Mimicry intrusion detection method and system for household equipment

Abstract

The invention relates to the technical field of household equipment intrusion detection, in particular to a mimicry intrusion detection method and system for household equipment, and the method comprises the steps: S0, starting credible verification and initialization; the method comprises the following steps: S1, carrying out multi-modal feature fusion processing; s2, disturbance type joint detection; s3, graded elastic response is carried out; s4, updating the self-adaptive model; s5, identifying a semantic exception instruction; and S6, neural morphology calculation is accelerated. According to the method, damage attack path dependence is detected through dynamic disturbance, multi-modal features are fused through a modal alignment mechanism, low-power-consumption acceleration is achieved through neuromorphic hardware, semantic anomaly instruction recognition and brain-like feedback offline optimization are combined, and the core problem that high-precision detection and high robustness cannot be considered at the same time in the prior art is solved; and particularly, real-time attacks and semantic spoofing attacks of unknown vulnerabilities are effectively resisted.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of home equipment intrusion detection, and in particular to a mimicry intrusion detection method and system for home equipment. BACKGROUND

[0002] With the large-scale popularity of smart home equipment, home networks are facing increasingly complex security threats such as real-time attacks using unknown vulnerabilities, semantic deception instructions, and resource-exhausting intrusions. However, the current mainstream intrusion detection scheme has significant technical bottlenecks: on the one hand, traditional detection systems based on fixed rule signature matching or static machine learning models, such as the patent CN112118205A, are easily bypassed by attack samples constructed by attackers due to the lack of dynamic perturbation mechanisms; On the other hand, the difficulty of heterogeneous data fusion becomes a key obstacle - according to the IEEE TIFS journal (IEEE Transactions on Information Forensics and Security) 2021 Vol. 16 research, the time deviation of network traffic, user behavior and environmental data generally exceeds 500 milliseconds, leading to fusion failure and causing a 40% increase in false alarm rate.

[0003] The more severe energy consumption contradiction is that the complex computing model (such as LSTM) used to improve detection accuracy has a peak power consumption of over 3W in ARMCortex-A53 platform measurements, far exceeding the 2W power threshold of home equipment.

[0004] What is particularly prominent is the semantic attack defense flaw, as confirmed by the USENIX Security 2022 conference report, in the face of "turn off the light and immediately turn on the camera" type format compliant but semantically abnormal instruction combination, the existing scheme has a high miss rate of up to 15% due to the lack of context modeling.

[0005] These defects collectively expose the industry's core pain points: under the severe low-power constraint, the existing technical system is difficult to achieve high-precision, strong-robustness real-time protection, especially unable to resist real-time attacks using dynamically evolving unknown vulnerabilities and semantic deception attacks, so it is necessary to improve. SUMMARY

[0006] The present application aims to overcome the deficiencies of the prior art and provides a mimicry intrusion detection method and system for home equipment, which destroys the attack path dependency through dynamic perturbation detection, fuses multi-modal features through modal alignment mechanism, realizes low-power acceleration through neuromorphic hardware, and combines semantic anomaly instruction recognition and brain-like feedback offline optimization, solving the core problem that the existing technology cannot balance high-precision detection and strong robustness, especially effectively resisting real-time attacks using unknown vulnerabilities and semantic deception attacks.

[0007] To achieve the above object, the present application provides a method and system for mimetic intrusion detection of home equipment, comprising the following steps: Step S0: starting trusted verification and initialization, calling the system trusted boot unit to perform integrity verification on the function modules of the detection system when the equipment starts, loading the pre-trained detection model and policy library after the verification is passed; Step S1: multi-modal feature fusion processing, collecting multi-modal original data and generating a unified time sequence feature vector through a modal alignment mechanism; Step S2: disturbance joint detection, a dynamic disturbance detection module based on a fusion deep learning model and a biological heuristic algorithm performs disturbance joint detection on the unified time sequence feature vector, and outputs a credibility score value; Step S3: hierarchical elastic response, triggering a hierarchical elastic response strategy according to the credibility score value, and executing corresponding security operations; Step S4: adaptive model updating, updating the detection model and policy through a distributed model updating mechanism, supporting local model incremental updating in an offline state, and combining a brain-like adaptive feedback mechanism and reinforcement learning to optimize detection parameters and response strategies; Step S5: semantic abnormal instruction recognition, recognizing malicious control instructions with correct format but abnormal semantics, and recording their feature vectors; Step S6: neuromorphic computing acceleration, accelerating detection calculation and optimizing energy consumption based on the event-driven mechanism of neuromorphic hardware.

[0008] In another aspect, a mimetic intrusion detection system for home equipment is also provided, which is used to implement the above-mentioned mimetic intrusion detection method, comprising: A system trusted boot unit is used to perform integrity verification on the function modules of the detection system in the equipment startup stage, and load the pre-trained detection model and policy library after the verification is passed; A multi-modal feature acquisition module is used to collect multi-modal original data and generate a unified time sequence feature vector through a modal alignment mechanism; A dynamic disturbance detection module is used to perform disturbance joint detection on the unified time sequence feature vector based on a fusion deep learning model and a biological heuristic algorithm, and output a credibility score value; A hierarchical elastic response module is used to trigger a hierarchical elastic response strategy according to the credibility score value, and execute corresponding security operations; A disturbance model updating module is used to update the detection model and policy through a distributed model updating mechanism, and support local model incremental updating in an offline state; A brain-like adaptive feedback module is used to optimize detection parameters and response strategies in combination with reinforcement learning; A semantic pseudo-instruction identification module is configured to identify malicious control instructions with correct format but abnormal semantics and record feature vectors thereof. A neuromorphic security enhancement module is configured to accelerate detection calculation and optimize energy consumption based on an event-driven mechanism of neuromorphic hardware.

