mechanisms include adversarial example generation methods and robust optimization objectives; and techniques for false alarm analysis mechanisms include false alarm source tracing and automated optimization strategies.

[0013] Preferred incremental learning algorithms include stochastic gradient descent and Bayesian update; adversarial example generation methods include fast gradient signification, projective gradient descent, and generative adversarial networks; false alarm attribution methods include feature attribution analysis and rule conflict detection; and automated optimization strategies include active learning and dynamic threshold adjustment.

[0014] The beneficial effects of this invention are that the artificial intelligence-based computer security protection system combines supervised learning, unsupervised learning, and reinforcement learning, integrating multiple learning paradigms. It can solve problems such as model obsolescence, attack bypass, and alarm fatigue, improve robustness, optimize accuracy, generate decision-making basis using tools such as SHAP, improve interpretability, and achieve human-machine collaboration. Key decisions retain human review, thereby improving security. Detailed Implementation

[0015] An artificial intelligence-based computer security protection system includes a data acquisition layer, a data processing and feature engineering layer, an AI model layer, a real-time detection and response layer, a feedback and adaptation layer, and a management visualization layer. In the data acquisition layer, data sources include network traffic, terminal behavior, system logs, external intelligence, and user behavior; The network traffic data includes firewall logs, NetFlow, and packet captures; The terminal behavior data includes process monitoring, file operations, and registry changes; The system log data includes operating system logs and application logs; The external intelligence data includes threat intelligence platforms and building databases; The user behavior data includes login mode, permission usage, and abnormal operations; In the data processing and feature engineering layer, the data processing steps include data normalization, feature extraction, and dimensionality reduction and encoding; Feature extraction targets include network layer features, host layer features, and user layer features; Network layer characteristics include traffic frequency, protocol distribution, and connection duration; Host-level characteristics include abnormal memory usage and changes in file hashes; User-level characteristics include login geographic location and operation time sequence; The AI ​​model layer includes the core model and model training; The core models include supervised learning, unsupervised learning, reinforcement learning, and graph neural networks; Supervised learning is used for malware detection and intrusion detection. Unsupervised learning functions include anomaly detection and cluster analysis. The function of reinforcement learning is to dynamically adjust defense strategies; The function of graph neural networks is to analyze the attacker's lateral movement path; The real-time monitoring and response layer includes a dynamic detection engine, a response module, and interpretive tools; The feedback and adaptive layer includes online learning mechanisms, adversarial training mechanisms, and false alarm analysis mechanisms; The management and visualization layer includes dashboards, alarm systems, and audit logs.

[0016] As a preferred approach, supervised learning includes classification algorithms and regression algorithms. Classification algorithms include logistic regression, decision trees, random forests, support vector machines, and neural networks, while regression algorithms include linear regression and time series models.

[0017] As a preferred approach, the unsupervised learning process includes input data, model training, and output results. Unsupervised learning algorithms include clustering algorithms, anomaly detection algorithms, and dimensionality reduction algorithms.

[0018] As preferred methods, reinforcement learning algorithms include value-based, policy-based, actor-critic, and multi-agent RL.

[0019] Preferably, graph neural networks include graph convolutional network models, graph attention network models, graph autoencoder models, spatiotemporal graph network models, and heterogeneous graph neural network models.

[0020] Preferably, the dynamic detection engine includes a data acquisition layer, a behavior analysis layer, a threat determination layer, a response layer, and a feedback optimization layer. The dynamic analysis methods used by the dynamic detection engine include sandbox execution, sequence modeling, graph behavior analysis, and differential analysis.

[0021] As a preferred embodiment, the response module includes a decision-making layer, an execution layer, an isolation layer, a repair layer, and a source tracing layer.

[0022] As a preferred approach, the core explanatory methods of explanatory tools include model-inherent explanatory and ex-post explanatory techniques. Inherent models include decision trees, linear regression, and rule engines, while ex-post explanatory techniques include LIME, SHAP, attention mechanisms, and adversarial example analysis.

