Intrusion prevention methods and systems based on cloud desktop traffic analysis and AI prediction
By acquiring and decoupling multi-source features in a cloud desktop environment, constructing an attack prediction field, and generating an adversarial defense scheme, the problem of defense lag and feature fragmentation in existing technologies is solved, enabling early identification and proactive defense against complex attacks and improving cloud desktop security.
Patent Information
- Application Number
- CN202611182857.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-05
- Publication Date
- 2026-11-17
AI Technical Summary
Existing cloud desktop environment intrusion prevention technologies lack the ability to dynamically predict multi-step penetration attacks, cannot effectively utilize the collaborative value of multi-source heterogeneous data, and the defense measures lag behind the attack evolution process, making it difficult to identify attack intentions and future behavioral trajectories.
By acquiring the communication behavior, process execution logic, and resource contention mode characteristics of the cloud desktop environment, cross-view semantic decoupling is performed, independent and coupled feature subspaces are constructed, mapped to the attack target-means binary semantic graph, the remaining attack steps are identified, and an attack evolution prediction field is constructed to generate an adversarial defense scheme.
It enables early and accurate prediction and proactive defense of attack behaviors in cloud desktop environments, significantly improving the recognition rate and prediction lead of complex attack chains, dynamically generating the probability distribution of being compromised and the heat map of attack intensity, and possessing adaptive adversarial capabilities, reducing the losses and spread caused by attacks.
Smart Images

Figure CN123122879A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of network security technology, and in particular to intrusion prevention methods and systems based on cloud desktop traffic analysis and AI prediction. Background Technology
[0002] In cloud desktop environments, existing intrusion prevention technologies typically rely on traditional network security protection systems, including rule-based intrusion detection systems, perimeter firewalls, and virtual machine-level access control lists. These solutions deploy static rules against known attack patterns, monitor communication behavior at thresholds, audit process execution logs, and employ isolation strategies for resource contention. Some systems introduce machine learning models to analyze traffic anomalies, but the training data often comes from single-modal features, such as using only network packet statistics or system call sequences. Defense response mechanisms primarily rely on blocking, alerts, and manual forensics, lacking the ability to infer attacker intent and future behavioral patterns.
[0003] Current conventional methods have significant drawbacks. First, defense decisions lack the ability to dynamically predict multi-step penetration attacks. Attackers often gradually compromise virtual desktop instances through progressive lateral movement, while traditional detection methods can only identify single events and cannot link attack steps across virtual machines to form a complete attack chain, causing defense measures to lag behind the attack evolution process. Second, feature extraction dimensions are fragmented, failing to effectively utilize the collaborative value of multi-source heterogeneous data in the cloud desktop environment. Communication behaviors, process logic, and resource contention patterns occur asynchronously on the timeline. Existing methods typically analyze each view independently, ignoring the implicit coupling relationships and semantic associations between different modalities. This makes attack intent inference susceptible to noise interference, making it difficult to accurately identify remaining attack paths and high-value targets. Summary of the Invention
[0004] This invention provides an intrusion prevention method and system for cloud desktop traffic analysis and AI prediction, which can solve the problems in the prior art.
[0005] A first aspect of the present invention provides an intrusion prevention method based on cloud desktop traffic analysis and AI prediction, comprising:
[0006] The system acquires communication behavior characteristics, process execution logic characteristics, and resource contention mode characteristics of the cloud desktop environment. After aligning them on a unified timeline, it performs cross-view semantic decoupling, extracts the coupling components of each view, and constructs independent feature subspaces and coupled feature subspaces.
[0007] The coupling components are mapped to the attack target-means binary semantic graph, and the set of target state nodes reachable by the current traffic behavior in the attack target-means binary semantic graph and the corresponding shortest means path are identified to obtain the intent reasoning result containing the remaining attack steps;
[0008] Based on the remaining attack steps and the shortest means path, and combining the vulnerability fingerprints of each virtual desktop instance in the independent feature subspace with the resource dependency tensor in the coupled feature subspace, an attack evolution prediction field of the attack behavior on the virtual desktop topology is constructed. Based on the attack evolution prediction field, the probability distribution of the compromise of each virtual desktop instance and the attack intensity heatmap at multiple future time points are obtained.
[0009] The probability distribution of being compromised and the heatmap of attack intensity are used as the game state space, the virtualized defense operation of the defender is used as the strategy space, and the ability of the attacker to adjust the attack path according to the defense situation is modeled as the adversarial strategy. An adversarial defense scheme is generated by solving the Nash equilibrium solution.
[0010] The characteristics of communication behavior, process execution logic, and resource contention patterns in the cloud desktop environment include:
[0011] The network traffic data streams of multiple virtual desktop instances in the cloud desktop environment are stripped layer by layer according to the encapsulation layer to obtain the initial field sequence of each protocol layer. A field dependency graph is constructed for the initial field sequence of each protocol layer. The field dependency graph is traversed to extract the protocol behavior sequence reflecting the communication session establishment mode, data transmission mode and session termination mode as the communication behavior feature.
[0012] The input parameters and return values of each system call in the system call sequence of multiple virtual desktop instances in the cloud desktop environment are constructed into a call data flow graph. The data object subgraph that forms a closed loop and the data object transfer path that crosses the process boundary are identified in the call data flow graph. The closed loop path in the data object subgraph is mapped to the intra-process call dependency chain, and the data object transfer path is mapped to the inter-process call dependency chain. The intra-process call dependency chain and the inter-process call dependency chain are merged to obtain a complete call dependency chain that reflects the inter-process cooperation logic and resource competition logic as the process execution logic feature.
[0013] Based on the time-series matching of virtual machine resource allocation events recorded in the resource scheduling log of the virtual machine management layer in the cloud desktop environment with the resource request timestamps in the inter-process call dependency chain, the time-series coupling pattern between virtual desktop instances caused by resource scheduling is extracted as the feature of the resource competition mode.
[0014] After alignment on a unified timeline, cross-view semantic decoupling is performed, extracting the coupling components of each view, and constructing independent feature subspaces and coupled feature subspaces, including:
[0015] After aligning the communication behavior features, process execution logic features, and resource contention mode features on a unified time axis, a cross-view feature association matrix is constructed.
[0016] By performing a rank decomposition operation on the cross-view feature association matrix, the cross-view feature association matrix is decomposed into a low-rank component that represents the shared semantics between views and a sparse component that represents the unique attributes of each view.
[0017] Based on the multi-view feature collaboration rules of legitimate business behaviors in the cloud desktop environment, constraint boundaries are determined to form a semantic constraint subspace;
[0018] The low-rank components are projected onto the semantic constraint subspace. By calculating the projection residuals of the low-rank components in the semantic constraint subspace, abnormal coupling patterns that violate the rules of legitimate business collaboration are identified. The associated components reflecting resource sharing and call dependencies between virtual desktop instances in the abnormal coupling patterns are extracted to form a resource dependency tensor as the coupling component. Tensor expansion is performed on the resource dependency tensor and the principal component directions are extracted as the basis vectors of the coupling feature subspace.
[0019] The feature components reflecting the configuration and exposure surface of each virtual desktop instance in the sparse component are extracted as vulnerability fingerprints. The vulnerability fingerprints are grouped according to the virtual desktop instance identifiers and orthogonalized to obtain the basis vectors of the independent feature subspace.
[0020] Mapping the coupling components to an attack target-means binary semantic graph, identifying the set of target state nodes reachable from the current traffic behavior in the attack target-means binary semantic graph and the corresponding shortest means path, yields an intent reasoning result including the remaining attack steps, including:
[0021] Historical attack events are acquired, and the attack targets in the historical attack events are abstracted into target state nodes and attack methods are abstracted into method operation nodes. The preconditions and post-effects between the attack targets and the attack methods are analyzed. A directed edge is established between the target state nodes and the method operation nodes. The directed edge carries the weight representing the degree to which the method operation satisfies the target state and the constraint conditions representing the preconditions of the resources required to execute the method operation, thus forming the attack target-method binary semantic graph.
[0022] Map the coupling components to the means operation nodes of the attack target-means binary semantic graph, calculate the matching degree between the coupling components and the constraints, and mark the currently executable means operation nodes;
[0023] Starting from the currently executable means operation node, the constraint satisfaction propagation algorithm is executed, propagating forward along the directed edge and accumulating the weights and constraints. When the propagation path reaches the target state node and all constraints are satisfied, it is marked as a reachable target state node and the propagation path is recorded as the shortest means path. The remaining attack steps are extracted based on the sequence of means operation nodes that have not been executed in the shortest means path to form the intent inference result.