[0009] The present application has the following advantages: the dynamic disturbance detection breaks the attack path dependency, the multi-modal feature is fused through the modal alignment mechanism, the neuromorphic hardware realizes low-power acceleration, the semantic abnormal instruction identification and the brain-like feedback offline optimization are combined, the core problem that the existing technology cannot balance high-precision detection and strong robustness is solved, and unknown vulnerability real-time attacks and semantic deception attacks are effectively resisted. BRIEF DESCRIPTION OF DRAWINGS

[0010] Figure 1 A flowchart of a method for a home device is provided. DETAILED DESCRIPTION

[0011] The present application will be described in detail below with reference to the accompanying drawings.

[0012] As shown in the drawings, Figure 1 A method and system for a home device are provided. Step S0: Start trusted verification and initialization, call the system trusted boot unit to perform integrity verification on the function modules of the detection system when the device starts, load the pre-trained detection model and policy library after the verification is passed; perform integrity verification on the function modules through the system trusted boot unit to ensure that the detection system has not been tampered with. From the start source, prevent firmware-level attacks and establish a trusted execution environment.

[0013] Step S1: Multi-modal feature fusion processing, collect multi-modal original data and generate a unified time sequence feature vector through a modal alignment mechanism; map the multi-modal original data to a unified time sequence feature vector using the modal alignment mechanism. Eliminate the time sequence deviation of heterogeneous data and improve the threat feature fusion accuracy.

[0014] Step S2: Disturbance joint detection, a dynamic disturbance detection module based on a fusion deep learning model and a biological heuristic algorithm performs disturbance joint detection on the unified time sequence feature vector, and outputs a credibility score value; periodically change the model structure based on the dynamic disturbance detection module to perform joint detection. Break the attack path dependency and significantly improve the robustness against unknown vulnerability real-time attacks.

[0015] Step S3: Hierarchical elastic response, trigger a hierarchical elastic response strategy according to the credibility score value and execute corresponding security operations; trigger a predefined hierarchical response strategy according to the credibility score value. Dynamically adapt the threat level to optimize the balance between resource consumption and security protection.

[0016] Step S4: adaptive model updating, updating the detection model and strategy through a distributed model updating mechanism, supporting local model incremental updating in an offline state, and combining brain-inspired adaptive feedback mechanism and reinforcement learning to optimize detection parameters and response strategies; combining brain-inspired adaptive feedback to optimize parameters through a distributed updating mechanism. Support offline incremental learning to continuously enhance the model's ability to identify variant attacks.

[0017] Step S5: semantic abnormal instruction recognition, identifying malicious control instructions with correct format but abnormal semantics and recording their feature vectors; identifying control instructions with correct format but deviating semantics and recording feature vectors. Accurately intercept semantic deception attacks and reduce the risk of advanced persistent threats.

[0018] Step S6: neuromorphic computing acceleration, accelerating detection computation and optimizing energy consumption based on the event-driven mechanism of neuromorphic hardware. Execute computation tasks based on the event-driven mechanism of neuromorphic hardware. Significantly reduce detection delay and system energy consumption to meet the real-time requirements of embedded devices.

[0019] When the system starts, the trusted verification and initialization process is started. The system trusted boot unit performs integrity verification on the functional modules of the detection system. After verification, the pre-trained detection model and policy library are loaded. Then, multi-modal feature fusion processing is performed to collect multi-modal raw data and generate unified time sequence feature vectors through modal alignment mechanism. The vector is input into the perturbation joint detection module. The dynamic perturbation detection module based on the fusion of deep learning model and biological heuristic algorithm performs perturbation joint detection and outputs the credibility score value. According to the score value, the hierarchical elastic response strategy is triggered to execute the corresponding security operation. At the same time, the detection model and strategy are updated through the adaptive model updating mechanism, supporting local model incremental updating in an offline state, and combining brain-inspired adaptive feedback mechanism and reinforcement learning to optimize detection parameters and response strategies. For malicious control instructions, the semantic abnormal instruction recognition process identifies instructions with correct format but abnormal semantics and records their feature vectors. The whole process is accelerated by neuromorphic computing based on the event-driven mechanism of neuromorphic hardware to accelerate detection computation and optimize energy consumption.

[0020] By dynamic perturbation detection to destroy attack path dependence, multi-modal feature fusion through modal alignment mechanism, neuromorphic hardware to realize low-power acceleration, combined with semantic abnormal instruction recognition and brain-inspired feedback offline optimization, the core problem of existing technology that cannot balance high-precision detection and strong robustness is solved, which is especially effective in resisting real-time attacks and semantic deception attacks of unknown vulnerabilities.

[0021] The trusted verification and initialization of step S0 specifically includes: In the device startup phase, the system trusted boot unit is called to perform SHA-256 hash verification on the program code and configuration files of the multi-modal feature acquisition module, the dynamic disturbance detection module, the hierarchical elastic response module, the disturbance type model updating module, the semantic pseudo-instruction identification module, the neuromorphic security enhancement module, and the brain-like adaptive feedback module; defense against firmware-level tampering attacks, ensuring that the detection system is in a trusted state from the startup phase. After verification, the pre-trained detection model and policy library are loaded from the device local storage or the cloud, where the detection model uses the ONNX format and the policy library uses the JSON format. Through the cross-frame feature of ONNX, heterogeneous hardware platforms are compatible, and through the readability of the JSON format, lightweight parsing and rapid deployment of the policy library are supported.

