Heterogeneous computing power terminal adaptive intrusion detection model compression method and system

Abstract

The invention discloses a heterogeneous computing power terminal adaptive intrusion detection model compression method and a heterogeneous computing power terminal adaptive intrusion detection model compression system, and aims at equipment with large computing power difference, such as a command vehicle, a portable terminal, an edge server and an unmanned aerial vehicle group, the method comprises the following steps: firstly, constructing a parameter quantity-calculated quantity-time delay three-dimensional hardware budget space on the ground side; cPU / GPU / NPU occupancy, memory, bandwidth, temperature and return flow portraits are collected in real time; the outer layer uses Gaussian process Bayesian optimization to quickly lock feasible budget points, the inner layer uses active learning and a lightweight agent model to screen elite strategies in layer pruning, quantification and distillation combination, and p95 / p99 delay, energy consumption and precision are verified in a real machine through a resource awareness acquisition function; all the triads are put in storage according to Pareto optimum, a locality sensitive hash index is established, and millisecond-level online retrieval is achieved; after deployment, the SLA is continuously monitored, reselection and hot update are carried out once a super tolerance is delayed or misreported, and the model is ensured to always run at low delay and high precision under finite computing power.

Description

Technical Field

[0001] This invention relates to the field of heterogeneous cluster security technology, and in particular to a method and system for compressing adaptive intrusion detection models for heterogeneous computing power terminals. Background Technology

[0002] With the rapid development of artificial intelligence technology, deep neural networks have achieved remarkable results in tasks such as intrusion detection and target recognition. However, high-performance AI models are typically computationally and storage-intensive, making them difficult to deploy directly on resource-constrained devices such as drones and edge servers. To alleviate this contradiction, model compression techniques such as pruning, quantization, and knowledge distillation have become a research hotspot.

[0003] Existing technologies largely focus on using one or more compression methods to generate lightweight models to adapt to specific hardware. While this has promoted the application of AI at the edge to some extent, the following key issues remain: First, existing methods mostly adopt a "one-time optimization" approach, generating "static point models" for specific hardware constraints, which are difficult to adapt to heterogeneous hardware and dynamic environmental changes. For example, a model optimized for a certain type of edge chip is difficult to migrate flexibly between devices with different performance levels. Once the hardware is upgraded or the task changes, the time-consuming optimization process needs to be repeated, which lacks agility.

[0004] Second, the optimization process often couples multi-dimensional policy parameters (such as pruning rate and quantization bit width) with hardware performance indicators (such as latency and power consumption) in the same search space, leading to the "curse of dimensionality," low search efficiency, and a tendency to get trapped in local optima. Existing methods mostly rely on trial and error adjustments, wasting computational resources and making it difficult to obtain the globally optimal policy.

[0005] Third, the level of automation is insufficient. Most solutions still rely on expert experience for manual parameter tuning, such as setting the upper limit of compression ratio and selecting pruning granularity. End-to-end full-process automation has not yet been achieved, which limits the standardization and large-scale application of the technology.

[0006] In summary, existing methods have significant shortcomings in terms of heterogeneity adaptability, search efficiency, and automation level, making it difficult to meet the needs of complex scenarios such as drone networks for efficient and agile deployment of AI models. Summary of the Invention

[0007] A method and system for compressing heterogeneous computing power terminal adaptive intrusion detection models. This method, through inner and outer layer optimization and a dynamic strategy library, enables dynamic adaptation to heterogeneous hardware resources and efficient model deployment, effectively solving the rigidity problem of the "one-time optimization" model. It also enables rapid and accurate search for compression strategies, avoiding the curse of dimensionality by decoupling hardware budget and strategy parameter space, significantly improving optimization efficiency. Furthermore, it automates the entire process from strategy generation to online adjustment, greatly reducing reliance on expert experience and enhancing the system's deployment agility and robustness in dynamic environments.

[0008] A method for compressing an adaptive intrusion detection model for heterogeneous computing power terminals includes the following steps: S1. In the constructed hardware budget space, select a hardware budget point; the hardware budget space is a three-dimensional continuous space composed of model parameters, floating-point operation counts and ground-end inference latency; the hardware budget point is any point located in the hardware budget space, corresponding to a specific set of budget targets. S2. Generate a candidate compression strategy set based on the selected hardware budget point, and use a lightweight proxy model to quickly evaluate the candidate compression strategy set and select an elite candidate set; the candidate compression strategy set is a collection of several candidate compression strategies, and the candidate compression strategy is a compression configuration vector generated by the Latin hypercube sampling method. Each compression configuration vector corresponds to a complete pruning strategy and is called a candidate compression strategy; the elite candidate set is a subset of the selected compression strategies. S3. Select one or more candidate compression strategies from the elite candidate set and perform real performance evaluation to obtain triples; the triples include at least the hardware budget point, the candidate compression strategy, and the performance evaluation result. S4. Dynamically incorporate triplet groups into the strategy library according to the multi-objective Pareto optimality principle; S5. Establish a multi-dimensional index structure for the policy library with hardware budget points as the primary key; S6. Utilize the multidimensional index structure to retrieve candidate compression strategies, select the final strategy according to the hierarchical decision rules, apply the final strategy to compress the intrusion detection model to generate a deployable model and go online. After deployment, make dynamic adjustments based on actual test feedback.