[0024] Based on the remaining attack steps and the shortest means path, and combining the vulnerability fingerprints of each virtual desktop instance in the independent feature subspace with the resource dependency tensor in the coupled feature subspace, an attack evolution prediction field for the attack behavior on the virtual desktop topology is constructed, including:
[0025] Extract the vulnerability type required by the attack method corresponding to each means operation node in the remaining attack steps, and calculate the semantic similarity with the configuration exposure surface features recorded in the vulnerability fingerprint to obtain the matching degree;
[0026] Based on the matching degree, each virtual desktop instance is marked with a feasibility score as the target of the means operation node, and the feasibility score is assigned to the corresponding instance node in the virtual desktop topology to form an initial attack reachability distribution;
[0027] Resource sharing and call dependency relationships between virtual desktop instances are extracted from the resource dependency tensor in the coupled feature subspace and mapped to form propagation edges between instance nodes in the virtual desktop topology. The propagation coefficient of the propagation edge is determined by the value of the corresponding element in the resource dependency tensor.
[0028] The initial attack reachability distribution is used as the initial condition to perform a diffusion propagation process along the propagation edge. During the diffusion propagation process, the attack reachability value of each instance node is updated and iterated according to the propagation coefficient and the attack reachability value of the adjacent nodes until convergence. The converged attack reachability value distribution is temporally correlated with the means operation node sequence in the shortest means path to form the attack evolution prediction field.
[0029] Based on the attack evolution prediction field, the probability distribution of the virtual desktop instance being compromised and the heatmap of attack intensity at multiple future time points are obtained, including:
[0030] The attack reachability values of each instance node in the virtual desktop topology at different time steps are extracted from the attack evolution prediction field. The attack reachability values are then weighted and fused with the attack success rate of the means operation node at the corresponding time step in the shortest means path. The attack success rate is obtained by statistically analyzing the historical successful execution records of the attack means corresponding to the means operation node. The probability of each virtual desktop instance being compromised at multiple future time points is obtained through the weighted fusion calculation. The probability of being compromised is then organized into the probability distribution of being compromised by time and space dimensions.
[0031] Calculate the rate of change of the probability of compromise of each virtual desktop instance at each time point in the compromise probability distribution, and multiply the rate of change of the probability of compromise with the attack reachability value of the corresponding instance node in the attack evolution prediction field to obtain the attack intensity value. The attack intensity value characterizes the urgency of the virtual desktop instance being attacked at a specific time point.
[0032] The attack strength value is mapped to a visual representation of the virtual desktop topology and different color codes are assigned according to the magnitude of the attack strength value to form an attack strength heatmap. The color depth in the attack strength heatmap is proportional to the attack strength value.
[0033] Using the probability distribution of being compromised and the heatmap of attack intensity as the game state space, the virtualized defense operations of the defender as the strategy space, and the attacker's ability to adjust the attack path according to the defense situation as the adversarial strategy, the adversarial defense scheme is generated by solving the Nash equilibrium solution, including:
[0034] The game state space is formed by vectorizing the probability distribution of being compromised and the heatmap of attack intensity to form a game state vector.
[0035] The virtualization defense operations of the defender are defined as including virtual desktop instance migration operations, resource isolation operations, and access control policy adjustment operations. Each virtualization defense operation corresponds to a defense action vector, which describes the influence coefficient of the virtualization defense operation on the game state vector. All defense action vectors are combined to form the policy space.
[0036] Based on the shortest means path and the resource dependency tensor, the attacker's countermeasure strategy is constructed. The countermeasure strategy takes the defensive action vector in the strategy space and the game state vector in the game state space as input, and outputs the attacker's adjusted attack target and attack path.
[0037] A two-player zero-sum game is constructed between the defender and the attacker. The payoff of the two-player zero-sum game is defined by the weighted difference between the decrease in the probability of being attacked and the resource expenditure. The defensive action vector sequence that maximizes the payoff is obtained by iteratively solving the game state space and the strategy space as the Nash equilibrium solution. The virtualized defensive operation sequence corresponding to the Nash equilibrium solution is used as the adversarial defense scheme.
[0038] A second aspect of this invention provides an intrusion prevention system for cloud desktop traffic analysis and AI prediction, comprising:
[0039] The feature decoupling unit is used to obtain the communication behavior features, process execution logic features and resource contention mode features of the cloud desktop environment, and after aligning them on a unified time axis, perform cross-view semantic decoupling, extract the coupling components of each view, and construct independent feature subspaces and coupled feature subspaces.
[0040] The intent reasoning unit is used to map the coupled components to the attack target-means binary semantic graph, identify the set of target state nodes that the current traffic behavior can reach in the attack target-means binary semantic graph and the corresponding shortest means path, and obtain the intent reasoning result containing the remaining attack steps.
[0041] An evolution prediction unit is used to construct an attack evolution prediction field on the virtual desktop topology based on the remaining attack steps and the shortest means path, combined with the vulnerability fingerprint of each virtual desktop instance in the independent feature subspace and the resource dependency tensor in the coupled feature subspace. Based on the attack evolution prediction field, the probability distribution of the compromise of each virtual desktop instance and the attack intensity heatmap at multiple future time points are obtained.
[0042] The game-theoretic defense unit is used to take the probability distribution of being compromised and the heat map of attack intensity as the game state space, the virtualized defense operation of the defender as the strategy space, and the ability of the attacker to adjust the attack path according to the defense situation as the adversarial strategy. The adversarial defense scheme is generated by solving the Nash equilibrium solution.
[0043] A third aspect of the present invention provides an electronic device, comprising:
[0044] processor;
[0045] Memory used to store processor-executable instructions;
[0046] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.
[0047] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.
[0048] The beneficial effects of this application are as follows:
[0049] The method provided by this invention enables early and accurate prediction and proactive defense against attacks within cloud desktop environments. By decoupling cross-view semantics and constructing independent and coupled feature subspaces, it deeply extracts shared attack patterns and unique vulnerability information from multiple views, making attack intent reasoning no longer dependent on a single feature plane. This significantly improves the recognition rate and prediction lead time for complex and covert attack chains. The introduction of the attack target-means binary semantic graph and intent reasoning mechanism allows for precise mapping between current traffic behavior and residual attack steps, enabling defenders to anticipate the attacker's next move rather than reacting passively.
[0050] Based on the attack evolution prediction field constructed using residual attack steps, vulnerability fingerprints, and resource dependency tensors, this system can dynamically generate the probability distribution of compromise and attack intensity heatmaps for each virtual desktop instance at multiple future time points. This prediction field integrates resource competition and dependencies within the virtual desktop topology, accurately depicting the lateral movement trend of attacks between parallel desktops and the law of attack effectiveness decay. This spatiotemporal predictive capability allows for the proactive deployment of defense resources along the highest-risk paths, significantly reducing the losses and spread caused by actual attacks.
[0051] By directly using the predicted probability distribution and heatmap as the game state space, and empowering virtualized defense operations and attacker path adjustments as adversarial strategies, the defense scheme generated by solving the Nash equilibrium possesses adaptive adversarial capabilities. The scheme not only considers the defender's current optimal strategy but also predicts the attacker's optimal adjustment path under different defense postures, thus forming a dynamic game equilibrium. This equilibrium solution resists attackers' attempts to proactively evade the defense plan, ensuring the defense effect remains stable and efficient in long-term adversarial operations, greatly enhancing the overall security resilience of the cloud desktop environment. Attached Figure Description
[0052] Figure 1 A flowchart illustrating an intrusion prevention method based on cloud desktop traffic analysis and AI prediction;
[0053] Figure 2 A schematic diagram of the process for forming an adversarial defense scheme. Detailed Implementation
[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.
[0055] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.
[0056] Figure 1 This is a flowchart illustrating the intrusion prevention method based on cloud desktop traffic analysis and AI prediction, as described in an embodiment of the present invention. Figure 1 As shown, the method includes:
[0057] The system acquires communication behavior characteristics, process execution logic characteristics, and resource contention mode characteristics of the cloud desktop environment. After aligning them on a unified timeline, it performs cross-view semantic decoupling, extracts the coupling components of each view, and constructs independent feature subspaces and coupled feature subspaces.
[0058] The coupling components are mapped to the attack target-means binary semantic graph, and the set of target state nodes reachable by the current traffic behavior in the attack target-means binary semantic graph and the corresponding shortest means path are identified to obtain the intent reasoning result containing the remaining attack steps;
[0059] Based on the remaining attack steps and the shortest means path, and combining the vulnerability fingerprints of each virtual desktop instance in the independent feature subspace with the resource dependency tensor in the coupled feature subspace, an attack evolution prediction field of the attack behavior on the virtual desktop topology is constructed. Based on the attack evolution prediction field, the probability distribution of the compromise of each virtual desktop instance and the attack intensity heatmap at multiple future time points are obtained.
[0060] The probability distribution of being compromised and the heatmap of attack intensity are used as the game state space, the virtualized defense operation of the defender is used as the strategy space, and the ability of the attacker to adjust the attack path according to the defense situation is modeled as the adversarial strategy. An adversarial defense scheme is generated by solving the Nash equilibrium solution.
[0061] In one optional implementation, obtaining the communication behavior characteristics, process execution logic characteristics, and resource contention mode characteristics of the cloud desktop environment includes:
[0062] The network traffic data streams of multiple virtual desktop instances in the cloud desktop environment are stripped layer by layer according to the encapsulation layer to obtain the initial field sequence of each protocol layer. A field dependency graph is constructed for the initial field sequence of each protocol layer. The field dependency graph is traversed to extract the protocol behavior sequence reflecting the communication session establishment mode, data transmission mode and session termination mode as the communication behavior feature.