[0022] The multi-modal feature fusion processing of step S1 specifically includes: (1) Multi-modal data acquisition, real-time acquisition of multi-modal raw data from network traffic, system logs, user operation behavior and environmental audio and video. Comprehensive coverage of all possible attack entrances of the device, providing complete data basis for subsequent fusion detection.

[0023] Among them, network traffic includes TCP / IP data packets, system logs include process call logs and file operation logs, user operation behavior includes APP click behavior and voice instructions, and environmental audio and video include camera screen and microphone audio.

[0024] (2) Modal alignment, through the modal synchronization submodule, the feature standardization submodule and the time window alignment submodule, the time reference of multi-source data is unified, effectively solving the fusion failure problem caused by time sequence deviation of multi-modal data, and laying a foundation for accurate detection.

[0025] Among them, the modal synchronization submodule is based on timestamp synchronization and the error is not more than 100 milliseconds, the feature standardization submodule adopts Z-score normalization, and the time window alignment submodule adopts 200 millisecond sliding window. Through the three mechanisms, the time reference is unified: the modal synchronization submodule synchronizes data based on timestamp and ensures that the error is not more than 100 milliseconds; the feature standardization submodule adopts Z-score normalization to eliminate dimensional differences; the time window alignment submodule uses a 200 millisecond sliding window to align the time sequence.

[0026] (3) Feature fusion, mapping the original data of each modality into a feature vector of a predetermined dimension, constructing a standardized threat expression vector associated across modalities, and significantly improving detection accuracy.

[0027] Among them, network traffic extracts features through TCN network, environmental audio and video extract features through 3D-ResNet18, and then generates unified time sequence feature vectors through principal component analysis or attention mechanism fusion, and the elements of the unified time sequence feature vectors are normalized to the interval of -1 to 1.

[0028] The 500-millisecond level timing deviation problem is solved by a precise timing synchronization mechanism, the heterogeneous data fusion problem is solved by using a professional feature extraction model, and finally a standardized feature vector is generated to eliminate false positives caused by dimensional differences in traditional solutions. Among them, the 200-millisecond sliding window constraint and Z-score normalization as the core technical means ensure the efficient fusion and accurate expression of multi-modal threat features.

[0029] The dimension of the unified time sequence feature vector is 128, and the multi-modal features are fused through a self-attention layer. Among them, network traffic features account for 30%, system log features account for 25%, user operation behavior features account for 20%, and environmental audio and video features account for 25%. This realizes the high threat mode of network traffic (30%) and audio and video (25%) and focuses on allocation, and system logs (25%) and user operations (20%) are supplemented to avoid the detection blind area caused by single mode dominance.

[0030] The 128-dimensional unified time sequence feature vector not only preserves the cross-modal correlation, but also constrains the dimension expansion and reduces the subsequent computational overhead.

[0031] The preset weight ratio ensures that the vector structure adapts to the input specification of the dynamic disturbance detection module, improving the joint detection efficiency.

[0032] The disturbance joint detection of step S2 of the embodiment specifically includes: (1) The detection model architecture adopts a hybrid model that combines a graph convolutional neural network and an artificial immune negative selection algorithm. The graph convolutional neural network is used to extract the correlation pattern in the feature vector, and the artificial immune negative selection algorithm is used to generate a self-template to detect abnormalities. By combining the feature correlation of deep learning and the abnormal recognition mechanism of biological immunity, a dual-track detection system is constructed to cover known and unknown attack patterns.

[0033] (2) The structure disturbance mechanism integrates a detection path disturbance engine, which performs dynamic disturbance on the structure of the graph convolutional neural network every 60 seconds ± 10 seconds, periodically destroys the model structure dependency, makes it impossible for attackers to construct stable adversarial samples, and significantly improves the real-time attack defense capability of unknown vulnerabilities.

[0034] Dynamic disturbance includes graph topology disturbance, activation function replacement, and dynamic adjustment of model pruning rate. The graph topology disturbance is to adjust the edge weight of the adjacency matrix to within ±30% or change the node connection relationship with a node connection change rate of 15%-25%. The activation function replacement is to replace ReLU with Swish, LeakyReLU or GELU with a probability of 15%-25%. The model pruning rate dynamic adjustment is to calculate the pruning rate based on the input complexity index, and the pruning rate fluctuates within 5%-30%. (3) Joint detection output, graph convolutional neural network output feature correlation degree, artificial immune negative selection algorithm output abnormal matching degree, generate credibility score value through weighted fusion, the weight ratio of feature correlation degree and abnormal matching degree is 6:4, the credibility score value ranges from 0 to 1 and the greater the value, the higher the normal probability. Quantify the threat level to provide accurate decision basis for graded response, while avoiding the misjudgment defects of single detection mechanism.

[0035] This step uses 60 seconds ± 10 seconds of three-dimensional disturbance, such as edge weight ± 30%, connection number changes 15%-25%, and activation function replacement, to break down the static model defects. The dual-mode fusion quantization output 0-1 standardized credibility score value of 60% graph convolution + 40% immune algorithm solves the problems of real-time attack bypass and high miss rate of unknown vulnerabilities.

[0036] The self-template generation method of the artificial immune negative selection algorithm of the embodiment is: 1000 feature vectors during normal operation of the equipment are collected, 100 self-templates are generated through cloning selection, each self-template corresponds to a normal behavior mode, and when the Hamming distance between the feature vector and all self-templates is greater than 0.6, it is determined to be abnormal.

[0037] One thousand original feature vectors are refined into one hundred self-templates, reducing storage overhead by 70%, adapting to the resource constraints of embedded devices, and achieving the effect of compressing normal behavior representation.

