AI-based enterprise network security operation and management system
By attaching scheduling tags to data packets and performing lightweight logging operations, the problem of feature loss caused by competition for shared computing resources is solved, achieving complete feature preservation and business continuity under high load conditions.
Patent Information
- Application Number
- CN202611469027.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-21
- Publication Date
- 2026-11-17
AI Technical Summary
Under conditions of competition for shared GPU computing resources, existing network security hosting platforms may skip deep feature extraction due to data packets exceeding the queuing timeout, resulting in the permanent loss of attack traces. Subsequently, the AI review module may misjudge normal business as an attack due to the lack of key time slice features, triggering an interruption.
By attaching scheduling tags to data packets, the load status of the shared computing pool is perceived, critical packets are marked and lightweight trace operations are performed, the original payload is randomly projected into low-dimensional measures and binary codes, encapsulated into feature packets and stored in an independent cache, and after the contention is alleviated, the approximate feature sequence is reconstructed and backfilled into the detection pipeline.
Under high-load scenarios, the original traffic characteristics of data packets are preserved, avoiding misjudgments and business interruptions, and ensuring the integrity of the detection pipeline and the continuity of production operations.
Smart Images

Figure CN123122972A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of cybersecurity technology, and more specifically to an artificial intelligence-based enterprise cybersecurity operation and management system. Background Technology
[0002] In recent years, enterprise cybersecurity operations have gradually evolved from locally deployed independent protection systems to cloud-hosted service models. In this model, traffic acquisition probes at the customer's network boundary continuously transmit inbound and outbound traffic, log data, and device alarm information through encrypted tunnels to the hosting service provider's cloud security operations platform. The platform then performs threat detection, analysis, and response on behalf of the customer. To improve detection efficiency and accuracy, mainstream hosting platforms have introduced deep learning-based feature extraction models into their detection pipelines. These models extract deep feature vectors from the packet payloads of tunnel access and identify attack behaviors based on the extracted feature sequences.
[0003] Under conditions of competition for shared GPU computing resources, existing network security hosting platforms may experience a situation where, when data packets are skipped from deep feature extraction due to queuing timeouts, the original payload features of that time window are directly discarded without any persistent storage, resulting in the permanent loss of attack traces. Subsequently, the AI review module, lacking key time slice features, relies solely on incomplete fragments before and after it to perform sequence completion inferences, which can easily misjudge normal business traffic as attack compensation behavior and trigger automated blocking, causing production business interruptions. Summary of the Invention
[0004] The purpose of this invention is to provide an artificial intelligence-based enterprise network security operation hosting system to solve the problems mentioned above.
[0005] The objective of this invention can be achieved through the following technical solutions: An AI-based enterprise cybersecurity operations hosting system includes: The traffic access and label generation module acquires data packets accessed via the tunnel and attaches scheduling labels to them based on the tenant priority and tunnel identifier. The load awareness and critical packet determination module continuously senses the current load status of the shared computing pool based on scheduling tags and dynamically tracks the historical retention time of preceding data packets in the same tunnel. If the current load status exceeds the high load threshold and the actual processing time of multiple consecutive preceding data packets in the same tunnel reaches the preset safety threshold, the current data packet is marked as a critical packet. The lightweight trace and feature caching module stops deep extraction on critical packets and performs lightweight trace: the original payload is randomly projected into a low-dimensional measure and a distance-preserving binary code is generated. The measure, binary code, timestamp, and tunnel identifier are encapsulated into a feature packet and stored in an independent cache. The feature reconstruction and backfilling module reads the feature package from the cache after the competition is relieved, reconstructs the approximate feature sequence of the critical package using the measure and sparse basis, and backfills it into the corresponding window of the detection pipeline along with the reconstruction confidence. The comprehensive assessment and disposal decision module assesses the comprehensive confidence of the feature sequence to be judged, including reconstruction confidence, before triggering disposal. If it is lower than the blocking threshold, it is downgraded to manual review and a confirmation instruction is sent. Otherwise, a control instruction is directly issued to the execution device.
[0006] As a further aspect of the present invention: the step of attaching a scheduling label to the tenant based on its priority and tunnel identifier specifically includes: Obtain the tunnel identifier carried by the data packet, query the tunnel's historical packet loss rate and latency fluctuation value within the current time window, and calculate the tunnel's current stability score based on the historical packet loss rate and latency fluctuation value; The weighting coefficients of tenant priorities are adjusted based on the current stability score to obtain dynamic priorities; The dynamic priority is associated with the tunnel identifier and written into the label field of the data packet as a scheduling label.