[0063] The input parameters and return values of each system call in the system call sequence of multiple virtual desktop instances in the cloud desktop environment are constructed into a call data flow graph. The data object subgraph that forms a closed loop and the data object transfer path that crosses the process boundary are identified in the call data flow graph. The closed loop path in the data object subgraph is mapped to the intra-process call dependency chain, and the data object transfer path is mapped to the inter-process call dependency chain. The intra-process call dependency chain and the inter-process call dependency chain are merged to obtain a complete call dependency chain that reflects the inter-process cooperation logic and resource competition logic as the process execution logic feature.
[0064] Based on the time-series matching of virtual machine resource allocation events recorded in the resource scheduling log of the virtual machine management layer in the cloud desktop environment with the resource request timestamps in the inter-process call dependency chain, the time-series coupling pattern between virtual desktop instances caused by resource scheduling is extracted as the feature of the resource competition mode.
[0065] In a cloud desktop environment, multiple virtual desktop instances run simultaneously, generating network traffic characterized by multi-protocol encapsulation and multi-layered nesting. To extract communication behavior characteristics, the network traffic data stream generated by each virtual desktop instance is peeled away layer by layer according to its encapsulation level. Starting from the outermost Ethernet frame, the process proceeds inward, parsing the IP layer, transport layer (TCP / UDP), and application layer protocols (such as RDP, VNC, HTTP, and other commonly used cloud desktop protocols), until the initial field sequence of each protocol layer is obtained. The field sequence of each layer contains all field values from the protocol header of that layer and their temporal arrangement. Examples include TCP layer fields such as sequence number, acknowledgment number, window size, and flags, as well as application layer fields such as session control fields and data payload length.
[0066] After obtaining the initial field sequences of each protocol layer, a field dependency graph is constructed for each layer. The field dependency graph uses field nodes as vertices and directed edges representing numerical constraints, temporal sequences, or logical implications between fields. For example, there is an incrementing numerical dependency between the TCP sequence number field and the acknowledgment number field, and a temporal dependency between the SYN flag during connection establishment and the subsequent ACK flag. The constructed field dependency graph is traversed, and field change sequences reflecting the communication session establishment mode, data transmission mode, and session termination mode are identified along the propagation path of the dependency edges. The session establishment mode corresponds to the field sequence change pattern during the three-way handshake phase; the data transmission mode corresponds to the throughput, packet interval, and window dynamic changes during the continuous data frame interaction phase; and the session termination mode corresponds to the field sequence characteristics of termination signals such as FIN / RST. The field change sequences corresponding to these three types of modes are integrated into a protocol behavior sequence, which is output as the communication behavior feature.
[0067] The extraction of process execution logic characteristics is based on system call sequences as the primary data source. In the cloud desktop environment, the operating system kernel of each virtual desktop instance records the system call sequence initiated by the process. Each record includes the system call name, call timestamp, input parameter list, and return value. The input parameters and return value of each system call are abstracted as data object nodes. If the return value of a call is used as an input parameter for a subsequent call, a directed passing edge is established between the two data object nodes, thus constructing an inter-call data flow graph. This graph fully depicts the data flow path in the system call sequence, revealing the data production and consumption relationships within a process, as well as data sharing and transfer behaviors across processes.
[0068] In the inter-call data flow graph, two key structures are identified. The first type is a subgraph of data objects forming a closed loop, i.e., a subgraph containing directed cycles. This structure reflects circular references and continuous operations on a resource or data object within a process, such as a process repeatedly reading and writing the same file handle or repeatedly using the same memory buffer. The directed cycles in the closed-loop subgraph are expanded into intra-process call dependency chains, where each node represents a system call and edges represent data dependencies, fully describing the call logic sequence within the process. The second type is a data object transfer path that crosses process boundaries, i.e., a directed path whose starting point belongs to a process. The endpoint belongs to the process. ( The data transfer is accomplished via inter-process communication mechanisms such as pipes, shared memory, and sockets. These paths are mapped to inter-process call dependency chains, which contain write operation call nodes from the source process, inter-process communication mechanism nodes, and read operation call nodes from the target process.
[0069] By merging intra-process call dependency chains and inter-process call dependency chains in timestamp order, a complete call dependency chain is obtained. This chain includes not only the call logic within a single process but also the cooperation logic and resource contention logic between multiple processes. For example, when multiple virtual desktop instance processes simultaneously compete to access the same shared storage resource, their call dependency chains will contain overlapping write operation nodes for the same resource object, forming a resource contention characteristic fragment. The complete call dependency chain, as a characteristic of process execution logic, provides a structured description of process behavior for subsequent semantic decoupling and attack intent inference.
[0070] Extracting the characteristics of resource contention patterns requires joint analysis of the resource scheduling logs of the hypervisor and inter-process call dependency chains. The hypervisor's resource scheduling logs record resource allocation events for each virtual desktop instance across dimensions such as CPU, memory, storage I / O, and network bandwidth. Each event record includes the event type (allocation, release, preemption), resource type, resource quantity, and event timestamp. And the virtual desktop instance identifiers involved. Each resource request node in the inter-process call dependency chain also carries a resource request timestamp. Resource type identifier.
[0071] The core of time-series matching lies in aligning two types of timestamp sequences on a unified timeline for identification. and The temporal relationship between them. If a process of a virtual desktop instance is in Initiate resource requests at any time The request, and the virtual machine management layer in time( ) resources If a virtual desktop instance preempts another virtual desktop instance and allocates it to this instance, it is considered that there is a temporal coupling relationship between the two, and the corresponding timing difference is... This reflects the response delay in resource contention. By systematically traversing the above-mentioned timing matching process among multiple virtual desktop instances, the timing coupling patterns generated by resource scheduling among virtual desktop instances are extracted, including typical categories such as periodic resource contention patterns, sudden resource preemption patterns, and long-term resource starvation patterns. These timing coupling patterns are output as resource contention pattern features.
[0072] The three types of features mentioned above—communication behavior features, process execution logic features, and resource contention pattern features—characterize the behavior of virtual desktop instances in a cloud desktop environment from three dimensions: network communication view, process execution view, and resource scheduling view, respectively. This provides comprehensive multi-view feature input for subsequent cross-view semantic decoupling on a unified timeline. Communication behavior features focus on describing externally observable traffic patterns, process execution logic features focus on describing the dependency structure of internal execution logic, and resource contention pattern features reveal the resource contention relationships among multiple instances from the virtualization infrastructure level. Together, these three features form the feature foundation for a comprehensive analysis of cloud desktop attack behavior.
[0073] In one optional implementation, after alignment on a unified timeline, cross-view semantic decoupling is performed, extracting the coupling components of each view, and constructing independent feature subspaces and coupled feature subspaces, including:
[0074] After aligning the communication behavior features, process execution logic features, and resource contention mode features on a unified time axis, a cross-view feature association matrix is constructed.
[0075] By performing a rank decomposition operation on the cross-view feature association matrix, the cross-view feature association matrix is decomposed into a low-rank component that represents the shared semantics between views and a sparse component that represents the unique attributes of each view.
[0076] Based on the multi-view feature collaboration rules of legitimate business behaviors in the cloud desktop environment, constraint boundaries are determined to form a semantic constraint subspace;
[0077] The low-rank components are projected onto the semantic constraint subspace. By calculating the projection residuals of the low-rank components in the semantic constraint subspace, abnormal coupling patterns that violate the rules of legitimate business collaboration are identified. The associated components reflecting resource sharing and call dependencies between virtual desktop instances in the abnormal coupling patterns are extracted to form a resource dependency tensor as the coupling component. Tensor expansion is performed on the resource dependency tensor and the principal component directions are extracted as the basis vectors of the coupling feature subspace.
[0078] The feature components reflecting the configuration and exposure surface of each virtual desktop instance in the sparse component are extracted as vulnerability fingerprints. The vulnerability fingerprints are grouped according to the virtual desktop instance identifiers and orthogonalized to obtain the basis vectors of the independent feature subspace.
[0079] After collecting characteristics of communication behavior, process execution logic, and resource contention patterns, these three types of features need to be aligned to the same timeline for time sequence alignment. Since the three types of features originate from different acquisition levels and have varying timestamp precision and sampling frequencies, a global monotonic clock is used as a reference to interpolate and resample the feature sequences of each view, ensuring that cross-view features at the same moment can correspond frame-by-frame. After time sequence alignment, the feature vectors of the three views are concatenated along the time dimension to construct a cross-view feature correlation matrix. ,in This represents the total dimension of the three types of view features. This represents the number of sampled frames after timing alignment. Matrix Each column corresponds to a snapshot of the joint features of the full view at a certain moment, and each row corresponds to the trend of change of a certain feature dimension over time, thus characterizing the semantic association across views in matrix form.