[0009] In some examples of the present invention, S1 specifically includes: The hardware budget space is constructed using the number of model parameters, the number of floating-point operations, and the inference latency as three coordinate axes; Within the hardware budget space, a feasible region is defined by online telemetry information and uplink backhaul traffic profiles, satisfying device resource and real-time constraints. Then, an outer optimization loop is executed. The hardware budget point is any point located within the feasible region; the outer optimization loop selects a new hardware budget point within the feasible region.

[0010] In some examples of the present invention, the hardware budget point is selected using a Bayesian optimization framework with Gaussian process regression as the surrogate model. A collection function integrating diversity rewards is used to balance the utilization of known high-performance regions and the exploration of unexplored regions. The diversity rewards are based on the distance between the hardware budget point and historically evaluated hardware budget points in the normalized space.

[0011] In some examples of the present invention, S2 specifically includes: For the hardware budget point selected by S1, an inner optimization loop is executed within a policy space consisting of compression parameters of each layer of the model. The inner optimization loop first generates an initial set of candidate compression strategies, and then pre-screens them through static analysis and resource estimation, eliminating candidate compression strategies that are predicted to consume more resources than the current hardware budget or may block the ground processing link. The lightweight proxy model is used to quickly evaluate the pre-screened candidate compression strategies. The proxy model can simultaneously predict the performance indicators and resource consumption of the candidate compression strategies and estimate the uncertainty of the prediction. Based on the results of the rapid evaluation, an elite candidate set with high prediction performance and meeting resource constraints is selected.

[0012] In some examples of the present invention, in the inner optimization loop, a resource-aware acquisition function selects candidate compression strategies from the elite candidate set for real performance evaluation. The acquisition function comprehensively considers the predictive performance, predictive uncertainty, and penalty for predictive resource consumption exceeding the current hardware budget point of the candidate compression strategies, thereby prioritizing the selection of strategies with high performance potential, high uncertainty, and low risk of exceeding the budget for evaluation.

[0013] In some examples of this invention, the maintenance mechanism of the strategy library in S4 is specifically as follows: When a new triplet is generated, it is compared with the existing records in the policy library for Pareto dominance. If the new triple is dominated by any record in the database, it is discarded; If the new triplet dominates one or more records in the library, then the dominated record is removed and stored in the new triplet; If the new triplet does not interact with any records in the database, it is stored directly. When the policy library capacity exceeds the limit, a pruning mechanism is activated to prioritize the retention of records that are located at a better Pareto front and are more dispersed on that front.

[0014] In some examples of the present invention, in step S5, the multidimensional index structure is constructed based on Locality Sensitive Hash (LSH) to achieve nearest neighbor retrieval with sublinear time complexity; the index structure supports incremental updates and includes a performance monitoring and adaptive adjustment mechanism, which can automatically adjust parameters or switch to a backup index algorithm when index performance degradation is detected.

[0015] In some examples of the present invention, in step S6, the hierarchical decision rules include, in order of priority: Prioritize the candidate compression strategy with the highest detection performance; If the detection performance is the same, the candidate compression strategy with the smallest tail latency index shall be selected first. If they are still the same, the candidate compression strategy with the fewest model parameters will be selected first. If they are still the same, the candidate compression strategy with the fewest floating-point operations will be selected first. If all the above indicators are the same, then a deterministic sorting will be performed based on the record metadata.

[0016] In some examples of the present invention, in step S6, the dynamic adjustment specifically refers to: After deployment, continuously monitor the model's measured performance on the target terminal; When the measured performance consistently exceeds the tolerance range of the target service level SLA within a preset time window, a more conservative hardware budget point is generated based on the deviation between the measured performance and the expected performance. The retrieval, selection, and deployment processes are re-executed using the corrected hardware budget point.

[0017] A heterogeneous computing power terminal adaptive intrusion detection model compression system includes: The strategy library management module is used to store and maintain the strategy library, and to perform dynamic inclusion and maintenance of the triples. An offline optimization engine is used to perform the selection of the hardware budget point, the generation and filtering of candidate compression strategy sets, and the acquisition of triples; the offline optimization engine includes: An outer optimizer is used to perform an outer optimization loop in the hardware budget space to select a hardware budget point; An inner optimizer is used to perform an inner optimization loop for a selected hardware budget point to generate and filter the elite candidate set; An online deployment engine is used to perform compression strategy retrieval, decision-making, model generation, and dynamic adjustment; the online deployment engine includes: An index retrieval module, connected to the strategy library management module, is used to retrieve candidate compression strategies using the multidimensional index structure. The strategy decision module is used to select the final strategy from the search results according to the hierarchical decision rules; A model generator is used to apply the final strategy to the original intrusion detection model for compression to generate the deployable model. An adaptive controller is used to trigger the dynamic adjustment based on the measured feedback.