[0007] As a further aspect of the present invention: marking the current data packet as a critical packet specifically includes: Based on the tunnel identifier recorded in the scheduling tag, retrieve the dwell time of each of the preceding data packets of the tunnel before the current time, and calculate the rate of change of the dwell time. The system senses the current load value of the shared computing pool and compares it with the historical average load value of the tunnel to obtain the degree of load offset. When the load offset exceeds the set offset threshold and the rate of change shows an accelerating growth trend, the current data packet is marked as a critical packet.
[0008] As a further aspect of the present invention: the storage in an independent cache specifically includes: Read the load length value of the critical package, dynamically determine the projection dimension based on the load length value, and perform random projection on the original load according to the projection dimension to obtain a low-dimensional measure; Using the tunnel identifier as a component of the random seed, a binary transformation is performed on the low-dimensional measure to generate a binary code that maintains distance similarity to the original payload; The low-dimensional measure, binary code, timestamp, and tunnel identifier are encapsulated together into a feature package and written to an independent cache.
[0009] As a further aspect of the present invention: the generation of binary code that maintains distance similarity with the original payload specifically includes: The tunnel identifier and the current timestamp are concatenated bit by bit, and then a hash operation is performed to obtain a random seed value; Starting with a random seed value, several projection directions are sequentially selected from a preset projection direction table; The low-dimensional measure is compared with each selected projection direction. Based on the comparison results, the binary bits are output bit by bit, and the binary bits are combined to obtain the binary code.
[0010] As a further aspect of the present invention: the reconstructed confidence level is backfilled to the corresponding window of the detection pipeline, specifically including: Low-dimensional measure and binary code are extracted from the feature package. The binary code is used as an index to find matching basis atom combinations in the sparse basis. The low-dimensional measure and the found basis atom combinations are linearly synthesized to obtain an approximate feature sequence. Calculate the deviation between the approximate feature sequence and the expected features of the original payload at the timestamp recorded in the feature package, and determine the reconstruction confidence based on the deviation. The approximate feature sequence and the reconstructed confidence are written together into the window position corresponding to the timestamp in the detection pipeline, overwriting the original empty marker of the corresponding window.
[0011] As a further aspect of the present invention: the process of generating the approximate feature sequence is as follows: The binary code is divided into multiple index segments with a fixed bit width, and each index segment corresponds to a decimal index value. Based on the decimal index value, the substrate atoms at the corresponding positions are sequentially extracted from the sparse substrate, and all the extracted substrate atoms are combined into a substrate atom combination according to the extraction order. Using the values of each dimension in the low-dimensional measure as the synthesis weights of the corresponding atoms in the base atom combination, the atoms are superimposed according to the synthesis weights to generate an approximate feature sequence.
[0012] As a further aspect of the present invention: the direct issuance of control commands to the execution device specifically includes: Extract all reconstructed confidence scores carried in the feature sequence to be judged, and take the median value of all reconstructed confidence scores as the overall correction coefficient of the feature sequence to be judged; The original confidence scores derived from the complete features in the feature sequence to be judged are weighted and fused with the overall correction coefficient to obtain the comprehensive confidence score. The overall confidence level is compared with the blocking threshold. If the threshold is reached or exceeded, a control command is directly sent to the execution device. Otherwise, a confirmation command is output and the automatic processing process is suspended.
[0013] The beneficial effects of this invention are: (1) This invention suspends deep feature extraction and performs lightweight trace retention on critical packets when resource contention intensifies in the shared computing pool. The original payload is encapsulated as a feature packet in the form of low-dimensional measure and binary code and stored in an independent cache. This ensures that even data packets that are skipped from deep extraction due to insufficient GPU memory resources under high load scenarios still retain their original traffic features in a compressed form. After the resource contention is alleviated, an approximate feature sequence is reconstructed through a sparse basis and backfilled into the corresponding window of the detection pipeline. This ensures that there is no time blind spot with missing features in the detection pipeline and preserves a complete data foundation for subsequent attack tracing.