[0080] Cross-view feature association matrix Perform rank decomposition to decompose it into lower-rank components. With sparse components That is, satisfying . low rank component Captures the shared semantic structure among multiple views, reflecting the collaborative patterns of communication behavior, process execution, and resource contention in normal business scenarios; sparse components This preserves attribute information unique to each view and not shared by other views, reflecting local anomalies or individual characteristics within a single view. Rank decomposition, by simultaneously applying nuclear norm constraints and... Norm constraints are used to achieve this: To impose a minimization constraint on the nuclear norm of a given number to induce its low rank, To minimize the sum of absolute values of each element, a constraint is imposed to promote sparsity. The solution process employs the alternating direction multiplier method (ADMM) iteratively, with each iteration involving the sum of absolute values of each element. and Perform singularity threshold shrinkage and soft threshold shrinkage operations until the update magnitude of both is lower than the preset convergence threshold.
[0081] The construction of the semantic constraint subspace relies on the collaborative patterns of historical multi-view features of legitimate business behaviors in the cloud desktop environment. Specifically, cross-view joint features are extracted from historical legitimate session data, the typical distribution range of low-rank components in legitimate scenarios is statistically analyzed, principal component analysis is used to extract the principal directions of legitimate low-rank components, and the linear subspace spanned by these principal directions is defined as the semantic constraint subspace. This subspace characterizes the normal range of changes in legitimate business at the multi-view collaboration level. Its boundary is determined by the variance coverage of legitimate low-rank components in historical data. Typically, the number of principal components that can cover more than 95% of the variance of legitimate samples is selected as the subspace dimension, thereby ensuring the rationality and completeness of the constraint boundary.
[0082] The low-rank components obtained from rank decomposition Projected to semantically constrained subspace Calculate the projection residual ,in express In subspace Orthogonal projection onto the surface. Residual. Frobenius norm This measures the degree to which the current low-rank components deviate from the rules of legitimate business collaboration. When Exceeding the preset threshold At that time, it was determined that an abnormal coupling pattern existed that violated the rules of legitimate business collaboration. Further analysis of the residual matrix... Column-oriented analysis was performed to identify at which times and between which pairs of virtual desktop instances significant anomalous coordination deviations occurred. Correlation components reflecting resource sharing and invocation dependencies between virtual desktop instances were extracted from the anomalous coupling patterns, and these correlation components were organized into third-order tensors according to instance, resource type, and time dimensions. ,in Indicates the number of virtual desktop instances. Indicates the number of resource types. The number of time-series frames is represented by this tensor, which is the resource dependency tensor and serves as a structured representation of the coupled components.
[0083] Resource dependency tensor Perform a tensor expansion operation, expanding it into a matrix along mode-1 (instance dimension). Then perform truncated singular value decomposition on the expanded matrix to extract the previous singular values. The left singular vectors are used as basis vectors of the coupled eigenspace, where The number of singular vectors is determined by the cumulative singular value energy percentage, typically taken as the number of singular vectors corresponding to an energy percentage of 90%. The subspace spanned by this set of basis vectors is the coupling feature subspace, which can compactly express the main changing patterns of resource dependencies between virtual desktop instances in a low-dimensional space, providing a structured coupling feature representation for subsequent attack evolution prediction.
[0084] Sparse components It contains feature information unique to each view, among which feature components related to the configuration and exposure surface of each virtual desktop instance are extracted as vulnerability fingerprints. Specifically, the sparse components... Grouping by virtual desktop instance identifier, and aggregating feature rows belonging to the same instance to form the original vulnerability feature vector of that instance. subscript Instance numbering. Vulnerability feature vector. It covers dimensions such as the distribution of open ports, historical vulnerability trigger records, frequency of abnormal process startup, and deviation of system calls, comprehensively depicting the attack exposure surface and inherent vulnerability of the instance in its current running state.
[0085] Vulnerability feature vector set for each instance Orthogonalization is performed by projecting and deduplicating these vectors sequentially using the Gram-Schmidt orthogonalization method, eliminating redundant components between vulnerability features of different instances, and obtaining a set of standard orthogonal basis vectors. The subspace spanned by this set of orthogonal basis vectors is the independent feature subspace. Its basis vectors guarantee the orthogonal separability of vulnerability features of each virtual desktop instance in the subspace, making the coordinate decomposition of vulnerability fingerprints of any instance unique and stable. Orthogonalization can also effectively suppress feature correlation interference introduced between different instances due to sharing the underlying image or configuration template, ensuring that the independent feature subspace truly reflects the differentiated vulnerability attributes of each instance.
[0086] Through the complete processing flow of rank decomposition, semantic constraint projection, tensor expansion and orthogonalization, we finally obtain a coupled feature subspace and an independent feature subspace with clear structure and explicit semantics. These subspaces respectively carry the resource dependencies between virtual desktop instances and the vulnerability attributes of each instance, providing a high-quality feature foundation for subsequent attack intent reasoning and attack evolution prediction.
[0087] In one optional implementation, the coupling components are mapped to an attack target-means binary semantic graph, and the set of target state nodes reachable by the current traffic behavior in the attack target-means binary semantic graph and the corresponding shortest means path are identified. The resulting intent reasoning result, including the remaining attack steps, is then obtained as follows:
[0088] Historical attack events are acquired, and the attack targets in the historical attack events are abstracted into target state nodes and attack methods are abstracted into method operation nodes. The preconditions and post-effects between the attack targets and the attack methods are analyzed. A directed edge is established between the target state nodes and the method operation nodes. The directed edge carries the weight representing the degree to which the method operation satisfies the target state and the constraint conditions representing the preconditions of the resources required to execute the method operation, thus forming the attack target-method binary semantic graph.
[0089] Map the coupling components to the means operation nodes of the attack target-means binary semantic graph, calculate the matching degree between the coupling components and the constraints, and mark the currently executable means operation nodes;
[0090] Starting from the currently executable means operation node, the constraint satisfaction propagation algorithm is executed, propagating forward along the directed edge and accumulating the weights and constraints. When the propagation path reaches the target state node and all constraints are satisfied, it is marked as a reachable target state node and the propagation path is recorded as the shortest means path. The remaining attack steps are extracted based on the sequence of means operation nodes that have not been executed in the shortest means path to form the intent inference result.
[0091] The construction of the attack target-means binary semantic graph uses historical attack events as the raw data source. After collecting historical attack events from publicly available vulnerability exploitation records, security incident response reports, and threat intelligence databases, the final or intermediate goals that the attacker attempted to achieve in each event are abstracted into target state nodes, such as "gaining domain administrator privileges," "lateral movement to the database server," and "persistent persistence." The specific operations actually performed by the attacker are abstracted into means operation nodes, such as "exploiting SMB protocol vulnerabilities to send malformed data packets," "injecting malicious DLLs into system processes," and "modifying registry startup items." Target state nodes and means operation nodes together constitute the vertex set of the graph. The two types of nodes are strictly distinguished at the semantic level, and the same node is not allowed to simultaneously play the dual roles of target and means.
[0092] By establishing directed edges between nodes, we analyze the precondition and post-effect relationships between each pair of target state nodes and means operation nodes. Preconditions describe the environmental states that a means operation node must satisfy before execution, such as "requiring a valid user credential" or "requiring the target host to have port 445 open." Post-effect relationships describe the degree to which the means operation node satisfies the target state node after successful execution, i.e., the extent to which the means operation promotes the realization of the corresponding target state. Directed edges point from means operation nodes to the target state nodes they can advance, and also from target state nodes to subsequent means operation nodes that depend on the target state being satisfied before execution, forming an alternating chain structure of "means → target → means → target." Each directed edge carries two types of attributes: weight and... The first is the degree to which the operation satisfies the target state node it directly targets; a larger value indicates a more significant contribution of the operation to the target. The second is the set of constraints. This describes the prerequisite resource conditions required to execute the attack method, including structured constraints such as the required system privilege level, network reachability, and the existence of specific services. Once the entire graph is constructed, a directed weighted constraint graph is formed, namely, an attack target-method binary semantic graph.
[0093] The coupling components are mapped to the attack target-means binary semantic graph. The coupling components themselves are feature representations extracted during cross-view semantic decoupling, reflecting the covariant relationships between multiple observation views. The coupling components are then matched with the feature descriptors of each means operation node, and the matching degree is... The calculation comprehensively considers the projection of the coupled components in the behavioral semantic space and the cosine similarity between the means operation node descriptors, as well as the resource occupancy patterns carried by the coupled components and the set of constraints of the means operation nodes. The degree of compliance with resource requirements. Specifically, for the means and operational nodes. (in (For the means of operating the node's index), its matching degree It is obtained by weighted merging of semantic similarity components and constraint compliance components. When Exceeding the preset executable threshold At that time, the node Nodes marked as currently executable means of operation are added to the propagation starting point set. The physical meaning of this step is: the coupled expression of currently observed network traffic behavior, process execution patterns, and resource contention characteristics, and which known methods of operation best match in both semantic and resource dimensions, thereby determining the set of operations that the attacker is currently executing or has the capability to execute.