[0018] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention.

[0019] The beneficial effects of this invention are: This invention systematically solves the core challenge of deploying high-performance AI models in resource-constrained, heterogeneous drone networks through an innovative "offline optimization-online deployment" paradigm and a two-layer automation framework. Compared with traditional static compression methods, this invention achieves a fundamental shift from "one-time optimization, single model" to "one-time optimization, multiple uses, and flexible adaptation." Specifically, in the offline phase, through decoupled Bayesian optimization and active learning mechanisms, a Pareto optimal policy library covering multi-dimensional hardware budgets is efficiently constructed, overcoming the bottlenecks of low search efficiency, susceptibility to local optima, and reliance on expert experience inherent in traditional methods. In the online phase, with index-based sub-second policy retrieval and deterministic hierarchical decision rules, customized optimal lightweight models can be instantly generated and deployed for scattered and heterogeneous computing terminals, achieving zero redundant optimization overhead and full-process automation in the deployment process. This closed-loop system not only significantly improves the initial deployment success rate and operational efficiency of the model on various devices, from high-end ground stations to low-power portable terminals, but also ensures long-term operational stability and real-time performance in dynamic task environments through explicit constraints and continuous adaptive adjustments to tail latency, traffic profiling, and resource fluctuations. Ultimately, this invention significantly reduces the application threshold and technical maintenance costs of AI models in complex edge scenarios in engineering practice, comprehensively improving the overall performance and total cost of ownership of the UAV network intrusion detection system and various edge intelligent applications from multiple dimensions, including model adaptation, deployment agility, resource utilization, and system robustness. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the method for compressing the adaptive intrusion detection model of heterogeneous computing power terminals provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the two-layer optimization process in the offline phase; Figure 3 This is a schematic diagram of the online deployment process; Figure 4 The system structure diagram of the heterogeneous computing power terminal adaptive intrusion detection model compression system provided in the embodiment of the present invention is shown. Detailed Implementation

[0022] 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.

[0023] In this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0024] This invention provides an adaptive intrusion detection model compression method for heterogeneous computing terminals. This invention can be applied to numerous general-purpose or specialized computing device environments or configurations, such as airborne computing devices in UAV swarms, ground-based computing terminals, edge computing infrastructure, cloud computing or central training platforms, and heterogeneous computing platforms.

[0025] This embodiment uses an intrusion detection neural network model (e.g., a CNN- or Transformer-based traffic classification model) for a drone swarm communication network as the original model to be compressed. The method follows a two-layer automated framework of "offline optimization-online deployment," aiming to decouple the computationally intensive compression strategy search from the lightweight, real-time model generation process. Figure 1 As shown, it includes the following steps: S1. In the constructed hardware budget space, select a hardware budget point; the hardware budget space is a three-dimensional continuous space composed of model parameters, floating-point operation counts and ground-end inference latency; the hardware budget point is any point located in the hardware budget space, corresponding to a specific set of budget targets. S2. Generate a candidate compression strategy set based on the selected hardware budget point, and use a lightweight proxy model to quickly evaluate the candidate compression strategy set and select an elite candidate set; the candidate compression strategy set is a collection of several candidate compression strategies, and the candidate compression strategy is a compression configuration vector generated by the Latin hypercube sampling method. Each compression configuration vector corresponds to a complete pruning strategy and is called a candidate compression strategy; the elite candidate set is a subset of the selected compression strategies. S3. Select one or more candidate compression strategies from the elite candidate set and perform real performance evaluation to obtain triples; the triples include at least the hardware budget point, the candidate compression strategy, and the performance evaluation result. S4. Dynamically incorporate triplet groups into the strategy library according to the multi-objective Pareto optimality principle; S5. Establish a multi-dimensional index structure for the policy library with hardware budget points as the primary key; S6. Utilize the multidimensional index structure to retrieve candidate compression strategies, select the final strategy according to the hierarchical decision rules, apply the final strategy to compress the intrusion detection model to generate a deployable model and go online. After deployment, make dynamic adjustments based on actual test feedback.

[0026] The above S1~S6 are composed of two main parts: the offline optimization phase and the online deployment phase, such as... Figure 2 As shown, the offline optimization stage includes the following specific steps: S1. An outer optimization loop is performed within a hardware budget space consisting of model parameters, floating-point operations, and ground-end inference latency. This budget space is constrained by the online telemetry and uplink backhaul traffic profile of the target ground terminal to determine the feasible region. The online telemetry includes at least CPU / GPU / NPU usage, memory / bandwidth margin, device temperature, and power supply mode. The uplink backhaul traffic profile includes at least packet rate, concurrent streams, burst rate, and packet loss retransmission status from the cluster. Within this feasible region, a Bayesian optimization framework using Gaussian process regression as a proxy is applied, and the next hardware budget point to be evaluated is selected based on a collection function that integrates diversity rewards. This diversity is measured in a normalized space to distance from historical points, encouraging uniform exploration of different traffic load slices. During the offline optimization phase, statistical analysis is performed on telemetry logs and link packet capture data from historical flight missions to construct a typical online telemetry distribution and uplink backhaul traffic profile. During the online deployment phase, the telemetry module of the ground control station continuously collects data from the current mission to update the feasible region of the hardware budget point.