[0023] Preferred methods for online learning mechanisms include incremental learning algorithms, streaming data processing, and model version control; techniques for adversarial training mechanisms include adversarial example generation methods and robust optimization objectives; and techniques for false alarm analysis mechanisms include false alarm source tracing and automated optimization strategies.

[0024] Preferred incremental learning algorithms include stochastic gradient descent and Bayesian update; adversarial example generation methods include fast gradient signification, projective gradient descent, and generative adversarial networks; false alarm attribution methods include feature attribution analysis and rule conflict detection; and automated optimization strategies include active learning and dynamic threshold adjustment.

[0025] NetFlow is a data exchange method that provides a session-level view of network traffic, recording information for each TCP / IP transaction.

[0026] A NetFlow flow is defined as a unidirectional stream of data packets transmitted between a source IP address and a destination IP address, with all packets sharing a common transport layer source and destination port number. NetFlow uses a standard switching pattern to process the first IP packet of the data stream and then generates a NetFlow cache. Subsequently, the same data is transmitted in the same data stream based on the cached information, without matching related access control policies. The NetFlow cache also contains statistical information for subsequent data streams.

[0027] Supervised learning is a paradigm in machine learning. Its core idea is to use labeled training data to allow the model to learn the mapping relationship between inputs (features) and outputs (labels), thereby making predictions about unknown data. In the field of computer security, supervised learning is often used for classification (such as malware detection) and regression (such as risk score prediction) tasks.

[0028] Supervised learning algorithms include two types: classification algorithms, which are used to output discrete labels, and regression algorithms, which are used to output continuous values. Classification algorithms include logistic regression, decision trees, random forests, support vector machines, and neural networks; Among them, the logistic regression algorithm is characterized by linear classification, output probability, and strong interpretability. It is mainly used for spam detection and login anomaly identification. Among them, the decision tree algorithm is characterized by rule-based splitting, easy interpretation, and is mainly used for malware classification and privilege abuse detection; Among them, the random forest algorithm is characterized by multi-decision tree ensemble and resistance to overfitting, and is mainly used for network intrusion detection; Among them, the support vector machine algorithm is characterized by high-dimensional space classification, which is suitable for small samples and is mainly used for ransomware behavior analysis; Among them, neural network algorithms are characterized by complex nonlinear relationships, require a large amount of data, and are mainly used for deep packet inspection; Regression algorithms include linear regression and time series models; Linear regression is characterized by predicting continuous values ​​and is mainly used for predicting the scale of DDoS attack traffic. The time series module is characterized by processing timestamped data and is mainly used to detect periodic attacks.

[0029] Example 1 Malware detection: Input features include PE file header information, API call sequence, and binary entropy value. The label is 0 or 1, where 0 indicates benign and 1 indicates malicious. The model uses random forest.

[0030] Network intrusion detection: Input features include traffic protocol type, packet length, and source IP, with labels such as normal, DoS, and port scan. The model uses a gradient boosting tree.

[0031] Phishing website identification: Input features include URL structure, domain age, and page keywords, with labels indicating whether it is phishing or legitimate. The model uses logistic regression.

[0032] User behavior analysis: Input features include login time, location, and operation frequency; labels are normal or internal threat; the model uses an isolated forest.

[0033] Unsupervised learning is a paradigm of machine learning characterized by data that is not pre-labeled; the model must autonomously discover hidden patterns, structures, or anomalies within the data. In the field of computer security, unsupervised learning is primarily used for anomaly detection, cluster analysis, and dimensionality reduction, and is suitable for discovering unknown threats and detecting zero-day attacks.

[0034] The basic process of unsupervised learning includes input data, model training, and output results. The input data includes only feature data and unlabeled data, such as network traffic logs, user behavior sequences, and system call records. Among them, model training is mainly used for the algorithm to automatically learn the distribution or similarity of data; The output structure includes clustering, anomaly detection, and dimensionality reduction. Clustering is used to group similar data together. Anomaly detection is used to identify data points that deviate from the normal pattern; Dimensionality reduction is used to compress feature dimensions, making visualization or subsequent analysis easier.