[0094] Constraint satisfaction propagation algorithm with set Starting from all executable operation nodes, the propagation proceeds forward along directed edges. The propagation process maintains a path state record structure; for each extending path, it records the sum of its accumulated weights. The set of satisfied constraints and the sequence of nodes traversed by the path. When propagating along a directed edge from the means operation node to the target state node, the weight of that edge is... Accumulate to And check the set of constraints for that edge. Check if all constraints have been satisfied by the execution effects of the preceding nodes in the current path. If any constraints are not satisfied, the propagation path is blocked here and will not extend further; if all constraints are satisfied, it continues to propagate along the directed edges to the next level of means operation nodes. When the propagation path first reaches a target state node... (in When the target state node is indexed by its number and all constraints on that path are satisfied, then... Mark the node as a reachable target state and record the current path as a set starting from the origin. arrive Shortest means path Here, "shortest" is defined as the path traversing the fewest means of operation nodes. When the number of nodes is the same, the cumulative weight is used. The principle of "maximum priority" is reflected in prioritizing the path with the strongest effect among multiple paths of equal length that reach the same attack target.
[0095] After the propagation process is completed, the set of reachable target state nodes is obtained. And the set of shortest means paths corresponding to each reachable target state node. For each shortest means path Extract the means operation nodes that have not yet been marked as executed, and arrange them in the order of the path to form the sequence of remaining attack steps. The criteria for determining which operation nodes have been executed come from the set. If a certain method's operation node has already appeared If the attacker has already executed or is currently executing the operation, then it is considered that the operation is still pending. Otherwise, it is considered a remaining step to be executed. All reachable target state nodes, their corresponding shortest means paths, and the remaining attack step sequences for each path are integrated to form the intent reasoning result. The intent reasoning result is stored in a structured form, containing the set of target states currently pursued by the attacker, the operation chains required to achieve each target, and a clear division between the completed and uncompleted parts of each chain, providing accurate intent input for the construction of the subsequent attack evolution prediction field.
[0096] In a real-world cloud desktop environment, the attack target-method binary semantic graph needs to support a dynamic update mechanism. When new threat intelligence arrives, new method operation nodes and target state nodes can be incrementally added to the graph, and corresponding directed edges and weights can be added to the new nodes. With the set of constraints This eliminates the need to rebuild the entire graph structure. The constraint satisfaction propagation algorithm is re-executed with each new coupled component input to ensure that the intent inference results always reflect the latest attack posture. For multi-tenant cloud desktop scenarios where different tenants face different attack intentions, an independent set of executable means operation node starting points can be maintained for each tenant. Constraint satisfaction propagation is then executed in parallel on the same attack target-means binary semantic graph, generating intent inference results for each tenant, thereby achieving refined inference of differentiated attack intentions in a multi-tenant environment.
[0097] In one optional implementation, based on the remaining attack steps and the shortest means path, and combining the vulnerability fingerprints of each virtual desktop instance in the independent feature subspace with the resource dependency tensor in the coupled feature subspace, an attack evolution prediction field for the attack behavior on the virtual desktop topology is constructed, including:
[0098] Extract the vulnerability type required by the attack method corresponding to each means operation node in the remaining attack steps, and calculate the semantic similarity with the configuration exposure surface features recorded in the vulnerability fingerprint to obtain the matching degree;
[0099] Based on the matching degree, each virtual desktop instance is marked with a feasibility score as the target of the means operation node, and the feasibility score is assigned to the corresponding instance node in the virtual desktop topology to form an initial attack reachability distribution;
[0100] Resource sharing and call dependency relationships between virtual desktop instances are extracted from the resource dependency tensor in the coupled feature subspace and mapped to form propagation edges between instance nodes in the virtual desktop topology. The propagation coefficient of the propagation edge is determined by the value of the corresponding element in the resource dependency tensor.
[0101] The initial attack reachability distribution is used as the initial condition to perform a diffusion propagation process along the propagation edge. During the diffusion propagation process, the attack reachability value of each instance node is updated and iterated according to the propagation coefficient and the attack reachability value of the adjacent nodes until convergence. The converged attack reachability value distribution is temporally correlated with the means operation node sequence in the shortest means path to form the attack evolution prediction field.
[0102] In constructing the attack evolution prediction field, it is necessary to analyze the sequence of remaining attack steps output from the intent reasoning phase and each means operation node in the shortest means path one by one, extracting the vulnerability types that the corresponding attack means rely on during execution for each means operation node. These vulnerability types cover various semantic categories such as configuration errors, privilege escalation, service exposure, and authentication bypass, and are expressed in the form of a structured tag set. At the same time, the vulnerability fingerprint of each virtual desktop instance is read from an independent feature subspace. This fingerprint contains the configuration exposure surface features of the instance in its current running state, specifically including open port configuration, service version information, access control policy gaps, historical vulnerability records, and other multi-dimensional descriptive information, which are stored in the form of feature vectors in the independent feature subspace.
[0103] For the matching relationship between the vulnerability type required for each operational node and the vulnerability fingerprint of each virtual desktop instance, a semantic similarity calculation method is used to quantify it. Let the... The semantic representation vector of the vulnerability types required for each method to operate the node is: , No. The vulnerability fingerprint feature vector of each virtual desktop instance is Then the semantic similarity between the two It is calculated using the cosine similarity metric, i.e. This similarity score reflects the semantic availability of the current virtual desktop instance to a specific attack method. A higher score indicates that the instance is more likely to become a target of the corresponding attack method's operation node. For cases where multiple attack method operation nodes are involved in the remaining attack steps, semantic similarity calculations are performed for each node with all virtual desktop instances, forming a complete similarity matrix between the attack method node and the instance.
[0104] Based on the semantic similarity calculation results above, each virtual desktop instance is assigned a feasibility score for its use as a specific means to operate a node to execute a target. Feasibility Score Taking into account both semantic similarity scores and the severity weights of exposed surfaces in vulnerability fingerprints, for the same method operation node, if a virtual desktop instance highly matches the attack requirements across multiple vulnerability dimensions, its feasibility score is correspondingly increased. The feasibility scores corresponding to each method operation node are assigned to the corresponding instance nodes in the virtual desktop topology graph, forming an initial attack reachability distribution. This initial attack reachability distribution is attached to each instance node in the topology graph as a node attribute vector, reflecting the initial probability estimate of the node being directly compromised under the current attack intent. For cases where multiple method operation nodes exist in the remaining attack steps, the weighted sum of the feasibility scores of each node is taken as the final initial attack reachability value. The weights are determined by the priority order of the corresponding method operation nodes in the shortest method path, with nodes earlier in the path assigned higher weights to reflect the temporal priority of attack progression.
[0105] Horizontal association information between virtual desktop instances is extracted from the resource dependency tensor in the coupled feature subspace. The resource dependency tensor stores multiple dependencies between the virtual desktop instance dimension, resource type dimension, and temporal dimension in a three-order structure. By aggregating the resource type dimension, the comprehensive association strength between any two virtual desktop instances at the resource sharing and invocation dependency level can be obtained. Let's consider an instance... With examples The strength of the resource dependency association between them is This value is determined by normalizing the corresponding slice element in the resource dependency tensor. Exceeding the preset propagation edge construction threshold At that time, in the virtual desktop topology graph at the instance node With instance nodes Establish a directed propagation edge between them, and take the propagation coefficient of the propagation edge directly. The numerical values are as follows. Resource sharing relationships and call dependency relationships correspond to different types of propagation edges: propagation edges generated by resource sharing relationships are bidirectional edges, indicating that attack reachability can spread bidirectionally along the shared resources; propagation edges generated by call dependency relationships are unidirectional edges, with the direction consistent with the call direction, indicating that attackers can penetrate laterally from the caller side along the call chain. Through the above mapping operations, the virtual desktop topology graph is transformed from a simple logical topology into a weighted directed graph structure carrying propagation semantics.
[0106] Starting with the initial attack reachability distribution as the initial condition, a diffusion propagation process is performed on the aforementioned weighted directed graph. The diffusion propagation process employs an iterative update mechanism; in each iteration, the instance nodes... Attack reachability value The update is performed by weighting the attack reachability values of all its predecessor nodes and their corresponding propagation coefficients. The specific update rules are as follows: ,in This represents the initial attack reachability value. This is the diffusion attenuation coefficient, used to control the rate at which attack reachability decays during propagation. This is the current iteration round number. Decay coefficient. The value is set between 0 and 1. A smaller value indicates that the attack reachability decays faster during propagation, which is suitable for virtual desktop environments with a high degree of isolation; a larger value indicates stronger propagation ability, which is suitable for resource-intensive cloud desktop deployment scenarios with high resource sharing. The iterative process continues until the change in the attack reachability value of all instance nodes in two adjacent iterations is lower than the convergence threshold. At this point, the diffusion and propagation process is considered to have reached a convergence state.