[0027] Specifically, the Bayesian optimization framework employs Gaussian Process Regression (GP) as a surrogate model. This model uses a Squared Exponential (SE) kernel function to measure the similarity between any two hardware budget points and, based on existing observation data, optimizes any new candidate hardware budget point. The performance yields a prediction that follows a posterior Gaussian distribution, the mean of which is... and variance The calculation formula is:

[0028]

[0029] In this formula, Calculated based on all observed points kernel matrix, This represents the number of observed points. It is the relationship between the new candidate point and all observed points. Covariance vector; These are the vectors of the actual performance values ​​corresponding to these observed points; The variance of the observed noise; It is then the identity matrix; The acquisition function, which integrates diverse rewards, is used to select the next hardware budget point to be evaluated. It combines Upper Confidence Bound (UCB) exploration with diversity rewards, and follows the formula below:

[0030] in, and These are the GP model at the hardware budget point. Predicted performance mean and standard deviation For the feasible region of the hardware budget space, and These are preset hyperparameters used to balance the weights of exploration-exploitation and to adjust the weights of diversity rewards, respectively. The diversity incentive is designed to encourage candidate budget points to be compared with the set of historically evaluated hardware budget points within the normalized space. Maintaining a greater distance results in higher bonus points. It is determined by calculating the normalized Euclidean distance from the candidate point to its nearest neighbor in the historical point set:

[0031] in, For the historical set of hardware budget points, Representative historical points Any budget point in the budget; It is a normalization function used to map dimensions of different units to a uniform and comparable numerical range; This means calculating the L2 norm between two vectors.

[0032] S2. For the hardware budget point selected in S1, an inner optimization loop based on active learning is executed within a policy space consisting of compression parameters of each layer of the model. This loop uses a lightweight agent model to quickly evaluate candidate compression strategies and selects an elite candidate set based on the evaluation results for subsequent selection. During the pre-screening, the constraints of the ground processing chain are considered: the data capture / unpacking / decryption / preprocessing pipeline is not blocked, the 95th / 99th percentile (p95 / p99) inference latency meets the real-time requirements under the backhaul peak, and the scheduling is isolated from other field task processes.

[0033] It should be noted that the outer optimization loop outputs a hardware budget point H each time; the inner optimization loop uses this H as a constraint to search for and compress a strategy S in the strategy space formed by the pruning rates of each layer, so that S satisfies H in terms of resource consumption and is as good as possible in terms of performance indicators such as accuracy.

[0034] Specifically, for the hardware budget point selected in S1, an inner optimization loop based on active learning is executed within a policy space consisting of compression parameters of each layer of the model. This loop uses a lightweight proxy model to quickly evaluate candidate compression strategies and selects an elite candidate set based on the evaluation results for subsequent selection, specifically including: Latin Hypercube Sampling (LHS) is used to generate an initial candidate set. Then, strategies whose resource consumption or end-to-end p95 / p99 latency exceeds the current ground budget can be determined through static analysis. At the same time, strategies that will cause packet capture, decryption, pipeline blockage, or / and mutual preemption with other services on the same machine in scheduling are also eliminated.

[0035] A lightweight surrogate model based on a feedforward neural network is used to quickly evaluate pre-selected candidate policies; this surrogate model utilizes Monte Carlo Dropout (MCDropout) to estimate the uncertainty of its predictions; a compression policy is applied for each input. The proxy model simultaneously predicts its performance metrics (such as accuracy) and dynamic resource consumption (such as inference latency, energy consumption per unit time, and peak memory usage); by applying the same strategy conduct The predicted mean can be calculated using a forward propagation with random inactivation. With the predicted standard deviation :

[0036]

[0037] in, The compression strategy to be evaluated; In the first In the second forward propagation with Dropout, the lightweight agent model applies the policy. The predicted value; This represents the number of repetitions in the forward propagation. For this The mean of the predicted values; The standard deviation of the prediction is used to measure the confidence level of the surrogate model in the prediction.

[0038] Based on the prediction results of the surrogate model, a multi-objective evaluation is performed on all candidate strategies to select an elite candidate set with high prediction performance and prediction resource consumption that meets the current hardware budget constraints. The size of this elite set is within a configurable range relative to the initial candidate set, so that S3 can make the final selection.