[0035] Unsupervised learning algorithms include clustering algorithms, anomaly detection algorithms, and dimensionality reduction algorithms; The clustering algorithms include K-means, DBSCAN, hierarchical clustering, and Gaussian mixture model. The K-means algorithm is characterized by dividing clusters based on distance, requiring the specification of a K value, and is mainly used for malware family classification. The DBSCAN algorithm is characterized by its density-based nature, its ability to discover clusters of arbitrary shapes, its suitability for noisy data, and its primary use for identifying cyberattack groups. Hierarchical clustering algorithms are characterized by generating tree-like structures, do not require a preset K value, and are mainly used for log pattern analysis; The Gaussian Mixture Model algorithm is characterized by assuming that the data follows a Gaussian distribution and soft clustering, and is mainly used for baseline modeling of user behavior.

[0036] Anomaly detection algorithms include Isolation Forest, One-Class SVM, Autoencoder, and LOF; The Isolation Forest algorithm is characterized by isolating outliers through random segmentation, making it suitable for high-dimensional data and primarily used for intrusion detection. The One-Class SVM algorithm is characterized by training with only normal data and identifying points that deviate from the boundary. It is mainly used for anomaly monitoring in industrial control systems. The characteristics of autoencoder algorithms are that they reconstruct the input through neural networks, resulting in high reconstruction errors for abnormal data. They are mainly used for malicious traffic detection. The LOF algorithm is characterized by being based on local density comparison, making it suitable for data in regions with different densities. It is mainly used for monitoring financial fraud transactions.

[0037] Dimensionality reduction algorithms include PCA, t-SNE, and UMAP; The PCA algorithm is characterized by linear dimensionality reduction and preservation of maximum variance, and is mainly used for network traffic feature compression. The t-SNE algorithm is characterized by non-linear dimensionality reduction and preservation of local similarity, and is mainly used for the visualization of malware samples. The UMAP algorithm is similar to the t-SNE algorithm but has higher computational efficiency and is mainly used for threat intelligence data exploration.

[0038] The advantages of unsupervised learning algorithms are: no need for labeled data, discovery of unknown threats, and high degree of automation.

[0039] Here, no labeled data is required, which addresses the problem of scarce labels in real-world security scenarios; Here, unknown threats can be discovered, and new types of attacks can be detected, such as zero-day vulnerabilities and APT attacks. This area is highly automated and suitable for preliminary analysis of large-scale data.

[0040] Example 2 Anomaly detection: The application scenario is to identify internal personnel data theft and account abuse. It uses isolated forest to analyze user login time and file access frequency, and clustering to discover abnormal operation patterns. Network intrusion detection: The application scenario is to detect unknown attacks, such as zero-day exploits. One-Class SVM is used to train a normal traffic baseline. Traffic that deviates from the baseline is considered abnormal. Autoencoder is used to reconstruct network traffic packets. Packets with high reconstruction errors may be attacks. Malware analysis: The application scenario is to discover new malware variants. K-means clustering API call sequences are used to group unknown families into new clusters. PCA dimensionality reduction is used to visualize PE file features and identify outlier samples. Log anomaly analysis: The application scenario is to locate potential attack traces from massive system logs. The LOF algorithm is used to detect abnormal log entries, and topic models, such as LDA, are used to extract log keyword patterns.

[0041] Reinforcement learning (RL) is a machine learning paradigm that learns optimal policies through interaction with the environment. Its core idea is that an agent learns to make optimal decisions in a given environment by learning through trial and error and receiving rewards. In the field of computer security, RL is suitable for dynamic adversarial scenarios, such as adaptive defense, automated penetration testing, and intrusion response.