[0014] (2) Before triggering the automated handling decision, this invention extracts the median value of all reconstructed confidence scores in the feature sequence to be judged as the overall correction coefficient, and performs a weighted fusion with the original confidence score to obtain a comprehensive confidence score. The comprehensive confidence score is then compared with the blocking threshold to determine whether to issue a control command. This mechanism automatically constrains the contribution of the time window containing reconstructed features to the final handling decision by the confidence weighting mechanism. Automatic blocking is only executed when the reconstructed features and the complete features jointly support a high confidence score decision, thus avoiding the misjudgment of normal business traffic as an attack and interruption due to minor deviations introduced by feature reconstruction, effectively ensuring the continuity of production business. Attached Figure Description
[0015] The invention will now be further described with reference to the accompanying drawings.
[0016] Figure 1 This is a system block diagram of the present invention; Figure 2 This is a flowchart illustrating how the current data packet is marked as a critical packet in this invention. Figure 3 This is a flowchart of the process of storing data in an independent cache in this invention. Detailed Implementation
[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0018] Please see Figure 1 As shown, this invention is an artificial intelligence-based enterprise network security operation and management system, comprising: The traffic access and label generation module acquires data packets accessed via the tunnel and attaches scheduling labels to them based on the tenant priority and tunnel identifier. The load awareness and critical packet determination module continuously senses the current load status of the shared computing pool based on scheduling tags and dynamically tracks the historical retention time of preceding data packets in the same tunnel. If the current load status exceeds the high load threshold and the actual processing time of multiple consecutive preceding data packets in the same tunnel reaches the preset safety threshold, the current data packet is marked as a critical packet. The lightweight trace and feature caching module stops deep extraction on critical packets and performs lightweight trace: the original payload is randomly projected into a low-dimensional measure and a distance-preserving binary code is generated. The measure, binary code, timestamp, and tunnel identifier are encapsulated into a feature packet and stored in an independent cache. The feature reconstruction and backfilling module reads the feature package from the cache after the competition is relieved, reconstructs the approximate feature sequence of the critical package using the measure and sparse basis, and backfills it into the corresponding window of the detection pipeline along with the reconstruction confidence. The comprehensive assessment and disposal decision module assesses the comprehensive confidence of the feature sequence to be judged, including reconstruction confidence, before triggering disposal. If it is lower than the blocking threshold, it is downgraded to manual review and a confirmation instruction is sent. Otherwise, a control instruction is directly issued to the execution device.
[0019] In the traffic access and label generation module, data packets accessed via the tunnel are acquired, and scheduling labels are attached to them based on the tenant priority and tunnel identifier. Specifically, this includes: The tunnel identifier is extracted by the tunnel access gateway and attached to the header extension field of the data packet when it passes through the tunnel access platform. Each data packet carries the tunnel identifier and the tenant priority identifier. The tenant priority identifier is set by the platform according to the service level agreement when the tenant signs up and is stored in the tenant configuration library.
[0020] To obtain the stability score of the current tunnel, the platform maintains a statistical record corresponding to each tunnel identifier in memory. This statistical record contains the packet loss count and latency record of all data packets transmitted in the tunnel in the past 60 seconds. The latency record is the difference between the timestamp of each data packet entering the tunnel and the timestamp received by the platform.
[0021] Whenever a new data packet arrives, the platform updates the statistics of the tunnel. The packet loss count is accumulated by detecting the continuity of the sequence number of the data packets. If a missing sequence number is detected, the number of missing numbers is added to the packet loss count. The latency record stores the latency values of the 100 most recently arrived data packets.
[0022] The platform reads the ratio of the packet loss count to the total number of data packets sent in the tunnel over the past 60 seconds from the statistical records as the historical packet loss rate, and at the same time calculates the standard deviation of the latency values of the 100 most recently arrived data packets in the tunnel as the latency fluctuation value.
[0023] Then, based on the historical packet loss rate and latency fluctuation value, the stability score of the tunnel is calculated. The stability score is calculated as follows: a full score of 100 is set, and the product of the historical packet loss rate and the first coefficient is subtracted from the full score, and the product of the latency fluctuation value and the second coefficient is subtracted from the full score. The difference is the stability score of the tunnel.
[0024] The first coefficient is set based on the following: when the historical packet loss rate increases by one-thousandth, the stability score decreases by 1 point, so the first coefficient is set to 1000. The second coefficient is set based on the following: when the latency fluctuation value increases by 1 millisecond, the stability score decreases by 2 points, so the second coefficient is set to 2. The above coefficient values are obtained by fitting the offline test statistics of multiple tunnels with different quality from the platform.
[0025] After obtaining the stability score, the platform reads the basic weight value corresponding to the tenant priority from the tenant configuration library. The basic weight value is an integer between 0 and 100.