[0107] The converged attack reachability distribution forms the spatial propagation result of the current attack intent across the entire virtual desktop topology, reflecting the relative probability of each virtual desktop instance being affected under the current attack path. This converged distribution is temporally correlated with the sequence of means operation nodes in the shortest means path. Specifically, the execution time window corresponding to each node in the means operation node sequence is correlated with snapshots of the reachability distribution at each stage of the propagation process, thus generating an attack reachability spatial distribution map for each time step. The spatial distribution maps of different time steps are stacked temporally to form an attack evolution prediction field with a time dimension. The attack evolution prediction field is stored in the form of a three-dimensional tensor, with the three dimensions corresponding to the virtual desktop instance number, time step number, and attack intensity value, respectively. The attack intensity value is obtained by standardizing the attack reachability value of the corresponding time step. Based on this prediction field, the probability distribution of each virtual desktop instance being compromised at multiple future time points can be directly read, and an attack intensity heatmap can be generated through heatmap visualization, providing spatialized and temporally sequenced attack situation input for the subsequent generation of game-theoretic defense schemes.
[0108] In one optional implementation, obtaining the probability distribution of compromise and attack intensity heatmaps of virtual desktop instances at multiple future time points based on the attack evolution prediction field includes:
[0109] The attack reachability values of each instance node in the virtual desktop topology at different time steps are extracted from the attack evolution prediction field. The attack reachability values are then weighted and fused with the attack success rate of the means operation node at the corresponding time step in the shortest means path. The attack success rate is obtained by statistically analyzing the historical successful execution records of the attack means corresponding to the means operation node. The probability of each virtual desktop instance being compromised at multiple future time points is obtained through the weighted fusion calculation. The probability of being compromised is then organized into the probability distribution of being compromised by time and space dimensions.
[0110] Calculate the rate of change of the probability of compromise of each virtual desktop instance at each time point in the compromise probability distribution, and multiply the rate of change of the probability of compromise with the attack reachability value of the corresponding instance node in the attack evolution prediction field to obtain the attack intensity value. The attack intensity value characterizes the urgency of the virtual desktop instance being attacked at a specific time point.
[0111] The attack strength value is mapped to a visual representation of the virtual desktop topology and different color codes are assigned according to the magnitude of the attack strength value to form an attack strength heatmap. The color depth in the attack strength heatmap is proportional to the attack strength value.
[0112] After extracting the attack reachability values of each virtual desktop instance node at different time steps from the attack evolution prediction field, these values need to be fused with the attack success rate of the corresponding time step of the shortest means path to obtain a more realistic probability estimate of compromise. The attack success rate is not a fixed constant but is statistically derived from the historical successful execution records of the attack means corresponding to that means operation node. Specifically, for each type of attack means, the number of successful executions and the total number of attempts of that means under the same or similar environments are statistically analyzed from historical alarm logs and sandbox execution records. The ratio of these two is used as the basic attack success rate estimate for that means operation node. To avoid estimation bias caused by the sparsity of historical data, a Laplace smoothing correction is introduced for means operation nodes with insufficient historical records, ensuring reasonable statistical robustness of the attack success rate estimate even with limited data.
[0113] Let the first The success rate of attacks manipulating nodes using these methods is [missing information]. Instance node At time step The attack reachability value is Then the instance at time step Corresponding methods to operate nodes The probability of being compromised is calculated using a weighted fusion method. The weighted fusion employs a product-weighted strategy, assigning weights to both the attack reachability value and the attack success rate. and ,in The two weights are configured based on their relative contribution to the actual compromise result according to the attack reachability and execution capability in the current network environment. The merged probability of compromise... Represented as And by using sigmoid normalization, the numerical values are constrained to... Within the interval, ensure that it possesses probabilistic semantics. When a time step in the shortest means path corresponds to multiple means operation nodes that can be executed in parallel, take the upper envelope value of the fusion result of each node, that is, take the maximum value as the probability of being compromised at that time step, so as to reflect the threat level of the most dangerous attack path.
[0114] The probability of each virtual desktop instance being compromised at multiple future time steps. Organized along the time and spatial dimensions, a probability distribution of vulnerability is formed. The time dimension covers all discrete time steps within the prediction window, and the spatial dimension covers the entire virtual desktop topology. Each instance node. Ultimately, the probability distribution of being compromised is stored in the form of a two-dimensional matrix, where the row index corresponds to the virtual desktop instance number. The column index corresponds to the time step. The matrix elements are the corresponding probability values of being compromised. This matrix supports both temporal risk evolution analysis of a single instance and horizontal risk comparison of all instances at the same point in time, providing complete state input for subsequent game modeling.
[0115] After constructing the probability distribution of compromise, the rate of change of the probability of compromise for each virtual desktop instance at each time step is further calculated. Rate of change Defined as the difference in the probability of being captured between two adjacent time steps, i.e. This reflects the dynamic evolution rate of the attack threat over time. When A positive value indicates that the attack threat faced by the instance is intensifying during that time period; a negative value indicates that the attack threat is in a decaying phase, because the preconditions for a certain means operation node in the attack chain have not yet been met, or the defensive measures have already blocked it. To eliminate noise interference, a moving average smoothing process is applied to the rate of change sequence, and the window length is adaptively adjusted according to the granularity of the prediction time step, usually taking 3 to 5 time steps.
[0116] Attack strength value It is obtained by multiplying the rate of change of the probability of being compromised by the attack reachability value of the corresponding instance node in the attack evolution prediction field, i.e. The physical meaning of this product operation is that the attack intensity value is significant only when the attack reachability of an instance node is high and the probability change rate is also large, thus accurately characterizing the urgency of an attack at a specific point in time. Simply having high reachability but a stable probability change, or a rapidly rising probability but extremely low reachability, is insufficient to determine a high-urgency threat. This product mechanism effectively avoids misjudgments caused by a single indicator. In the case of a negative value, The threshold is set to zero because the urgency of the attack threat decay phase is considered a non-positive threat and is not included in the intensity mapping range of the heatmap.
[0117] The attack intensity heatmap is generated based on a visual representation of the virtual desktop topology. The spatial locations and network connections of each instance node in the virtual desktop topology are mapped onto a two-dimensional plane, forming a topology layout diagram. On this layout diagram, with each instance node's location as the center, the attack intensity value at the current time step is used as the basis for the heatmap. Assign corresponding color codes. The color coding uses a continuous gradient color spectrum, assigning cool tones (such as blue or green) to areas with lower attack intensity values and warm tones (such as orange or red) to areas with higher attack intensity values. The color depth is... The values are directly proportional, and the specific mapping relationship is achieved through linear normalization, which normalizes the values of all instance nodes at the current time step. Value normalized to The interval is then mapped to the chromatogram.
[0118] To enhance the visual continuity of the heatmap, a kernel density estimation method is used for smooth interpolation of color transition areas between adjacent instance nodes. The bandwidth parameter of the interpolation kernel is adaptively set based on the average spacing of instance nodes in the topology graph. This allows the heatmap to not only clearly identify individual high-risk instance nodes but also visually present the spatial diffusion of attack threats throughout the virtual desktop topology, enabling security operations personnel to quickly identify attack hotspots and their spread directions. The heatmap supports dynamic refresh on a time-step basis, forming a time-series dynamic heatmap sequence that intuitively displays the joint distribution patterns of attack evolution in both time and space, providing a visual basis for the dynamic adjustment of defense strategies.
[0119] In actual deployment, the probability distribution of compromise and the heatmap of attack intensity are simultaneously output to the security situation awareness platform. The platform automatically triggers alarms based on high-intensity areas in the heatmap and overlays the time-series curves of the compromise probability of the corresponding instance nodes as a line graph. This helps security personnel determine whether the attack is in a rapid spread phase, thus providing accurate state-space input for solving subsequent Nash equilibrium game defense schemes.
[0120] In one optional implementation, the probability distribution of being compromised and the heatmap of attack intensity are used as the game state space, the virtualized defense operations of the defender are used as the strategy space, and the attacker's ability to adjust the attack path according to the defense situation is modeled as an adversarial strategy. An adversarial defense scheme is generated by solving for the Nash equilibrium, including:
[0121] The game state space is formed by vectorizing the probability distribution of being compromised and the heatmap of attack intensity to form a game state vector.
[0122] The virtualization defense operations of the defender are defined as including virtual desktop instance migration operations, resource isolation operations, and access control policy adjustment operations. Each virtualization defense operation corresponds to a defense action vector, which describes the influence coefficient of the virtualization defense operation on the game state vector. All defense action vectors are combined to form the policy space.
[0123] Based on the shortest means path and the resource dependency tensor, the attacker's countermeasure strategy is constructed. The countermeasure strategy takes the defensive action vector in the strategy space and the game state vector in the game state space as input, and outputs the attacker's adjusted attack target and attack path.
[0124] A two-player zero-sum game is constructed between the defender and the attacker. The payoff of the two-player zero-sum game is defined by the weighted difference between the decrease in the probability of being attacked and the resource expenditure. The defensive action vector sequence that maximizes the payoff is obtained by iteratively solving the game state space and the strategy space as the Nash equilibrium solution. The virtualized defensive operation sequence corresponding to the Nash equilibrium solution is used as the adversarial defense scheme.