[0039] It should be noted that the resource consumption predictions are all given by the same lightweight proxy model, which simultaneously predicts the detection performance of the compression strategy and multi-dimensional resource consumption. It is a multi-output feedforward neural network.

[0040] The training process of this lightweight agent model consists of two parts: In the offline phase, we first collect the performance and resource consumption metrics of the original model measured on the ground under different compression strategies and different hardware budget points. Using the compression strategy and its corresponding hardware budget point as input, and the above-mentioned real performance and resource data as supervision labels, we perform initial training on the lightweight proxy model. During the online phase, after each real-world performance evaluation on the ground, the obtained "hardware budget point - compression strategy - performance evaluation results and resource consumption" is used as a new sample to continuously and incrementally update the same surrogate model. Combined with the validation set and early stopping criteria to control the training process, the prediction accuracy is continuously improved.

[0041] S3. In the inner optimization loop, one or more candidate compression strategies are selected from the elite candidate set screened in S2 through a resource-aware acquisition function for ground-based real-world performance evaluation. This function jointly considers agent prediction performance, uncertainty, and pre-penalty for over-budget risk, and gives higher weight to tail latency index to cope with traffic bursts and jitter caused by air-to-ground links.

[0042] Specifically, the performance evaluation results after S3 real performance evaluation are used as new training samples to incrementally update the surrogate model, thereby continuously improving its prediction accuracy in the iterative cycle; the training process of the surrogate model is monitored by the validation set performance and the early stopping criterion; in addition, the termination condition of the inner optimization loop is linked to the convergence state of the outer Bayesian optimization. When the overall performance gain of the outer optimization is lower than a preset threshold, the inner loop can be terminated in advance to save computing resources.

[0043] In the inner optimization loop, a resource-aware acquisition function selects one or more candidate compression strategies from the elite candidate set filtered by S2 for real-world performance evaluation. This acquisition function integrates the predictive performance of the strategy, prediction uncertainty, and pre-penalty for expected resource consumption exceeding the current hardware budget, specifically including: This function selects the optimal strategy by solving the following optimization problem. : ; in, It is an elite candidate set selected by S2; and These are the mean and standard deviation of the agent model's predictions of the performance of policy S, respectively. It is a hyperparameter that balances exploration and utilization. This represents the current hardware budget point; The aforementioned pre-requisite resource penalty item This is used to penalize, in the selection phase, those projects that predict resource consumption exceeding the current hardware budget point. The constraint strategy is defined using a piecewise linear hinge form: ; In this formula, , and These are the predicted values ​​of the number of parameters, number of floating-point operations, and inference latency of the agent model for policy S, respectively. , , The target component of the current hardware budget point H; , , The corresponding penalty weights are used to control the severity of penalties for exceeding limits in different resource dimensions; It is a piecewise linear hinge function used to penalize only when the resource forecast exceeds the budget limit.

[0044] Furthermore, to support parallel computing, this selection process can be extended to a batch mode, which selects the top K highest-scoring strategies from the elite candidate set at once, provided they meet certain difference constraints, for parallel real-world performance evaluation.

[0045] S4. Dynamically incorporate and maintain the strategy library for each "hardware budget point – compression strategy – real performance" triplet that has been evaluated for real performance, according to the multi-objective Pareto optimality principle. The library records metadata such as ground terminal platform family, Network Interface Card (NIC) / driver and kernel version, deployment topology (single machine / cluster / hot standby), traffic profile tags and timestamp / version hash. The maintenance mechanism ensures both optimality and diversity under limited capacity.

[0046] Specifically, S4 includes: The policy library records a structured definition: each triple that has undergone real evaluation is formalized as a parameter containing a hardware budget point. Compression strategy Actual performance indicators and metadata strategy library records .

[0047] Among them, the actual performance indicators At least including accuracy, inference latency, and the number of model parameters; the metadata is environmental information collected synchronously during real-world performance evaluation on the ground. It should include at least timestamps, version checksum hashes, terrestrial terminal equipment family / accelerator type, operating system / kernel and NIC / driver information, traffic profile tags (packet rate range, concurrent session volume, burst coefficient), and deployment topology tags. This metadata serves as additional information for subsequent retrieval and sorting.

[0048] Dynamic update rule based on Pareto optimality: When a new record When it arises, it is processed according to the following Pareto dominance: Dominated: If any record exists in the policy library Its performance indicators No inferior to in all optimization objectives And is strictly superior to at least one objective. ,but If the solution is a passive solution, it will not be stored in the database.

[0049] Dominate: if If one or more records in the policy repository are dominated, then those dominated records are removed from the repository, and... Add to the database.

[0050] Non-dominance: if If it does not control any records in the database, then directly... Add to the database.

[0051] A pruning mechanism based on capacity and diversity is implemented: When the policy pool's capacity exceeds a preset limit, a pruning mechanism based on the Non-dominated Sorting Genetic Algorithm II (NSGA-II) is activated. First, all records in the pool are subjected to a fast non-dominated sort to create different Pareto fronts. Then, all records of the optimal front (Front1) are retained first, followed by the second-best front (Front2), and so on, until the capacity limit is reached. For the last partially retained front, records are sorted according to crowding distance, prioritizing records with larger crowding distances to ensure a wider distribution and greater diversity of solutions in the target space.