[0026] The platform uses the average of the stability score and the basic weight value as the dynamic priority. The value of this dynamic priority and the tunnel identifier are written together into the scheduling tag area in the data packet tag field. This scheduling tag area is always transmitted with the data packet during the subsequent flow of the data packet and is used for the queuing order and resource allocation decisions of the data packets in each subsequent processing stage.
[0027] Please see Figure 2 As shown, in the load awareness and critical packet determination module, based on scheduling tags, the current load status of the shared computing pool is continuously sensed, and the historical latency of preceding data packets within the same tunnel is dynamically tracked. If the current load status exceeds the high load threshold and the cumulative actual processing time of multiple consecutive preceding data packets within the same tunnel reaches a preset safety threshold, the current data packet is marked as a critical packet. Specifically, this includes: The scheduling tag records the tunnel identifier and the enqueue timestamp and dequeue timestamp of each data packet under that tunnel. The enqueue timestamp is the moment when the data packet enters the shared computing pool queue, and the dequeue timestamp is the moment when the data packet is allocated to the computing resources to start execution. The difference between the two is the dwell time of the data packet.
[0028] The platform maintains a latency queue for each tunnel in memory. This queue stores the latency of the 50 most recently arrived data packets before the current time. Whenever a data packet completes the allocation of computing resources, the platform appends the latency of the data packet to the tail of the latency queue of the corresponding tunnel and removes the earliest record at the head of the queue, keeping the queue length at 50.
[0029] The platform reads all 50 latency records from the latency queue corresponding to the tunnel, divides the difference between two adjacent latency by the collection time interval according to the arrival order of the data packets, and obtains multiple rate of change values. The average of these multiple rate of change values is then taken as the rate of change of the latency of the tunnel. This rate of change is used to characterize the acceleration or deceleration trend of the data packet queuing waiting time in the tunnel over time.
[0030] When sensing the current load value of the shared computing pool, the platform calculates the ratio of the number of data packets currently executing inference tasks to the total number of computing units in the shared computing pool in real time as the current load value, which is between 0 and 1.
[0031] The platform also maintains the historical average load value of the tunnel, which is the arithmetic mean of the load values collected every second over the past 300 seconds. Subtracting the historical average load value from the current load value yields the load offset. A positive load offset indicates that the current load is higher than the historical average load value of the tunnel, while a negative value indicates that the load is lower than the historical average load value.
[0032] When determining the critical packet, the platform first checks whether the load offset exceeds the set offset threshold, which is set to 0.2. The basis for this setting is that when the load offset exceeds 0.2, according to the platform's historical operation data, the queuing latency of the shared computing pool begins to show a non-linear growth trend.
[0033] If the load offset exceeds 0.2, it is further determined whether the rate of change shows an accelerating growth trend. The method for determining the accelerating growth trend is as follows: calculate the average of the 5 most recent stay duration records in the stay duration queue and the 5 stay duration records before that. If the average of the 5 most recent records is greater than the average of the 5 records before that, it is determined to be an accelerating growth trend.
[0034] When both conditions are met simultaneously—that the load offset exceeds the set offset threshold and the rate of change shows an accelerating growth trend—the platform marks the current data packet as a critical packet.
[0035] Please see Figure 3 As shown, in the lightweight tracekeeping and feature caching module, deep extraction is stopped for the critical packet, and lightweight tracekeeping is performed: the original payload is randomly projected into a low-dimensional measure and a distance-preserving binary code is generated. The measure, binary code, timestamp, and tunnel identifier are encapsulated into a feature packet and stored in an independent cache, specifically including: Once a critical package is marked, the platform suspends the deep feature extraction process for that critical package and instead performs a lightweight trace retention operation.
[0036] The platform reads the payload length value of the critical packet and records it as the first value. This first value represents the data size of the original payload in bytes.
[0037] The platform dynamically determines the projection dimension based on the first value. The projection dimension is determined by dividing the first value by the compression base and adding the first offset. The compression base is 8, and the first offset is the integer value obtained by dividing the last byte value in the tunnel identifier of the critical packet by the second coefficient. The second coefficient is 4.
[0038] The result of the projection dimension calculation is denoted as the second value, which is the target dimension for subsequent random projection operations.
[0039] When performing a random projection operation on the original load according to the above projection dimension, the platform pre-generates a random projection matrix with a dimension of the second value multiplied by the first value. Each element in the random projection matrix follows a standard normal distribution.
[0040] The original load is considered as a column vector of the first numerical dimension. The random projection matrix is multiplied by the column vector to obtain a column vector of the second numerical dimension. This column vector is used as the low-dimensional measure.