[0125] like Figure 2 As shown, the method includes:
[0126] After obtaining the compromise probability distribution and attack intensity heatmap for each virtual desktop instance, these two types of prediction results need to be transformed into a structured representation that can be used for game solving. Specifically, the compromise probability value and attack intensity value of each virtual desktop instance at all prediction time steps are expanded according to the instance number and time step order, and concatenated to form a high-dimensional vector, called the game state vector. . The first half consists of the probability values of each instance being compromised at each time step. The components are arranged in row-major order, with the latter half consisting of the corresponding attack strength values. They are arranged and pieced together in the same order. All the game state vectors that appear together constitute the game state space. The dimension of the game state space is determined by the number of virtual desktop instances. Total number of predicted time steps Jointly decided, dimension is This vectorized encoding method embeds both the spatial distribution information and temporal evolution information contained in the heatmap of attack probability distribution and attack intensity into a single vector representation, which is convenient for subsequent game-solving frameworks to process directly.
[0127] The defender's strategy space consists of three types of virtualization defense operations: virtual desktop instance migration, resource isolation, and access control policy adjustment. Migration refers to moving high-risk instances from their current physical host to another host to cut off the attacker's established lateral movement path. Resource isolation refers to restricting access to the target instance's network bandwidth, shared memory, or storage channels, blocking resource dependency tensors. The identified highly correlated propagation edges; access control policy adjustment operations refer to dynamically modifying the access permission matrix between instances, raising the precondition thresholds that attackers must meet to advance the remaining attack steps. Each virtualization defense operation corresponds to a defense action vector. ,in This is a numbered index for defensive actions. Each component describes the effect of the defensive operation on the game state vector. The influence coefficient at the corresponding position represents the expected change in the probability of being compromised and the expected change in attack intensity after the operation is performed. The influence coefficient is estimated based on historical defense operation effectiveness data and the resource scheduling model of the virtualization platform. All defense action vectors The set of elements constitutes the defender's strategy space. .
[0128] The attacker's adversarial strategy is modeled as a conditional mapping function. Its input is the current game state vector. The defensive action vector chosen by the defender The output is the new set of attack target nodes adjusted by the attacker and the updated attack path. The construction of this mapping function depends on the shortest means path. With resource dependency tensor When the defender performs migration or isolation operations, the resource dependency tensor... The corresponding correlation strength component will change, causing the preconditions of some means operation nodes on the original shortest means path to no longer be met. The attacker needs to search for alternative paths to bypass the defense barrier again in the attack target-means binary semantic graph. By updating the resource dependency tensor Re-perform the reachability analysis to obtain the optimal attack path that the attacker can still advance under the current defensive posture, which serves as the attacker's response strategy output. This modeling approach reflects the attacker's dynamic adaptive capabilities and avoids the risk of defense scheme failure caused by treating attack behavior as static.
[0129] In constructing the game theory framework, the defender and attacker are modeled as two participants in a two-player zero-sum game. The defender selects a defensive action vector. The attacker uses countermeasures Response. Single-step payoff function in a game. Defined as the weighted difference between the decrease in the probability of being breached and the cost of defensive resources. Specifically, assuming a defensive action is performed... The post-game state vector from Change to The overall decrease in the probability of being attacked for and The weighted sum of the differences in the probability components of all instances being compromised, with the weights determined by the business importance level of each instance; defense resource overhead. This is an estimated cost-benefit analysis of the computational resources, network bandwidth, and management latency consumed in executing this defensive action. The revenue function is... ,in The weight of the benefit based on the reduction in the probability of being attacked. The penalty weights for resource overhead are both pre-configured hyperparameters. In a zero-sum game framework, the attacker's payoff is exactly equal to the negative of the defender's payoff; the attacker's goal is to minimize... The defender's goal is to maximize .
[0130] The solution to the Nash equilibrium is obtained by using an iterative optimal response method in the game state space. With strategy space The process proceeds alternately. In each iteration, the attacker's current adversarial strategy is fixed. The defender traverses the strategy space. For all defensive action vectors, select the one that maximizes the benefit function. Maximum defensive action vector As the current optimal response; subsequently, the choice of the fixed defender. The attacker used Update its attack path to obtain a new game state vector. This is then used as the initial state for the next iteration. This process continues iteratively until the defender's optimal response remains unchanged for several consecutive rounds, or the change in the payoff function falls below the convergence threshold. If the game reaches a certain equilibrium, then the game is considered to have reached a Nash equilibrium. At this point, the optimal defensive action vector sequence for the defender under each game state is the Nash equilibrium solution.
[0131] The defensive action vector sequence corresponding to the Nash equilibrium solution This is reduced to a specific sequence of virtualized defense operations, which constitutes the final adversarial defense scheme. For each defense action vector in the sequence... ( (Assign operation step numbers to the sequence), determine whether it is a migration operation, resource isolation operation, or access control policy adjustment operation based on its corresponding operation type identifier, and assign the instance number corresponding to the component with the largest absolute value in the influence coefficient. Determine the target instance for the operation. The target host machine for the migration operation is determined by the real-time resource status and resource dependency tensor of the virtualization platform. The distribution of association strength in the data jointly determines the priority of selecting the host machine with the lowest association strength with high-risk instances and sufficient remaining computing resources as the migration destination. The isolation granularity of resource isolation operations is determined by the magnitude of the influence coefficient of the resource type component in the corresponding defense action vector; the larger the absolute value of the influence coefficient, the stronger the isolation of that resource type. The access control policy adjustment operation updates the communication permission entries of the corresponding instance pair in the access control matrix of the virtualization platform according to the permission tightening direction indicated by the influence coefficient. The entire adversarial defense scheme is output in the form of an operation sequence, which can directly drive the execution of the virtualization platform's management interface, realizing dynamic adaptive defense against the cloud desktop environment.
[0132] A second aspect of this invention provides an intrusion prevention system for cloud desktop traffic analysis and AI prediction, comprising:
[0133] The feature decoupling unit is used to obtain the communication behavior features, process execution logic features and resource contention mode features of the cloud desktop environment, and after aligning them on a unified time axis, perform cross-view semantic decoupling, extract the coupling components of each view, and construct independent feature subspaces and coupled feature subspaces.
[0134] The intent reasoning unit is used to map the coupled components to the attack target-means binary semantic graph, identify the set of target state nodes that the current traffic behavior can reach in the attack target-means binary semantic graph and the corresponding shortest means path, and obtain the intent reasoning result containing the remaining attack steps.
[0135] An evolution prediction unit is used to construct an attack evolution prediction field on the virtual desktop topology based on the remaining attack steps and the shortest means path, combined with the vulnerability fingerprint of each virtual desktop instance in the independent feature subspace and the resource dependency tensor in the coupled feature subspace. Based on the attack evolution prediction field, the probability distribution of the compromise of each virtual desktop instance and the attack intensity heatmap at multiple future time points are obtained.
[0136] The game-theoretic defense unit is used to take the probability distribution of being compromised and the heat map of attack intensity as the game state space, the virtualized defense operation of the defender as the strategy space, and the ability of the attacker to adjust the attack path according to the defense situation as the adversarial strategy. The adversarial defense scheme is generated by solving the Nash equilibrium solution.
[0137] A third aspect of the present invention provides an electronic device, comprising:
[0138] processor;
[0139] Memory used to store processor-executable instructions;
[0140] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.
[0141] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.
[0142] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.
[0143] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. An intrusion prevention method based on cloud desktop traffic analysis and AI prediction, characterized in that, include: The system acquires communication behavior characteristics, process execution logic characteristics, and resource contention mode characteristics of the cloud desktop environment. After aligning them on a unified timeline, it performs cross-view semantic decoupling, extracts the coupling components of each view, and constructs independent feature subspaces and coupled feature subspaces. The coupling components are mapped to the attack target-means binary semantic graph, and the set of target state nodes reachable by the current traffic behavior in the attack target-means binary semantic graph and the corresponding shortest means path are identified to obtain the intent reasoning result containing the remaining attack steps; Based on the remaining attack steps and the shortest means path, and combining the vulnerability fingerprints of each virtual desktop instance in the independent feature subspace with the resource dependency tensor in the coupled feature subspace, an attack evolution prediction field of the attack behavior on the virtual desktop topology is constructed. Based on the attack evolution prediction field, the probability distribution of the compromise of each virtual desktop instance and the attack intensity heatmap at multiple future time points are obtained. The probability distribution of being compromised and the heatmap of attack intensity are used as the game state space, the virtualized defense operation of the defender is used as the strategy space, and the ability of the attacker to adjust the attack path according to the defense situation is modeled as the adversarial strategy. An adversarial defense scheme is generated by solving the Nash equilibrium solution.