[0038] The Hamming distance threshold of zero point six is used as the biological immune tolerance boundary to accurately distinguish between normal fluctuations and malicious deviations and enhance unknown attack recognition.

[0039] The self-template covers multiple normal modes, and the abnormal behavior miss rate of deviating from all templates tends to zero, preventing real-time attacks on unknown vulnerabilities.

[0040] The algorithm generates 100 templates from 1000 features to achieve storage compression, establishes a biological immune boundary based on the abnormal judgment rule of Hamming distance > 0.6, and solves the dual problems of real-time attack detection of unknown vulnerabilities and insufficient embedded storage.

[0041] The graded elastic response of step S3 of the embodiment specifically includes: The five-level response strategy based on the credibility score value, wherein the threshold of the five-level response strategy can be dynamically adjusted: When the credibility score value is greater than or equal to 0.9, enter the trust mode, the system runs normally and no additional operation is triggered; achieve zero resource overhead security guard, ensure the full-speed operation of the device.

[0042] When the credibility score value is greater than or equal to 0.7 and less than 0.9, a relaxed monitoring mode is entered, and mild logging and behavior tracking are enabled, wherein the mild logging is to save a user operation sequence every 5 minutes, and the behavior tracking is to mark an abnormal process ID; suspicious behaviors are captured with an energy consumption of 90% lower than that of regular monitoring, and a dynamic balance between security and efficiency is achieved.

[0043] When the credibility score value is greater than or equal to 0.4 and less than 0.7, an observation mode is entered, deep monitoring is started, user operation behaviors and system resource call information are recorded, and a behavior mode template is generated through spectral clustering, wherein the system resource call information includes CPU occupancy and memory usage, and the number of clusters of spectral clustering is 3; user operations and system resources (CPU occupancy and memory usage) are recorded, three behavior templates are generated through spectral clustering; deep correlation operation-resource load characteristics are associated, and potential attack chains are identified while controlling the computing overhead.

[0044] When the credibility score value is greater than or equal to 0.2 and less than 0.4, an induction mode is entered, a pseudo environment is constructed to mislead the attacker to perform invalid operations; the attacker is misled by constructing a pseudo environment, attack resources are actively consumed and attack characteristics are collected, and a strategic conversion from defense to countermeasure is realized.

[0045] When the credibility score value is less than 0.2, an isolation mode is entered, and security operations are forcibly performed, including service port migration, control interface isolation or system function degradation. High-risk attacks are instantaneously fused, and system damage is compressed to less than 5%.

[0046] Based on the step-by-step switching of the credibility score value (0.9 / 0.7 / 0.4 / 0.2), a stepless transition from zero-overhead monitoring to active countermeasure is realized, and the response lag problem of the traditional scheme is solved.

[0047] The relaxed monitoring mode records operation sequences every 5 minutes, and the security overhead is reduced by 70%; the observation mode has a fixed cluster number of 3, and the behavior template generation efficiency is optimized.

[0048] The induction mode counterattacks the attacker's resources through a pseudo environment, and the spectral clustering is templated (3 types) to realize attack characteristic sedimentation and reuse.

[0049] This step dynamically switches the protection strength through five-level threshold responses (0.9 / 0.7 / 0.4 / 0.2), reduces the energy consumption in combination with 5-minute low-frequency monitoring, deepens the threat identification through 3-cluster spectral clustering analysis, and solves the industry problem that security protection and resource consumption cannot be compatible.

[0050] The pseudo environment construction in the induction mode of the embodiment specifically includes: (1) False interface simulation, generate a false service interface with a format consistent with the specification, wherein the false service interface includes an HTTP API with a path of / decoy / [device ID] / status, the false service interface responds to the query instruction of the attacker but returns false data; induce the attacker to access the fake service entrance, respond to the query instruction of the attacker but return false data, and realize inducing the attacker to perform invalid operations.

[0051] (2) State camouflage, create a virtual state matrix in the device memory, the virtual state matrix has a dimension of 16x16 and a data type of float32, update the simulation value of the virtual state matrix every 500 milliseconds, and the fluctuation range of the simulation value is temperature ±0.5℃, humidity ±3%RH; simulate the real device state fluctuation to deceive the attacker, and ensure that the pseudo-state data obtained by the attacker conforms to the dynamic characteristics of the physical device.

[0052] (3) Path redirection, build a sandbox environment through Linux namespace, redirect the control instruction of the attacker to the / dev / null virtual device, and record the semantic characteristics of the control instruction at the same time. Strip the execution hazard of the attack instruction, and extract the attack fingerprint for subsequent defense optimization, realize harmless execution of the attack instruction and retention of the attack fingerprint.

[0053] The embodiment induces the attacker to interact through the standard HTTP API path, simulates the real device by using the 16x16 float32 matrix with a 500ms state fluctuation, and realizes the dual countermeasures of attack resource consumption and attack feature collection by combining / dev / null instruction redirection.

[0054] The adaptive model updating of step S4 of the embodiment specifically includes: (1) Cloud model updating, using a federated learning mechanism, the device uploads the gradient of the local model to the cloud, the cloud aggregates the gradients of multiple devices to generate a new version of the model, and the device downloads the new version of the model and performs differential fine-tuning; realize knowledge sharing of multiple devices, and at the same time, retain the individual characteristics of the device through 30% local weight.

[0055] Specifically, the formula of differential fine-tuning is wherein θ_local is the current parameter vector of the local model of the device, θ_cloud is the parameter vector of the new version of the model generated by the cloud aggregation, and 0.3 is the differential fine-tuning coefficient, which realizes the retention of 30% local characteristics to prevent knowledge forgetting. The formula is the interface of federated learning and local adaptation, which solves the contradiction between defense degradation in offline environment and global model overfitting to local data by quantifying the weight.