[0042] The basic framework of reinforcement learning includes agent, environment, state, action, reward, and policy; In this context, the intelligent agent is the security system, such as a firewall or intrusion detection system; the environment is the network, host, or attacker behavior; the state is the current system snapshot, such as traffic load or vulnerability status; the action is the defensive measure, such as blocking IPs or isolating terminals; the reward is the feedback on the effectiveness of the defense; and the policy is the mapping from state to action, such as decision rules.

[0043] Among the classifications of reinforcement learning algorithms, the value-based representative algorithm is DQN, which is characterized by learning the state-action value function and is mainly used for dynamic firewall rule optimization. The representative policy-based algorithm is PPO, which is characterized by directly optimizing the policy without requiring a Q-table and is mainly used for adaptive honeypot trapping strategies. The representative algorithm of Actor-Critic is SAC, which combines values ​​and strategies, has high stability, and is mainly used in automated intrusion response systems. The representative algorithm for multi-agent RL is MADDPG, which is characterized by cooperation and adversarial among multiple agents and is mainly used for red-blue team adversarial simulation.

[0044] Example 3 Adaptive Intrusion Detection and Response: The application scenario is to dynamically adjust the detection threshold to reduce false alarms. It adopts the PPO algorithm, which adjusts the detection sensitivity and switches the detection model according to the current attack frequency, false alarm rate and system load. If a real attack is detected, the value increases by 1, the false alarm decreases by 0.2, and the missed attack decreases by 1. Automated penetration testing: The application scenario is to simulate an attacker searching for vulnerable paths. It uses the DDPG algorithm and selects attack tools based on the current target network topology, open ports, and known vulnerabilities. When privileges are granted +10, an alert is triggered -5. Dynamic defense: The application scenario is to confuse attackers with randomized system configuration. It adopts the SAC algorithm and switches the virtual IP and changes the service port according to the current configuration and the attacker's probing behavior. Attack failure +1, service interruption -2; Malware countermeasures: The application scenario is to generate adversarial samples to bypass AI detection. The PPO algorithm is used to insert redundant code and obfuscated strings based on malware characteristics, such as API call sequences. If the malware bypasses detection, the score increases by 1; if it is identified, the score decreases by 1.

[0045] The advantages of reinforcement learning are dynamic adaptation, long-term planning, and automated response.

[0046] Among these features, dynamic adaptation can adjust strategies in real time to deal with new types of attacks, long-term planning can consider multiple decision-making processes, and automated response can reduce human intervention.

[0047] Graph Neural Networks (GNNs) are deep learning models specifically designed for processing graph-structured data. They learn and reason by capturing the relationships between nodes, edges, and the global graph. In the field of computer security, GNNs are suitable for scenarios such as attack path analysis, malware association, and social network fraud detection.

[0048] In this graph, nodes are entities, such as IP addresses, user accounts, and processes; edges are relationships between entities, such as communication traffic and data flow; and the graph is a collection of nodes and edges, which may have attributes such as weight and direction.

[0049] The core ideas of graph neural networks are neighbor aggregation, state update, and graph-level output.

[0050] Among them, neighbor aggregation is used to collect features of neighboring nodes, such as mean and maximum value; State updates are used to combine their own features and aggregation results to generate new representations; Graph-level output is generated by pooling to produce features for the entire graph, such as classification.

[0051] Among them, the representative algorithm of graph convolutional network model is GCN, which is characterized by weighted aggregation of neighbor information, suitable for homogeneous graphs, and mainly used for network intrusion detection; The representative algorithm of graph attention network model is GAT, which is characterized by dynamically learning the importance of neighbors and is mainly used for key point recognition. The representative algorithm of the graph autoencoder model is GAE, which is characterized by unsupervised learning graph embedding and is used for anomaly detection, mainly for internal threat behavior analysis. The representative algorithm of the spatiotemporal graph network model is DCRNN, which is characterized by its ability to process dynamic graphs and is mainly used for DDoS attack prediction. The representative algorithm of heterogeneous graph neural network model is HAN, which is characterized by its ability to handle multiple node and edge types and is mainly used for APT attack knowledge graph reasoning.