[0052] S5. Establish a multi-dimensional index structure for the strategy library, using hardware budget points (which can be combined with discrete tags or hashes from traffic profiles) as the primary key to achieve sublinear online retrieval and have the ability to adaptively adjust to changes in data distribution.

[0053] Specifically, a multi-dimensional index structure supporting efficient retrieval is established for the strategy library; the index structure aims to reduce the complexity of online queries from linear scanning to sublinear levels and has the ability to adaptively adjust to changes in data distribution, specifically including: Main index construction—based on Locality-Sensitive Hashing (LSH): using the hardware budget point of each record in the policy library. As the index key, an index based on locality-sensitive hashing is constructed. By designing a family of LSH functions, a high-dimensional budget space vector is mapped to multiple hash tables, so that vectors that are close in the original space collide with the same hash bucket with a high probability. This enables the querying of a small number of hash buckets to replace a full database scan when an online query request is received, thereby achieving nearest neighbor retrieval with the expected sublinear time complexity.

[0054] Dynamic maintenance and incremental updates of the index: The index structure supports online incremental updates. When the strategy library in S4 changes (including adding or deleting records), it is not necessary to completely rebuild the index. Instead, only the affected hash buckets are partially updated, thus ensuring the timeliness and low maintenance cost of the index.

[0055] Adaptive Index Degradation Monitoring and Switching Mechanism: A built-in index performance monitoring module evaluates the performance of the current LSH index by periodically sampling queries or monitoring the load balancing factor of the hash bucket. When index performance (such as precision or query latency) degrades below a preset threshold due to uneven data distribution or increased budget dimensionality, an adaptive adjustment mechanism is triggered. Parameter adaptation: Prioritize online adjustment of LSH hyperparameters (such as the number of hash functions, bucket width, etc.) to adapt to the new data distribution.

[0056] Structure switching: If parameter adjustments cannot restore performance, the index structure will be automatically and seamlessly switched to a backup index algorithm that is not sensitive to data distribution, such as graph-based HNSW (HierarchicalNavigableSmallWorld) or tree-based KD-Tree variant, to ensure query efficiency and robustness in extreme cases.

[0057] like Figure 3 As shown, the online deployment phase includes the following specific steps: S6. In the online deployment phase, based on the hardware budget point and SLA target of the target ground terminal, candidate strategies are efficiently retrieved using the index of S5, and the final strategy is selected according to the hierarchical decision rules. It is then applied to the original ground intrusion detection model to generate a deployable model and put it online. After deployment, continuous monitoring is performed based on the actual test feedback from the ground terminal (inference latency distribution, throughput / packet loss, false alarms / missed alarms, energy consumption and thermal status). When the SLA is not met due to traffic bursts or resource fluctuations, dynamic adjustments are triggered (including re-retrieval, hot parameter updates, or enabling ground-side distributed inference / elastic scaling when conditions permit).

[0058] Specifically, during the online deployment phase, based on the received terminal hardware budget and service level objectives, candidate policies are efficiently retrieved from the policy library using an index built by S5, and a final policy is selected according to a hierarchical decision rule. The selected final policy is applied to the original network for compression to generate a deployable model, and based on the actual performance feedback from the deployed edge, dynamic adjustments to the policy are triggered when the service level objectives are not met, including: Retrieval: In response to receiving a hardware budget point and a service level agreement (SLA) threshold from a target terminal, a candidate policy set is retrieved from the policy library using the index structure established in S5, and the candidate policy set is filtered to form an SLA-compatible candidate set that only contains records in the library whose performance meets the SLA threshold. Selection: Applying a deterministic, multi-level decision rule, a final strategy is selected from the SLA-compatible candidate set, wherein the multi-level decision rule includes: Prioritize candidate strategies with the highest detection performance (e.g., accuracy / recall); If they are the same, the candidate strategy with the lowest inference latency of p95 / p99 will be selected first (to ensure stability during peak periods). If they are still the same, the candidate compression strategy with the fewest model parameters will be selected first. If they are still the same, the candidate compression strategy with the fewest FloatingPointOperations (FLOPs) will be selected first. If all of the above are the same, then a deterministic sort is performed based on the record's metadata to ensure uniqueness (such as version hash / timestamp / device family priority).

[0059] Deployment: Based on the final strategy, generate a deployable ground model and deploy it to the target terminal; Monitoring and Adaptive Adjustment: After deployment, continuously monitor the measured performance of the deployable model on the target terminal, and when it is determined that the measured performance continuously exceeds a tolerance range of the SLA threshold within a preset time window, generate a corrected and more conservative hardware budget point based on the deviation between the measured performance and the performance recorded in the library of the final strategy, and use the corrected hardware budget point to repeat the previous steps to select and deploy a new deployable model to the target terminal.