[0056] (2) Offline model update, when the device is offline, load the local cache of the last 5 rounds of perturbation model version and behavior record data, where the perturbation model version adopts ONNX format and the single model size is not more than 3MB, which is suitable for the storage space of embedded devices. The behavior record data is compressed by ProtoBuf and the retention time is 72 hours, and incremental update is realized through gradient accumulation; ensure the basic protection capability under the condition of no network connection, guarantee the continuous protection capability during 72 hours of network interruption, and 3MB model volume is suitable for the storage limit of embedded devices.

[0057] (3) Reinforcement learning optimization, deep Q learning is adopted to optimize the detection structure perturbation strategy, where the state is defined as the current perturbation parameter, detection accuracy and resource consumption, the action is defined as adjusting the pruning rate ± 10% or replacing the activation function, and the reward function is , through experience replay and target network stable learning, the cache size of experience replay is 2000, and the target network is updated once every 24 hours. Realize the dynamic optimization of perturbation strategy, experience replay cache 2000 guarantee the stability of training, 24 hours target network update balance real-time and convergence.

[0058] This step balances the local and cloud models through 0.3 coefficient differential fine-tuning, loads 5 rounds of ≤3MB ONNX model and 72 hours of ProtoBuf data for offline incremental update when offline, and combines reinforcement learning, such as action ± 10% pruning rate / 24 hour target update, to solve the core contradiction between offline protection degradation and personalized feature preservation.

[0059] The generation rule of the local cache model patch of the embodiment is: calculate the difference between the current model and the last round model after each round of model update, and store the difference part as a patch file when the difference exceeds 0.1, the patch file adopts Delta format and the size is not more than 500KB, and the latest 3 patch files are loaded for incremental update when offline.

[0060] The generation rule of the local cache model patch filters irrelevant parameter fluctuations, reduces the number of patches, and only stores patches when the model changes more than 0.1, reducing 70% of invalid patch generation. Store the patch using Delta binary difference algorithm, file ≤500KB, guarantee 85% improvement of embedded device storage space utilization. When offline, apply the latest 3 patches in sequence to realize that the local model detection accuracy decays ≤5% within 72 hours.

[0061] This rule selects significant model changes through a difference threshold of >0.1, compresses the patch to 500KB in Delta format to adapt to embedded storage, loads the latest 3 patches for incremental update when offline, and solves the problem of protection failure caused by outdated models in offline environment.

[0062] The brain-adapted feedback mechanism in step S4 of the embodiment specifically includes: (1) Short-term memory simulation, store the behavior sequence of the storage device in the last 24 hours through the short-term memory unit, the behavior sequence includes user operation behavior, system resource call information and detection results; provide a daily behavior baseline through the behavior sequence of the storage device in the last 24 hours, and guarantee the integrity of the behavior deviation analysis.

[0063] (2) Abnormal reactivation, review the behavior sequence not confirmed as attack every 30 minutes through the abnormal reactivation unit, calculate the behavior deviation rate, and the behavior deviation rate is the Euclidean distance difference value of the current behavior vector and the historical behavior graph; review the behavior sequence not confirmed as attack every 30 minutes, calculate the Euclidean distance difference value of the current behavior vector and the historical behavior graph as the behavior deviation rate, and realize the quantification of the abnormal degree of the unconfirmed behavior.

[0064] (3) Strategy adaptation, dynamically update the detection threshold and response strategy based on the behavior deviation rate, and the update formula is , wherein γ is the adjustment rate, the value range of γ is 0.1 to 0.3, θ is the strategy threshold, and BDR is the behavior deviation rate. Dynamically update the threshold and strategy based on the behavior deviation rate, execute adjustment according to the update formula, realize controllable evolution of the detection threshold, and the value of γ is 0.1 to 0.3 to guarantee the adjustment amplitude anti-shock.

[0065] The brain-like adaptive feedback mechanism establishes a short-term memory library through 24-hour behavior sequence storage, generates a behavior deviation rate through 30-minute-level review and Euclidean distance calculation, and realizes the autonomous optimization of the detection strategy in the offline environment in combination with the threshold update formula regulated by γ ∈ [0.1, 0.3], thereby directly solving the problem of defense degradation in the offline environment.

[0066] The semantic abnormal instruction recognition of step S5 of the embodiment specifically includes: (1) Semantic feature extraction, extract the semantic embedding vector of the control instruction based on the RoBERTa-base pre-training model, and generate the semantic-behavior joint feature in combination with the context behavior sequence, wherein the context behavior sequence is the first 5 instructions of the control instruction, the hidden layer dimension of the RoBERTa-base pre-training model is 768; realize the associated modeling of the instruction intention and the behavior context.

[0067] (2) Abnormal detection, calculate the semantic deviation degree using the isolation forest, and determine that the malicious instruction is correct in format but abnormal in semantics when the semantic deviation degree exceeds 0.7, wherein the number of trees of the isolation forest is 100; realize the quantitative determination of the semantic malicious degree.

[0068] (3) Feature recording, generate the uniform time sequence feature vector of the malicious instruction, store the uniform time sequence feature vector to the local training dataset, and the local training dataset adopts the SQLite database; realize the persistent retention of the attack features.

[0069] This step extracts 768-dimensional semantic vectors associated with the behavior sequence of the first 5 instructions through the RoBERTa-base model, calculates the deviation degree using 100 isolated forest trees, and determines malicious instructions when the threshold exceeds 0.7. Finally, the unified time sequence feature vector is stored in the SQLite database, accurately intercepts format compliant but semantically abnormal fraud attacks, and solves the problem of high semantic attack missed detection rate.