[0041] Each dimension of this low-dimensional measure is a weighted sum of the original loads in each random projection direction.
[0042] The value of each element in the random projection matrix is generated by reading a pre-stored random seed sequence when the platform starts. Each tunnel identifier corresponds to a different random seed sequence to ensure that the data distribution of different tunnels in the projection space is independent of each other.
[0043] After obtaining the low-dimensional measure, the platform concatenates the tunnel identifier with the current timestamp bit by bit and performs a hash operation to obtain a random seed value. This hash operation uses a secure hash algorithm.
[0044] Starting with this random seed value, the platform sequentially selects several projection directions from a preset projection direction table. Each projection direction is a vector with the same dimension as the low-dimensional measure dimension.
[0045] The projection direction table is constructed as follows: during platform initialization, a direction set containing 256 projection direction vectors is randomly generated. Each element in each projection direction vector is taken from a standard normal distribution, and the projection direction vectors are orthogonalized.
[0046] The number of projection directions is the same as the dimension value of the low-dimensional measure, that is, the number of projection directions is equal to the number of projection dimensions.
[0047] The low-dimensional measure is compared with each selected projection direction, and the binary bits are output bit by bit according to the comparison results.
[0048] For each projection direction, the platform performs a dot product operation between the projection direction vector and the low-dimensional measure vector to obtain a comparison value. If the comparison value is greater than zero, the binary bit 1 is output; if the comparison value is less than or equal to zero, the binary bit 0 is output.
[0049] The comparison and determination rule adopts a symbolic function form, and the calculation expression is as follows: ; in Indicates the first The low-dimensional measure component and the first The binary result output after comparing each projection direction. Represents the first dimension in low-dimensional measures Each dimension value Indicates the first In the projection direction vector, the one that is related to the first The element value at each position in each dimension and The product of is the first Each component contributes a value, and the sum of the component contributions from all dimensions yields the dot product. For sign removal operation, the value is 1 when the dot product result is greater than zero, and 0 when the dot product result is less than or equal to zero.
[0050] The binary bits corresponding to each projection direction are combined and arranged in the order of selection of the projection direction to obtain a binary sequence. This binary sequence is the binary code that maintains distance similarity with the original payload. This distance similarity is guaranteed by the locality-sensitive hashing theory. That is, the closer the distance between the original payload vectors, the closer the Hamming distance between the binary codes obtained after the above projection and binary conversion.
[0051] The low-dimensional measure, binary code, timestamp, and tunnel identifier are encapsulated together into a feature package and written to an independent cache.
[0052] This independent cache is a pre-defined circular buffer in the platform's memory, with a capacity of 1000 feature packets. When the number of feature packets exceeds the capacity, new feature packets overwrite the oldest feature packets.
[0053] Each feature packet is accompanied by a timestamp of the write time when it is written, which is used to identify the time window to which the feature packet belongs when it is read from the cache later.
[0054] This cache is independent of the shared computing pool's memory resources, and its read and write operations do not consume the shared computing pool's computing resources, thus not increasing the load pressure on the shared computing pool.
[0055] In the feature reconstruction and backfilling module, after contention mitigation, feature packets are read from the cache, and approximate feature sequences of the critical packets are reconstructed using measures and sparse basis functions. These sequences, along with the reconstructed confidence scores, are then backfilled into the corresponding window of the detection pipeline. Specifically, this includes: The platform monitors the current load value of the shared computing pool. When the current load value is lower than the low load threshold for three consecutive samples, it is determined that the competition in the shared computing pool has been alleviated. The low load threshold is 30% of the total number of computing units. This value is derived from the platform's load statistics experience value in multi-tunnel concurrent scenarios.
[0056] Once the competition subsides, the platform reads the feature package from an independent cache. This feature package encapsulates a low-dimensional measure, binary code, timestamp, and tunnel identifier, where the binary code is a binary sequence with a fixed length of 256 bits.
[0057] After extracting the binary code from the feature packet, the platform splits the binary code into multiple index segments with a fixed bit width of 4 bits, meaning that every 4 bits is an index segment. The 256-bit binary code is split into a total of 64 index segments.
[0058] The platform converts each index segment from binary to decimal value to obtain the corresponding decimal index value, which ranges from 0 to 15.
[0059] The platform pre-constructs a sparse basis, which is a two-dimensional lookup table with 16 x 64 = 1024 rows and 1024 columns. Each element in the two-dimensional lookup table is taken from a standard normal distribution, and each row vector is normalized by a unit length.