[0060] Through the aforementioned S1-S6 implementation steps, this invention enables adaptive compression and deployment of intrusion detection models for heterogeneous computing power terminals in UAV networks. It achieves a Pareto optimal balance between model performance and resource consumption through a two-layer framework of "offline optimization-online deployment." Compared to traditional static compression methods focused on a single hardware point or manual parameter tuning methods requiring repeated optimization, this invention proactively constructs a strategy library covering multi-dimensional hardware budgets. It instantly generates the optimal lightweight model based on the terminal's real-time resource status and business requirements, thereby significantly improving the model's deployment success rate, operational efficiency, and overall system security in heterogeneous computing power environments.

[0061] and Figure 1 Corresponding to the method described above, this embodiment of the invention also provides an adaptive intrusion detection model compression system for heterogeneous computing power terminals, used for compressing... Figure 1 The specific implementation of the method is illustrated in the structural diagram of the heterogeneous computing power terminal adaptive intrusion detection model compression system provided in this embodiment of the invention. Figure 4 As shown, it specifically includes: The strategy library management module 100 is used to store and maintain the strategy library, and to perform dynamic inclusion and maintenance of the triples; An offline optimization engine 200 is used to perform the selection of the hardware budget point, the generation and filtering of candidate compression strategy sets, and the acquisition of triples; the offline optimization engine includes: Outer optimizer 210 is used to perform an outer optimization loop in the hardware budget space to select a hardware budget point; Inner optimizer 220 is used to perform an inner optimization loop for a selected hardware budget point to generate and filter the elite candidate set; An online deployment engine 300 is used to perform retrieval, decision-making, model generation, and dynamic adjustment of compression strategies; the online deployment engine includes: The index retrieval module 310 is connected to the strategy library management module and is used to retrieve candidate compression strategies using the multidimensional index structure. The strategy decision module 320 is used to select the final strategy from the search results according to the hierarchical decision rules; Model generator 330 is used to apply the final strategy to the original intrusion detection model for compression to generate the deployable model; An adaptive controller 340 is used to trigger the dynamic adjustment based on the measured feedback.

[0062] The specific working processes of the strategy library management module 100, offline optimization engine 200, outer optimizer 210, inner optimizer 220, online deployment engine 300, index retrieval module 310, strategy decision module 320, model generator 330, and adaptive controller 340 in the heterogeneous computing power terminal adaptive intrusion detection model compression system disclosed in the above embodiments of the present invention can be found in the corresponding content of the heterogeneous computing power terminal adaptive intrusion detection model compression method disclosed in the above embodiments of the present invention, and will not be repeated here.

[0063] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple, and relevant parts can be referred to the descriptions in the method embodiments. The systems and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0064] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in connection with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both.

[0065] To clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been generally described in terms of functionality above. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0066] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for compressing an adaptive intrusion detection model for heterogeneous computing power terminals, characterized in that, Includes the following steps: S1. In the constructed hardware budget space, select a hardware budget point; the hardware budget space is a three-dimensional continuous space composed of model parameters, floating-point operation counts and ground-end inference latency; the hardware budget point is any point located in the hardware budget space, corresponding to a specific set of budget targets. S2. Generate a candidate compression strategy set based on the selected hardware budget point, and use a lightweight proxy model to quickly evaluate the candidate compression strategy set and select the elite candidate set; the candidate compression strategy set is a collection of several candidate compression strategies, and the candidate compression strategy is a compression configuration vector generated by the Latin hypercube sampling method. Each compression configuration vector corresponds to a complete pruning strategy and is called a candidate compression strategy. The elite candidate set is a subset of compression strategies that have been selected. S3. Select one or more candidate compression strategies from the elite candidate set and perform real performance evaluation to obtain triples; the triples include at least the hardware budget point, the candidate compression strategy, and the performance evaluation result. S4. Dynamically incorporate triplet groups into the strategy library according to the multi-objective Pareto optimality principle; S5. Establish a multi-dimensional index structure for the policy library with hardware budget points as the primary key; S6. Utilize the multidimensional index structure to retrieve candidate compression strategies, select the final strategy according to the hierarchical decision rules, apply the final strategy to compress the intrusion detection model to generate a deployable model and go online. After deployment, make dynamic adjustments based on actual test feedback.

2. The method for compressing the adaptive intrusion detection model of heterogeneous computing power terminals according to claim 1, characterized in that, S1 specifically includes: The hardware budget space is constructed using the number of model parameters, the number of floating-point operations, and the inference latency as three coordinate axes; Within the hardware budget space, a feasible region is defined by online telemetry information and uplink backhaul traffic profiles, satisfying device resource and real-time constraints. Then, an outer optimization loop is executed. The hardware budget point is any point located within the feasible region; the outer optimization loop selects a new hardware budget point within the feasible region.