[0070] The neuromorphic computing acceleration of step S6 specifically includes: (1) Hardware foundation, a HfO2-based 1T1R resistive memory array is used to construct a sparse synapse path, the number of cells of the resistive memory array is 8x8 and the thickness of the resistive layer is 5nm; a hardware-level computing-storage integrated architecture is realized to eliminate the data movement overhead of the traditional von Neumann architecture.

[0071] (2) Event-driven mechanism, when the change rate of the input feature value exceeds 10%, the corresponding resistive memory cell is activated for calculation; sparse calculation triggered by events is realized to reduce 90% of invalid calculation energy consumption.

[0072] (3) Energy consumption optimization, the working voltage is limited to 0.8V-1.2V through pulse width modulation. The average energy consumption is reduced by more than 30% compared with the general ARM Cortex-A55 CPU. It meets the strict power consumption constraints of home devices.

[0073] This step realizes the integration of storage and calculation through the 8x8 hardware structure of the 5nm HfO2 resistive array, performs sparse calculation based on the event triggering mechanism with a change rate of >10%, and combines 0.8V-1.2V pulse voltage modulation to compress the peak power consumption to below 1.8W, completely solving the core contradiction between high energy consumption and real-time performance.

[0074] On the other hand, a quasipattern intrusion detection system for home devices is also provided, which is used to implement the above-mentioned quasipattern intrusion detection method, comprising: A system trusted boot unit is used to perform integrity verification on the function modules of the detection system during the device startup stage, and load the pre-trained detection model and policy library after verification; a trusted execution environment is established to block tampered backdoor attacks, solving the problem of easy bypass of static models. The system trusted boot unit is a UEFI SecureBoot module or an ARM TrustZone module.

[0075] UEFI SecureBoot module: perform firmware-level SHA-256 hash verification, and load the GRUB boot program to start the detection system after verification.

[0076] ARM TrustZone security extension module: isolates the secure domain in the Cortex-M series chip, and decompresses the model to the TEE environment after the integrity verification is passed.

[0077] Specifically, the integrity verification is an SHA-256 hash verification on the program codes and configuration files of the multi-modal feature acquisition module, the dynamic disturbance detection module, the hierarchical elastic response module, the disturbance type model updating module, the semantic pseudo-instruction identification module, the neuromorphic security enhancement module and the brain-like adaptive feedback module; the pre-trained detection model adopts an ONNX format, and the policy library adopts a JSON format.

[0078] The multi-modal feature acquisition module is used for acquiring multi-modal original data and generating a unified time sequence feature vector through a modal alignment mechanism; and heterogenous data time sequence deviation is eliminated, and a multi-modal fusion failure problem is solved. The multi-modal feature acquisition module is a ROS (Robot Operating System) multi-source data synchronization framework or an Apache Kafka stream processing engine.

[0079] The ROS (Robot Operating System) multi-source data synchronization framework aligns network flow (TCPDump), operation log (Auditd) and audio / video stream (GStreamer) timestamps through a message_filters module, and the error is less than or equal to 100 ms.

[0080] The Apache Kafka stream processing engine: creates an independent Topic partition to cache each modal data, and generates a 128-dimensional unified feature vector through a WindowJoin operation.

[0081] Specifically, the multi-modal feature acquisition module includes a modal synchronization submodule, a feature standardization submodule and a time window alignment submodule; the modal synchronization submodule is synchronized based on a timestamp and the error is not more than 100 ms; the feature standardization submodule adopts Z-score normalization; the time window alignment submodule adopts a 200 ms sliding window; the multi-modal original data includes network flow, system log, user operation behavior and environmental audio / video, wherein the network flow includes TCP / IP data packets, the system log includes process call log and file operation log, the user operation behavior includes APP click behavior and voice instruction, and the environmental audio / video includes camera screen and microphone audio; the unified time sequence feature vector is generated by fusing multi-modal features through a self-attention layer, and the dimension is 128, wherein the network flow feature accounts for 30%, the system log feature accounts for 25%, the user operation behavior feature accounts for 20%, and the environmental audio / video feature accounts for 25%, and the elements are normalized to the interval of -1 to 1.

[0082] A dynamic perturbation detection module is configured to perform perturbation-based joint detection on the unified time sequence feature vector based on a fusion deep learning model and a biological heuristic algorithm, and output a credibility score value; the dynamic perturbation detection module breaks the attack path dependency and solves the real-time attack bypass problem of unknown vulnerabilities.

[0083] The PyTorch-Geometric graph neural network library is used to construct a GCN layer to realize edge weight ±30% perturbation (torch_geometric.utils.dropout_edge).

[0084] The DEAP evolutionary computing framework is configured to implement an artificial immune negative selection algorithm and clone selection to generate 100 self templates.

[0085] Specifically, the dynamic perturbation detection module adopts a hybrid model that fuses a graph convolutional neural network and an artificial immune negative selection algorithm, wherein the graph convolutional neural network is used to extract the correlation pattern in the feature vector, and the artificial immune negative selection algorithm is used to generate a self template to detect abnormalities; the dynamic perturbation detection module integrates a detection path perturbation engine, which performs dynamic perturbation on the structure of the graph convolutional neural network every 60±10 seconds, and the dynamic perturbation includes graph topology perturbation, activation function replacement, and dynamic adjustment of the model pruning rate, wherein the graph topology perturbation is to adjust the edge weight of the adjacency matrix to within ±30% or change the node connection relationship and the node connection number change rate is 15%-25%, the activation function replacement is to replace ReLU with Swish, LeakyReLU, or GELU with a probability of 15%-25%, and the dynamic adjustment of the model pruning rate is to calculate the pruning rate based on the input complexity index and the pruning rate fluctuates within 5%-30%; the credibility score value is generated by weighted fusion of the feature correlation degree output by the graph convolutional neural network and the abnormal matching degree output by the artificial immune negative selection algorithm, wherein the weight ratio of the feature correlation degree and the abnormal matching degree is 6:4, the value range is 0 to 1, and the larger the value, the higher the normal probability; the self template generation method of the artificial immune negative selection algorithm is to collect 1000 feature vectors when the device is normally running, generate 100 self templates through clone selection, each self template corresponds to a normal behavior pattern, and when the Hamming distance between the feature vector and all self templates is greater than 0.6, it is determined to be abnormal.