[0052] Example 4 Network intrusion detection: Input a network traffic graph, use GCN to learn the embedding representation of nodes, classify malicious nodes or edges, construct a traffic graph based on NSL-KDD data, and the graph neural network detection rate is 15% higher than that of traditional methods; Malware analysis: Input code behavior graph, learn the semantic features of API call graph through graph neural network, and classify malware families; Social network fraud detection: Input a relationship graph and use GAT to identify abnormal subgraphs, such as star-shaped structures where fake accounts are clustered, and combine node features for detection. APT attack tracing: Input the attack chain graph, use RGCN to construct a heterogeneous graph, and predict the attacker's next action or key vulnerabilities; IoT security devices: Input device communication diagram, use STGNN to detect abnormal communication patterns, such as data leakage.

[0053] The advantages of graph neural networks include relational modeling capabilities, interpretability, and end-to-end learning. Among these, the ability to model relationships is manifested in the direct processing of non-Euclidean data, such as social networks and code structures; Interpretability is demonstrated by locating key nodes or edges through attention weights; End-to-end learning means that no manual extraction of graph features is required.

[0054] Dynamic detection engines are a core component of security protection systems, identifying threats through real-time behavior monitoring, dynamic analysis, and immediate response. Compared to static detection (such as signature matching), its core advantage lies in its ability to detect advanced threats such as unknown attacks and zero-day exploits. The following is an in-depth analysis of its architecture, key technologies, and typical application scenarios.

[0055] The dynamic detection engine adopts a layered design to achieve closed-loop processing from data acquisition to response.

[0056] The data acquisition layer is primarily responsible for capturing behavioral data in real time, such as processes, networks, files, and the registry. The behavior analysis layer's main function is to dynamically parse behavior sequences and identify suspicious patterns. The threat assessment layer's main function is comprehensive scoring and classification, such as malicious, suspicious, or normal. The response layer's main functions are to perform blocking, alarming, and isolation. The feedback optimization layer's main function is to record false alarms or missed alarms and iterate the model and rules.

[0057] The dynamic analysis method used by the dynamic detection engine employs a sandbox execution method that runs samples in an isolated environment, recording API calls, network requests, etc. The principle of sequence modeling is to use LSTM to model process behavior sequences; The principle of graph behavior analysis is to construct a process-file-network relationship graph and detect abnormal subgraphs; The principle of differential analysis is to compare normal or abnormal behavior baselines, such as a sudden increase in CPU usage.

[0058] The response module is the "operational hub" of the security protection system, responsible for executing rapid and accurate response measures to detected threats. Its core objective is to minimize the impact of attacks and restore the system to a normal state. Unlike the "analysis" function of the detection module, the response module emphasizes "execution," requiring a balance between automation efficiency and operational security. The following is an in-depth analysis of its technical architecture, strategy design, and key challenges.

[0059] The decision-making layer's function is to select a response strategy based on the threat level and context. The function of the execution layer is to call the underlying interface to perform specific operations; The function of the isolation layer is to limit the spread of attacks, such as network isolation or process suspension; The repair layer's function is to automatically restore damaged assets; The function of the source tracing layer is to record evidence of the attack chain for subsequent analysis.

[0060] The key response strategies and technologies of the response module include real-time blocking, isolation and containment, repair and recovery, and deception and countermeasures. Real-time blocking includes network layer blocking and host layer termination. Network layer blocking includes coordinated firewalls and dynamically updated DNS blacklists. Host-level termination includes ending malicious processes and their child process trees, and disabling suspicious accounts.

[0061] Isolation and containment include unisolated and sandbox redirection. Micro-segmentation restricts lateral movement using SDN technology, allowing only necessary communication. Sandbox redirection redirects file operations to a virtual environment.

[0062] The repair and recovery categories include automated repair and credential reset. Automated repair includes rolling back tampered files and clearing persistent items, while credential reset forces multi-factor authentication to re-verify high-risk sessions.