2. The method according to claim 1, characterized in that, The characteristics of communication behavior, process execution logic, and resource contention patterns in the cloud desktop environment include: The network traffic data streams of multiple virtual desktop instances in the cloud desktop environment are stripped layer by layer according to the encapsulation layer to obtain the initial field sequence of each protocol layer. A field dependency graph is constructed for the initial field sequence of each protocol layer. The field dependency graph is traversed to extract the protocol behavior sequence reflecting the communication session establishment mode, data transmission mode and session termination mode as the communication behavior feature. The input parameters and return values of each system call in the system call sequence of multiple virtual desktop instances in the cloud desktop environment are constructed into a call data flow graph. The data object subgraph that forms a closed loop and the data object transfer path that crosses the process boundary are identified in the call data flow graph. The closed loop path in the data object subgraph is mapped to the intra-process call dependency chain, and the data object transfer path is mapped to the inter-process call dependency chain. The intra-process call dependency chain and the inter-process call dependency chain are merged to obtain a complete call dependency chain that reflects the inter-process cooperation logic and resource competition logic as the process execution logic feature. Based on the time-series matching of virtual machine resource allocation events recorded in the resource scheduling log of the virtual machine management layer in the cloud desktop environment with the resource request timestamps in the inter-process call dependency chain, the time-series coupling pattern between virtual desktop instances caused by resource scheduling is extracted as the feature of the resource competition mode.
3. The method according to claim 1, characterized in that, After alignment on a unified timeline, cross-view semantic decoupling is performed, extracting the coupling components of each view, and constructing independent feature subspaces and coupled feature subspaces, including: After aligning the communication behavior features, process execution logic features, and resource contention mode features on a unified time axis, a cross-view feature association matrix is constructed. By performing a rank decomposition operation on the cross-view feature association matrix, the cross-view feature association matrix is decomposed into a low-rank component that represents the shared semantics between views and a sparse component that represents the unique attributes of each view. Based on the multi-view feature collaboration rules of legitimate business behaviors in the cloud desktop environment, constraint boundaries are determined to form a semantic constraint subspace; The low-rank components are projected onto the semantic constraint subspace. By calculating the projection residuals of the low-rank components in the semantic constraint subspace, abnormal coupling patterns that violate the rules of legitimate business collaboration are identified. The associated components reflecting resource sharing and call dependencies between virtual desktop instances in the abnormal coupling patterns are extracted to form a resource dependency tensor as the coupling component. Tensor expansion is performed on the resource dependency tensor and the principal component directions are extracted as the basis vectors of the coupling feature subspace. The feature components reflecting the configuration and exposure surface of each virtual desktop instance in the sparse component are extracted as vulnerability fingerprints. The vulnerability fingerprints are grouped according to the virtual desktop instance identifiers and orthogonalized to obtain the basis vectors of the independent feature subspace.
4. The method according to claim 1, characterized in that, Mapping the coupling components to an attack target-means binary semantic graph, identifying the set of target state nodes reachable from the current traffic behavior in the attack target-means binary semantic graph and the corresponding shortest means path, yields an intent reasoning result including the remaining attack steps, including: Historical attack events are acquired, and the attack targets in the historical attack events are abstracted into target state nodes and attack methods are abstracted into method operation nodes. The preconditions and post-effects between the attack targets and the attack methods are analyzed. A directed edge is established between the target state nodes and the method operation nodes. The directed edge carries the weight representing the degree to which the method operation satisfies the target state and the constraint conditions representing the preconditions of the resources required to execute the method operation, thus forming the attack target-method binary semantic graph. Map the coupling components to the means operation nodes of the attack target-means binary semantic graph, calculate the matching degree between the coupling components and the constraints, and mark the currently executable means operation nodes; Starting from the currently executable means operation node, the constraint satisfaction propagation algorithm is executed, propagating forward along the directed edge and accumulating the weights and constraints. When the propagation path reaches the target state node and all constraints are satisfied, it is marked as a reachable target state node and the propagation path is recorded as the shortest means path. The remaining attack steps are extracted based on the sequence of means operation nodes that have not been executed in the shortest means path to form the intent inference result.
5. The method according to claim 1, characterized in that, Based on the remaining attack steps and the shortest means path, and combining the vulnerability fingerprints of each virtual desktop instance in the independent feature subspace with the resource dependency tensor in the coupled feature subspace, an attack evolution prediction field for the attack behavior on the virtual desktop topology is constructed, including: Extract the vulnerability type required by the attack method corresponding to each means operation node in the remaining attack steps, and calculate the semantic similarity with the configuration exposure surface features recorded in the vulnerability fingerprint to obtain the matching degree; Based on the matching degree, each virtual desktop instance is marked with a feasibility score as the target of the means operation node, and the feasibility score is assigned to the corresponding instance node in the virtual desktop topology to form an initial attack reachability distribution; Resource sharing and call dependency relationships between virtual desktop instances are extracted from the resource dependency tensor in the coupled feature subspace and mapped to form propagation edges between instance nodes in the virtual desktop topology. The propagation coefficient of the propagation edge is determined by the value of the corresponding element in the resource dependency tensor. The initial attack reachability distribution is used as the initial condition to perform a diffusion propagation process along the propagation edge. During the diffusion propagation process, the attack reachability value of each instance node is updated and iterated according to the propagation coefficient and the attack reachability value of the adjacent nodes until convergence. The converged attack reachability value distribution is temporally correlated with the means operation node sequence in the shortest means path to form the attack evolution prediction field.
6. The method according to claim 5, characterized in that, Based on the attack evolution prediction field, the probability distribution of the virtual desktop instance being compromised and the heatmap of attack intensity at multiple future time points are obtained, including: The attack reachability values of each instance node in the virtual desktop topology at different time steps are extracted from the attack evolution prediction field. The attack reachability values are then weighted and fused with the attack success rate of the means operation node at the corresponding time step in the shortest means path. The attack success rate is obtained by statistically analyzing the historical successful execution records of the attack means corresponding to the means operation node. The probability of each virtual desktop instance being compromised at multiple future time points is obtained through the weighted fusion calculation. The probability of being compromised is then organized into the probability distribution of being compromised by time and space dimensions. Calculate the rate of change of the probability of compromise of each virtual desktop instance at each time point in the compromise probability distribution, and multiply the rate of change of the probability of compromise with the attack reachability value of the corresponding instance node in the attack evolution prediction field to obtain the attack intensity value. The attack intensity value characterizes the urgency of the virtual desktop instance being attacked at a specific time point. The attack strength value is mapped to a visual representation of the virtual desktop topology and different color codes are assigned according to the magnitude of the attack strength value to form an attack strength heatmap. The color depth in the attack strength heatmap is proportional to the attack strength value.
7. The method according to claim 1, characterized in that, Using the probability distribution of being compromised and the heatmap of attack intensity as the game state space, the virtualized defense operations of the defender as the strategy space, and the attacker's ability to adjust the attack path according to the defense situation as the adversarial strategy, the adversarial defense scheme is generated by solving the Nash equilibrium solution, including: The game state space is formed by vectorizing the probability distribution of being compromised and the heatmap of attack intensity to form a game state vector. The virtualization defense operations of the defender are defined as including virtual desktop instance migration operations, resource isolation operations, and access control policy adjustment operations. Each virtualization defense operation corresponds to a defense action vector, which describes the influence coefficient of the virtualization defense operation on the game state vector. All defense action vectors are combined to form the policy space. Based on the shortest means path and the resource dependency tensor, the attacker's countermeasure strategy is constructed. The countermeasure strategy takes the defensive action vector in the strategy space and the game state vector in the game state space as input, and outputs the attacker's adjusted attack target and attack path. A two-player zero-sum game is constructed between the defender and the attacker. The payoff of the two-player zero-sum game is defined by the weighted difference between the decrease in the probability of being attacked and the resource expenditure. The defensive action vector sequence that maximizes the payoff is obtained by iteratively solving the game state space and the strategy space as the Nash equilibrium solution. The virtualized defensive operation sequence corresponding to the Nash equilibrium solution is taken as the adversarial defense scheme.
8. A cloud desktop traffic analysis and AI prediction intrusion prevention system, used to implement the method as described in any one of claims 1-7, characterized in that, include: The feature decoupling unit is used to obtain the communication behavior features, process execution logic features and resource contention mode features of the cloud desktop environment, and after aligning them on a unified time axis, perform cross-view semantic decoupling, extract the coupling components of each view, and construct independent feature subspaces and coupled feature subspaces. The intent reasoning unit is used to map the coupled components to the attack target-means binary semantic graph, identify the set of target state nodes that the current traffic behavior can reach in the attack target-means binary semantic graph and the corresponding shortest means path, and obtain the intent reasoning result containing the remaining attack steps. An evolution prediction unit is used to construct an attack evolution prediction field on the virtual desktop topology based on the remaining attack steps and the shortest means path, combined with the vulnerability fingerprint of each virtual desktop instance in the independent feature subspace and the resource dependency tensor in the coupled feature subspace. Based on the attack evolution prediction field, the probability distribution of the compromise of each virtual desktop instance and the attack intensity heatmap at multiple future time points are obtained. The game-theoretic defense unit is used to take the probability distribution of being compromised and the heat map of attack intensity as the game state space, the virtualized defense operation of the defender as the strategy space, and the ability of the attacker to adjust the attack path according to the defense situation as the adversarial strategy. The adversarial defense scheme is generated by solving the Nash equilibrium solution.
9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.
10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.