[0060] The sparse basis is constructed as follows: during platform initialization, a random matrix with 1024 rows and 1024 columns is generated. Unit length normalization is performed on each row of the random matrix so that the magnitude of each row vector is equal to 1, thus obtaining the sparse basis.
[0061] Using each decimal index value as the basis for searching, the substrate atoms at the corresponding positions are sequentially extracted from the sparse substrate.
[0062] The specific extraction method is as follows: For the decimal index value corresponding to the first index segment, add 1 to the decimal index value to obtain the first row number, and extract the whole row vector of the first row number from the sparse basis as the first basis atom; for the decimal index value corresponding to the second index segment, add 1 to the decimal index value to obtain the second row number, and extract the whole row vector of the second row number from the sparse basis as the second basis atom; and so on, until the 64th index segment, a total of 64 basis atoms are extracted. The 64 extracted basis atoms are arranged in the order of the index segments and combined into a basis atom combination.
[0063] Each basis atom is a 1024-dimensional row vector, so the combination of basis atoms is essentially a matrix of 64 rows by 1024 columns.
[0064] The platform extracts a low-dimensional measure from the feature package. This low-dimensional measure is a 64-dimensional column vector. The values of each dimension in this low-dimensional measure are used as the synthesis weights of the corresponding base atoms in the base atom combination. The base atoms are then superimposed according to the synthesis weights to generate an approximate feature sequence.
[0065] The specific superposition method is as follows: multiply the first base atom in the base atom combination with the value of the first dimension of the low-dimensional measure to obtain the first weighted base atom; multiply the second base atom with the value of the second dimension of the low-dimensional measure to obtain the second weighted base atom; and so on, until 64 weighted base atoms are obtained. Then, add these 64 weighted base atoms together using vector addition to obtain a 1024-dimensional vector, which is the approximate feature sequence.
[0066] After generating the approximate feature sequence, the platform calculates the deviation between the approximate feature sequence and the original expected load features, where the original expected load features are the standard feature vectors of the tunnel at the corresponding timestamps, which are statistically obtained by the platform based on a large amount of normal traffic data.
[0067] The deviation magnitude is calculated as follows: subtract the dimension values at the same position in the approximate feature sequence from the dimension values at the same position in the original load expected features one by one to obtain the difference of each dimension. Take the sum of the absolute values of the differences of each dimension, divide the sum by the total number of dimensions of the approximate feature sequence, 1024, to obtain the average absolute deviation value, which is the deviation magnitude.
[0068] The platform determines the reconstruction confidence level based on the deviation magnitude. The reconstruction confidence level is calculated by subtracting the ratio of the deviation magnitude to the maximum allowable deviation value from 1. The maximum allowable deviation value is three times the standard deviation of each dimension value in the standard feature vector. This value is based on the principle in statistics that 99.7% of the data under a normal distribution are within the range of the mean plus or minus three standard deviations.
[0069] The platform writes the approximate feature sequence and reconstruction confidence together into the window position corresponding to the timestamp in the detection pipeline, overwriting the original empty mark of the window. The empty mark is a pre-set special value used to indicate that the window was not extracted due to timeout.
[0070] In the comprehensive assessment and response decision-making module, before triggering response, the comprehensive confidence level of the target feature sequence, including the reconstruction confidence level, is assessed. If it is lower than the blocking threshold, it is downgraded to manual review and a confirmation instruction is sent; otherwise, a control instruction is directly issued to the execution device, specifically including: Before triggering automated processing decisions, the platform first reads the feature sequence to be judged from the detection pipeline. This feature sequence consists of feature vectors from multiple time windows arranged in chronological order. The feature vectors of some time windows are high-dimensional feature vectors extracted in real time, while the feature vectors of other time windows are approximate feature sequences reconstructed from feature packets.
[0071] For each reconstructed time window, the reconstruction confidence score is stored within the window. This reconstruction confidence score has been written into the empty marker position of the window during the backfilling step. For time windows that are fully extracted in real time, their confidence score is marked as the full confidence score, with a fixed value of 1.
[0072] The platform iterates through all time windows in the feature sequence to be judged, filters out all windows carrying reconstruction confidence, and extracts the reconstruction confidence values stored in each of these windows.
[0073] These reconstructed confidence scores are collected into a reconstructed confidence score set. All scores in the set are sorted in ascending order, and the score in the middle position after sorting is taken as the overall correction coefficient of the feature sequence to be judged.