[0063] Deception and countermeasures include honeypot linkage and attacker interference. Honeypot linkage involves directing attacker traffic to the decoy system, while attacker interference refers to injecting fake data.

[0064] Explanatory tools are technologies and methods used to reveal the decision-making logic of AI models, helping security analysts understand "why the model judged a sample as malicious." In the security field, model interpretability is crucial—it not only needs to detect threats but also provide credible evidence for human verification, compliance audits, and attack attribution. The following is an in-depth analysis of its core methods, tools, and security scenarios.

[0065] Explanatory tools are primarily designed to reduce false positive rates, distinguish between real threats and model biases, such as over-reliance on irrelevant features, meet the requirements of GDPR, AI ethics and other regulations on "algorithmic transparency", locate key attack features, such as suspicious API calls in malware, and discover feature or data defects, such as misjudgments caused by outdated threat intelligence.

[0066] The LIME method in post-hoc interpretation techniques is based on local linear approximation, which uses a simple model to explain a single prediction of a complex model. It is suitable for malware classification and abnormal login detection. The SHAP method is based on game theory and calculates the contribution of each feature to the prediction. It is suitable for network traffic analysis and UEBA behavior scoring. The principle of the attention mechanism method is to display the key input parts that the neural network focuses on, such as malicious fragments in code, and it is applicable to NLP models and graph neural networks. The principle of adversarial example analysis is to generate small perturbation inputs and observe the changes in the model output, which is suitable for verifying the robustness of the model.

[0067] Online learning, adversarial training, and false positive analysis are the three core mechanisms in an AI-driven security system, addressing issues such as dynamic model updates, adversarial attack defense, and result optimization, respectively.

[0068] Online learning is an incremental learning method where the model is updated in real time as new data arrives, eliminating the need for full retraining and making it suitable for dynamically changing threat environments.

[0069] Stochastic gradient descent is used to update model parameters on a sample-by-sample basis, while Bayesian update is used to dynamically adjust the probability distribution, such as in threat scoring models. Streaming data processing uses Flink to process real-time log streams; Model version control is used for rolling updates or hot switching.

[0070] In adversarial training mechanisms, the fast gradient sign method is mainly used to add perturbations along the gradient direction; Projected gradient descent is primarily used for iterative optimization of adversarial examples; Generative adversarial networks (GANs) are primarily used by generators to create adversarial examples and by discriminators to improve models.

[0071] The goal of robustness optimization is to minimize the difference in loss between real samples and adversarial samples.

[0072] Feature attribution analysis is mainly used to locate high-weight misclassified features; Rule conflict detection is used to compare model decisions with manually entered rules; Automatic learning is used to iteratively train manually labeled high-uncertainty samples. The threshold is dynamically adjusted based on the flat curve of false alarm rate and false negative rate.

[0073] The collaborative application of online learning, adversarial training, and false positive analysis enables the real-time reception of new data, the periodic generation of adversarial examples to update the model, and the manual review of false positive cases, with the goal of correcting feature engineering through SHAP analysis.

[0074] Here, online learning addresses the problem of model obsolescence and adapts to dynamic threats; adversarial training addresses the problem of attack bypass and improves robustness; and false alarm analysis addresses the problem of alarm fatigue and optimizes accuracy.

[0075] Compared with existing technologies, this AI-based computer security protection system combines supervised learning, unsupervised learning, and reinforcement learning, integrating multiple learning paradigms. It can solve problems such as model obsolescence, attack bypass, and alarm fatigue, improve robustness, optimize accuracy, generate decision-making basis using tools such as SHAP, improve interpretability, and achieve human-machine collaboration. Key decisions retain human review, thus enhancing security.