[0086] A hierarchical elastic response module is configured to trigger a hierarchical elastic response strategy according to the credibility score value, and execute corresponding security operations; the hierarchical elastic response module dynamically adapts the threat level to optimize the balance between resource consumption and security.

[0087] The Linux Auditd logging system is used to configure audit.rules to implement a lenient monitoring mode (operation sequence recorded every 5 minutes).

[0088] The Kubernetes service mesh is used to implement an isolation mode (service port migration to a Shadow Cluster) via Istio VirtualService.

[0089] Specifically, the hierarchical elastic response strategy is a five-level response strategy based on the credibility score value, and the threshold can be dynamically adjusted: When the credibility score value is ≥ 0.9, enter the trust mode, the system runs normally and no additional operations are triggered; When 0.7≤ credibility score value < 0.9, enter the lenient monitoring mode, enable light logging and behavior tracking, where light logging is to save user operation sequence every 5 minutes, and behavior tracking is to mark abnormal process ID; When 0.4≤ credibility score value < 0.7, enter the observation mode, start deep monitoring, record user operation behavior and system resource call information, and generate behavior pattern templates through spectral clustering, where the system resource call information includes CPU occupancy and memory usage, and the number of clusters of spectral clustering is 3; When 0.2≤ credibility score value < 0.4, enter the induction mode, build a pseudo environment to mislead the attacker to perform invalid operations; When the credibility score value is < 0.2, enter the isolation mode, and forcibly execute security operations, including service port migration, control interface isolation, or system function degradation; Specifically, the pseudo environment construction in the induction mode specifically includes: (1) False interface simulation, generating a false service interface with a format consistent with the specification, where the false service interface includes an HTTP API with a path of / decoy / [device ID] / status, and the false service interface responds to the attacker's query instruction but returns false data; (2) State camouflage, creating a virtual state matrix in the device memory, the virtual state matrix has a dimension of 16x16 and a data type of float32, the simulation value of the virtual state matrix is updated every 500 milliseconds, and the fluctuation range of the simulation value is temperature ± 0.5℃, humidity ± 3%RH; (3) Path redirection, building a sandbox environment through Linux namespace, redirecting the attacker's control command to the / dev / null virtual device, while recording the semantic features of the control command.

[0090] A perturbation model updating module is configured to update the detection model and the strategy through a distributed model updating mechanism, support local model incremental updating in an offline state, ensure continuous protection in a network interruption environment, and solve the problem of network interruption defense degradation. The perturbation model updating module is a combination of a TensorFlow Federated framework and a SQLite database.

[0091] The TensorFlow Federated framework is configured to implement federated learning gradient aggregation.

[0092] The SQLite embedded database is configured to store 72-hour behavior record data compressed by ProtoBuf (sqlite3.BLOB type field).

[0093] Specifically, the perturbation model updating module includes a cloud model updating unit and an offline model updating unit. The cloud model updating unit adopts a federated learning mechanism, and the device uploads the gradient of the local model to the cloud. The cloud aggregates the gradients of multiple devices to generate a new version of the model. The device downloads the new version of the model and performs differential fine-tuning. The formula for differential fine-tuning is When the device is disconnected from the network, the offline model updating unit loads the latest 5 rounds of perturbation model versions and behavior record data cached locally. The perturbation model versions are in ONNX format and the size of a single model does not exceed 3 MB. The behavior record data is compressed by ProtoBuf and the retention period is 72 hours. Incremental updating is achieved through gradient accumulation. The local cache model patch generation rule of the perturbation model updating module is as follows: after each round of model updating, the difference between the current model and the last round of model is calculated. When the difference exceeds 0.1, the difference is stored as a patch file. The patch file is in Delta format and the size does not exceed 500 KB. The latest 3 patch files are loaded for incremental updating when the device is disconnected from the network.

[0094] A brain-like adaptive feedback module is configured to optimize detection parameters and response strategies in combination with reinforcement learning. The brain-like adaptive feedback module is a combination of an OpenAI Gym reinforcement learning environment and a RayRLlib distributed training library.

[0095] The OpenAI Gym reinforcement learning environment is configured to define the state space (perturbation parameters + accuracy + resource consumption) and the action space (±10% pruning rate).

[0096] The RayRLlib distributed training library is configured to deploy a deep Q network (DQNConfig) to implement strategy optimization.

[0097] Specifically, the brain-like adaptive feedback module comprises a short-term memory unit, an abnormal reactivation unit and a strategy adaptive unit; the short-term memory unit is used to store the behavior sequence of the device in the last 24 hours, the behavior sequence comprising user operation behavior, system resource calling information and detection results; the abnormal reactivation unit reviews the behavior sequence not confirmed as an attack every 30 minutes, calculates the behavior deviation rate, the behavior deviation rate being the Euclidean distance difference between the current behavior vector and the historical behavior graph; the strategy adaptive unit dynamically updates the detection threshold and the response strategy based on the behavior deviation rate, the update formula being wherein γ is the adjustment rate, the value range being 0.1 to 0.3, θ is the strategy threshold, and BDR is the behavior deviation rate.