[0074] If the number of values in the reconstructed confidence set is even, the average of the two middle values is taken as the overall correction coefficient. The average value is calculated by adding the two middle values and then dividing by 2.
[0075] The platform reads the original confidence level derived from the complete features from the detection results corresponding to the feature sequence to be judged. This original confidence level is the judgment confidence value output by the platform when judging whether the current traffic contains attack behavior based on all complete feature vectors after completing feature extraction for each time window. This value is between 0 and 1.
[0076] The overall correction coefficient is used as a multiplier and multiplied with the original confidence level to obtain a weighted confidence level value. This weighted confidence level value is then used as the overall confidence level of the feature sequence to be judged.
[0077] Under this weighted fusion method, when the feature sequence to be judged contains a large number of reconstructed features, the overall correction coefficient becomes smaller because the reconstruction confidence value is generally lower than the fixed value of 1, which in turn reduces the overall confidence relative to the original confidence.
[0078] The platform compares the overall confidence level with a preset blocking threshold. This threshold is set based on a sample set of historical operational data that has been confirmed as genuine attack events and false alarms. The platform statistically analyzes the distribution curves of the correct handling rate and false handling rate under different thresholds, and takes the threshold corresponding to the largest difference between the correct and false handling rates as the blocking threshold. This blocking threshold is set to 0.75. If the overall confidence level reaches or exceeds 0.75, the platform directly issues a control command to the executing device. This control command includes the five-tuple information of the traffic to be blocked and the blocking duration. If the overall confidence level is below 0.75, the platform does not send a control command but instead outputs a confirmation command to the administrator. This confirmation command carries a summary of the feature sequence to be judged and the decision-making basis, and the automated handling process is paused until manual feedback is received from the administrator.
[0079] The working principle of this invention is as follows: Data packets accessed via tunnels are acquired, and scheduling tags are attached to them based on their tenant priority and tunnel identifier. Based on these scheduling tags, the current load status of the shared computing pool is continuously monitored, and the historical latency of preceding data packets within the same tunnel is dynamically tracked. When the load offset exceeds a set offset threshold and the latency rate shows an accelerating growth trend, the current data packet is marked as a critical packet. Deep feature extraction is stopped for critical packets, and the original load is randomly projected into a low-dimensional measure, generating a binary code that maintains distance similarity. The low-dimensional measure, binary code, timestamp, and tunnel identifier are encapsulated into a feature packet and stored in an independent cache. The shared computing pool is then... After pool contention is alleviated, feature packets are read from the cache. Using binary code as an index, basis atom combinations are extracted from the sparse basis. The values of each dimension in the low-dimensional measure are used as synthetic weights for linear superposition to reconstruct the approximate feature sequence of the critical packet. The reconstruction confidence is determined based on the deviation between the approximate feature sequence and the expected features. The reconstruction confidence is then backfilled into the corresponding window of the detection pipeline. Before triggering automated processing, the median value of all reconstruction confidences in the feature sequence to be judged is extracted as the overall correction coefficient. This is then weighted and fused with the original confidence to obtain the comprehensive confidence. If the comprehensive confidence is lower than the blocking threshold, it is downgraded to manual review; otherwise, a control command is issued to the execution device.
[0080] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.
Claims
1. An enterprise network security operation and hosting system based on artificial intelligence, characterized in that: include: The traffic access and label generation module acquires data packets accessed via the tunnel and attaches scheduling labels to them based on the tenant priority and tunnel identifier. The load awareness and critical packet determination module continuously senses the current load status of the shared computing pool based on scheduling tags and dynamically tracks the historical retention time of preceding data packets in the same tunnel. If the current load status exceeds the high load threshold and the actual processing time of multiple consecutive preceding data packets in the same tunnel reaches the preset safety threshold, the current data packet is marked as a critical packet. The lightweight trace and feature caching module stops deep extraction on critical packets and performs lightweight trace: the original payload is randomly projected into a low-dimensional measure and a distance-preserving binary code is generated. The measure, binary code, timestamp, and tunnel identifier are encapsulated into a feature packet and stored in an independent cache. The feature reconstruction and backfilling module reads the feature package from the cache after the competition is relieved, reconstructs the approximate feature sequence of the critical package using the measure and sparse basis, and backfills it into the corresponding window of the detection pipeline along with the reconstruction confidence. The comprehensive assessment and handling decision module evaluates the comprehensive confidence of the feature sequence to be judged, including the reconstruction confidence, before triggering the handling. If it is lower than the blocking threshold, it is downgraded to manual review and a confirmation instruction is sent. Otherwise, a control instruction is directly issued to the execution device.