[0076] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A computer security protection system based on artificial intelligence, characterized in that, It includes a data acquisition layer, a data processing and feature engineering layer, an AI model layer, a real-time detection and response layer, a feedback and adaptation layer, and a management and visualization layer; In the data acquisition layer, data sources include network traffic, terminal behavior, system logs, external intelligence, and user behavior; The network traffic data includes firewall logs, NetFlow, and packet captures; The terminal behavior data includes process monitoring, file operations, and registry changes; The system log data includes operating system logs and application logs; The external intelligence data includes threat intelligence platforms and building databases; The user behavior data includes login mode, permission usage, and abnormal operations; In the data processing and feature engineering layer, the data processing steps include data normalization, feature extraction, and dimensionality reduction and encoding; Feature extraction targets include network layer features, host layer features, and user layer features; Network layer characteristics include traffic frequency, protocol distribution, and connection duration; Host-level characteristics include abnormal memory usage and changes in file hashes; User-level characteristics include login geographic location and operation time sequence; The AI ​​model layer includes the core model and model training; The core models include supervised learning, unsupervised learning, reinforcement learning, and graph neural networks; Supervised learning is used for malware detection and intrusion detection. Unsupervised learning functions include anomaly detection and cluster analysis. The function of reinforcement learning is to dynamically adjust defense strategies; The function of a graph neural network is to analyze the attacker's lateral movement path; The real-time monitoring and response layer includes a dynamic detection engine, a response module, and interpretive tools; The feedback and adaptive layer includes online learning mechanisms, adversarial training mechanisms, and false alarm analysis mechanisms; The management and visualization layer includes dashboards, alarm systems, and audit logs.

2. The computer security protection system based on artificial intelligence as described in claim 1, characterized in that, Supervised learning includes classification algorithms and regression algorithms. Classification algorithms include logistic regression, decision trees, random forests, support vector machines, and neural networks, while regression algorithms include linear regression and time series models.

3. The artificial intelligence-based computer security protection system as described in claim 1, characterized in that, The process of unsupervised learning includes input data, model training, and output results. Unsupervised learning algorithms include clustering algorithms, anomaly detection algorithms, and dimensionality reduction algorithms.

4. The computer security protection system based on artificial intelligence as described in claim 1, characterized in that, Reinforcement learning algorithms include value-based, policy-based, actor-critic, and multi-agent RL.

5. The computer security protection system based on artificial intelligence as described in claim 1, characterized in that, Graph neural networks include graph convolutional network models, graph attention network models, graph autoencoder models, spatiotemporal graph network models, and heterogeneous graph neural network models.

6. The computer security protection system based on artificial intelligence as described in claim 1, characterized in that, The dynamic detection engine includes a data acquisition layer, a behavior analysis layer, a threat determination layer, a response layer, and a feedback optimization layer. The dynamic analysis methods used by the dynamic detection engine include sandbox execution, sequence modeling, graph behavior analysis, and differential analysis.

7. The computer security protection system based on artificial intelligence as described in claim 1, characterized in that, The response module includes a decision-making layer, an execution layer, an isolation layer, a repair layer, and a source tracing layer.

8. The computer security protection system based on artificial intelligence as described in claim 1, characterized in that, The core explanatory methods of explanatory tools include model-inherent explanatory and ex-post explanatory techniques. Inherent models include decision trees, linear regression, and rule engines, while ex-post explanatory techniques include LIME, SHAP, attention mechanisms, and adversarial example analysis.

9. The computer security protection system based on artificial intelligence as described in claim 1, characterized in that, Online learning mechanisms include incremental learning algorithms, streaming data processing, and model version control; adversarial training mechanisms include adversarial example generation methods and robust optimization objectives; and false alarm analysis mechanisms include false alarm source tracing and automated optimization strategies.

10. The computer security protection system based on artificial intelligence as described in claim 9, characterized in that, Incremental learning algorithms include stochastic gradient descent and Bayesian update; adversarial example generation methods include fast gradient signification, projective gradient descent, and generative adversarial networks; false alarm attribution methods include feature attribution analysis and rule conflict detection; and automated optimization strategies include active learning and dynamic threshold adjustment.