3. The method for compressing the adaptive intrusion detection model of heterogeneous computing power terminals according to claim 2, characterized in that: The hardware budget point is selected using a Bayesian optimization framework with Gaussian process regression as the surrogate model. A collection function that integrates diversity rewards balances the utilization of known high-performance regions and the exploration of unexplored regions. The diversity rewards are based on the distance between the hardware budget point and historically evaluated hardware budget points in the normalized space.

4. The method for compressing the adaptive intrusion detection model of heterogeneous computing power terminals according to claim 1, characterized in that, S2 specifically includes: For the hardware budget point selected by S1, an inner optimization loop is executed within a policy space consisting of compression parameters of each layer of the model. The inner optimization loop first generates an initial set of candidate compression strategies, and then performs pre-screening through static analysis and resource estimation to eliminate candidate compression strategies that are predicted to consume more resources than the current hardware budget or may block the ground processing link. The lightweight proxy model is used to quickly evaluate the pre-screened candidate compression strategies. The proxy model can simultaneously predict the performance indicators and resource consumption of the candidate compression strategies and estimate the uncertainty of the prediction. Based on the results of the rapid evaluation, an elite candidate set with high prediction performance and meeting resource constraints is selected.

5. The method for compressing the adaptive intrusion detection model of heterogeneous computing power terminals according to claim 4, characterized in that: In the inner optimization loop, a resource-aware acquisition function selects candidate compression strategies from the elite candidate set for real performance evaluation. The acquisition function comprehensively considers the predictive performance, predictive uncertainty, and penalty for predictive resource consumption exceeding the current hardware budget point of the candidate compression strategies, thereby prioritizing the selection of strategies with high performance potential, high uncertainty, and low risk of exceeding the budget for evaluation.

6. The method for compressing the adaptive intrusion detection model of heterogeneous computing power terminals according to claim 1, characterized in that, In S4, the maintenance mechanism of the strategy library is as follows: When a new triplet is generated, it is compared with the existing records in the policy library for Pareto dominance. If the new triple is dominated by any record in the database, it is discarded; If the new triplet dominates one or more records in the library, then the dominated record is removed and stored in the new triplet; If the new triplet does not interact with any records in the database, it is stored directly. When the policy library capacity exceeds the limit, a pruning mechanism is activated to prioritize the retention of records that are located at a better Pareto front and are more dispersed on that front.

7. The method for compressing the adaptive intrusion detection model of heterogeneous computing power terminals according to claim 1, characterized in that: In S5, the multidimensional index structure is built based on Locality Sensitive Hash (LSH) to achieve nearest neighbor retrieval with sublinear time complexity. The index structure supports incremental updates and includes a performance monitoring and adaptive adjustment mechanism. When index performance degradation is detected, it can automatically adjust parameters or switch to a backup index algorithm.

8. The method for compressing the adaptive intrusion detection model of heterogeneous computing power terminals according to claim 1, characterized in that, In step S6, the hierarchical decision rules, in order of priority, include: Prioritize the candidate compression strategy with the highest detection performance; If the detection performance is the same, the candidate compression strategy with the smallest tail latency index shall be selected first. If they are still the same, the candidate compression strategy with the fewest model parameters will be selected first. If they are still the same, the candidate compression strategy with the fewest floating-point operations will be selected first. If all the above indicators are the same, then a deterministic sorting will be performed based on the record metadata.

9. The method for compressing the adaptive intrusion detection model of heterogeneous computing power terminals according to claim 1, characterized in that, In S6, the dynamic adjustment specifically refers to: After deployment, continuously monitor the model's measured performance on the target terminal; When the measured performance consistently exceeds the tolerance range of the target service level SLA within a preset time window, a more conservative hardware budget point is generated based on the deviation between the measured performance and the expected performance. The retrieval, selection, and deployment processes are re-executed using the corrected hardware budget point.

10. A system for implementing the adaptive intrusion detection model compression method for heterogeneous computing power terminals as described in any one of claims 1 to 9, characterized in that, The system includes: The strategy library management module is used to store and maintain the strategy library, and to perform dynamic inclusion and maintenance of the triples. An offline optimization engine is used to perform the selection of the hardware budget point, the generation and filtering of candidate compression strategy sets, and the acquisition of triples; the offline optimization engine includes: An outer optimizer is used to perform an outer optimization loop in the hardware budget space to select a hardware budget point; An inner optimizer is used to perform an inner optimization loop for a selected hardware budget point to generate and filter the elite candidate set; An online deployment engine is used to perform compression strategy retrieval, decision-making, model generation, and dynamic adjustment; the online deployment engine includes: An index retrieval module, connected to the strategy library management module, is used to retrieve candidate compression strategies using the multidimensional index structure. The strategy decision module is used to select the final strategy from the search results according to the hierarchical decision rules; A model generator is used to apply the final strategy to the original intrusion detection model for compression to generate the deployable model. An adaptive controller is used to trigger the dynamic adjustment based on the measured feedback.

Images