2. The enterprise network security operation and hosting system based on artificial intelligence according to claim 1, characterized in that, The process of attaching scheduling tags to tenants based on their priority and tunnel identifier specifically includes: Obtain the tunnel identifier carried by the data packet, query the tunnel's historical packet loss rate and latency fluctuation value within the current time window, and calculate the tunnel's current stability score based on the historical packet loss rate and latency fluctuation value; The weighting coefficients of tenant priorities are adjusted based on the current stability score to obtain dynamic priorities; The dynamic priority is associated with the tunnel identifier and written into the label field of the data packet as a scheduling label.
3. The enterprise network security operation and hosting system based on artificial intelligence according to claim 1, characterized in that, The step of marking the current data packet as a critical packet specifically includes: Based on the tunnel identifier recorded in the scheduling tag, retrieve the dwell time of each of the preceding data packets of the tunnel before the current time, and calculate the rate of change of the dwell time. The system senses the current load value of the shared computing pool and compares it with the historical average load value of the tunnel to obtain the degree of load offset. When the load offset exceeds the set offset threshold and the rate of change shows an accelerating growth trend, the current data packet is marked as a critical packet.
4. The enterprise network security operation hosting system based on artificial intelligence according to claim 1, characterized in that, The storage in the independent cache specifically includes: Read the load length value of the critical package, dynamically determine the projection dimension based on the load length value, and perform random projection on the original load according to the projection dimension to obtain a low-dimensional measure; Using the tunnel identifier as a component of the random seed, a binary transformation is performed on the low-dimensional measure to generate a binary code that maintains distance similarity to the original payload; The low-dimensional measure, binary code, timestamp, and tunnel identifier are encapsulated together into a feature package and written to an independent cache.
5. The enterprise network security operation and hosting system based on artificial intelligence according to claim 4, characterized in that, The generation of binary code that maintains distance similarity to the original payload specifically includes: The tunnel identifier and the current timestamp are concatenated bit by bit, and then a hash operation is performed to obtain a random seed value; Starting with a random seed value, several projection directions are sequentially selected from a preset projection direction table; The low-dimensional measure is compared with each selected projection direction. Based on the comparison results, the binary bits are output bit by bit, and the binary bits are combined to obtain the binary code.
6. The enterprise network security operation hosting system based on artificial intelligence according to claim 1, characterized in that, The process of reconstructing confidence levels and backfilling them to the corresponding window in the detection pipeline specifically includes: Low-dimensional measure and binary code are extracted from the feature package. The binary code is used as an index to find matching basis atom combinations in the sparse basis. The low-dimensional measure and the found basis atom combinations are linearly synthesized to obtain an approximate feature sequence. Calculate the deviation between the approximate feature sequence and the expected features of the original payload at the timestamp recorded in the feature package, and determine the reconstruction confidence based on the deviation. The approximate feature sequence and the reconstructed confidence are written together into the window position corresponding to the timestamp in the detection pipeline, overwriting the original empty marker of the corresponding window.
7. The enterprise network security operation and hosting system based on artificial intelligence according to claim 6, characterized in that, The process of generating the approximate feature sequence is as follows: The binary code is divided into multiple index segments with a fixed bit width, and each index segment corresponds to a decimal index value. Based on the decimal index value, the substrate atoms at the corresponding positions are sequentially extracted from the sparse substrate, and all the extracted substrate atoms are combined into a substrate atom combination according to the extraction order. Using the values of each dimension in the low-dimensional measure as the synthesis weights of the corresponding atoms in the base atom combination, the atoms are superimposed according to the synthesis weights to generate an approximate feature sequence.
8. The enterprise network security operation and hosting system based on artificial intelligence according to claim 1, characterized in that, The direct issuance of control commands to the execution device specifically includes: Extract all reconstructed confidence scores carried in the feature sequence to be judged, and take the median value of all reconstructed confidence scores as the overall correction coefficient of the feature sequence to be judged; The original confidence scores derived from the complete features in the feature sequence to be judged are weighted and fused with the overall correction coefficient to obtain the comprehensive confidence score. The overall confidence level is compared with the blocking threshold. If the threshold is reached or exceeded, a control command is directly sent to the execution device. Otherwise, a confirmation command is output and the automatic processing process is suspended.