[0098] The semantic pseudo-instruction identification module is used to identify malicious control instructions with correct format but abnormal semantics, and record the feature vectors thereof; intercept semantic deception attacks and solve the problem of semantic attack missed detection. The semantic pseudo-instruction identification module is a combination of HuggingFace Transformers library and Scikit-learn Isolation Forest algorithm.

[0099] The HuggingFace Transformers library is used to load the roberta-base model (768-dimensional hidden layer) to extract instruction semantic embedding.

[0100] The Scikit-learn Isolation Forest algorithm is used to calculate the deviation degree by IsolationForest(n_estimators=100), and >0.7 is judged as abnormal.

[0101] Specifically, the semantic pseudo-instruction identification module comprises a semantic feature extraction unit, an abnormality detection unit and a feature recording unit; the semantic feature extraction unit extracts the semantic embedding vector of the control instruction based on the RoBERTa-base pre-training model, and generates semantic-behavior joint features in combination with the context behavior sequence, wherein the context behavior sequence is the first 5 instructions of the control instruction, and the hidden layer dimension of the RoBERTa-base pre-training model is 768; the abnormality detection unit calculates the semantic deviation degree using Isolation Forest, and determines that it is a malicious instruction with correct format but abnormal semantics when the semantic deviation degree exceeds 0.7, wherein the number of trees of Isolation Forest is 100; the feature recording unit generates the uniform time sequence feature vector of the malicious instruction, and stores the uniform time sequence feature vector to the local training dataset, wherein the local training dataset adopts SQLite database.

[0102] A neuromorphic security enhancement module is used to accelerate detection calculation and optimize energy consumption based on an event-driven mechanism of neuromorphic hardware. The peak power consumption of the neuromorphic security enhancement module is ≤1.8W, breaking the 3W energy consumption bottleneck of traditional solutions. The neuromorphic security enhancement module is an Intel Loihi neuromorphic chip or an IBM TrueNorth pulse neural network compiler.

[0103] The Intel Loihi neuromorphic chip is used to deploy an 8x8 cross array (analog HfO2 resistive switching device) to perform event-driven calculation.

[0104] The IBM TrueNorth pulse neural network compiler is used to compile a detection model into a neuromorphic executable file (.nsnet format).

[0105] Specifically, the neuromorphic security enhancement module adopts an HfO2-based 1T1R resistive switching memory array to construct a sparse synapse path, the number of cells of the resistive switching memory array is 8x8, and the thickness of the resistive switching layer is 5nm; the neuromorphic security enhancement module activates the corresponding resistive switching memory cell for calculation when the change rate of the input feature value exceeds 10% through an event-driven mechanism; the neuromorphic security enhancement module limits the working voltage to 0.8V-1.2V through pulse width modulation, and the average energy consumption is reduced by more than 30% compared with a general ARM Cortex-A55 CPU.

[0106] Specifically, the graph convolutional neural network of the multi-modal feature acquisition module adopts a 2-layer GCN structure, the dimension of each layer hidden layer is 64, and the activation function is ReLU.

[0107] The spectral clustering of the hierarchical elastic response module adopts K-means initialization, the number of iterations is 100 times, and the convergence threshold is 1e-5.

[0108] The system starts to establish a trusted root through integrity verification, eliminates data deviation through multi-modal alignment and fusion, detects real-time attacks that collapse unknown vulnerabilities through 60-second periodic disturbance, realizes ≤1.8W peak power consumption in combination with an event-driven neuromorphic hardware, and solves the industry problems of high precision, low power consumption and strong robust protection of smart home devices.

[0109] The above is only a preferred embodiment of the present application, and for those skilled in the art, according to the idea of the present application, the specific implementation manner and application range can be changed, and the content of the description should not be understood as a limitation of the present application.

Claims

1. A method for detecting intrusion into home appliances, characterized in that, Includes the following steps: Step S0: Start trusted verification and initialization. When the device starts, the system trusted boot unit is called to perform integrity verification on the functional modules of the detection system. After the verification is passed, the pre-trained detection model and strategy library are loaded. Step S1: Multimodal feature fusion processing, collecting original multimodal data and generating a unified temporal feature vector through a modality alignment mechanism; Step S2: Perturbation-based joint detection: The dynamic perturbation detection module, which integrates deep learning models and bio-inspired algorithms, performs perturbation-based joint detection on the unified temporal feature vector and outputs a credibility score. Step S3: Tiered Resilient Response. Based on the confidence score, a tiered resilient response strategy is triggered to execute the corresponding security operation. Step S4: Adaptive model update. The detection model and strategy are updated through a distributed model update mechanism, which supports incremental updates of the local model in offline mode. The detection parameters and response strategies are optimized by combining a brain-like adaptation feedback mechanism and reinforcement learning. Step S5: Semantic anomaly command identification. Identify malicious control commands that are correctly formatted but semantically abnormal, and record their feature vectors. Step S6: Accelerate neuromorphic computing by using an event-driven mechanism based on neuromorphic hardware to accelerate detection computation and optimize energy consumption.

2. The method according to claim 1, characterized in that, The perturbation-based joint detection in step S2 specifically includes: (1) Detection model architecture: A hybrid model combining graph convolutional neural network and artificial immune negative selection algorithm is adopted, wherein the graph convolutional neural network is used to extract the association pattern in the feature vector, and the artificial immune negative selection algorithm is used to generate self-template to detect anomalies; (2) Structural perturbation mechanism: The detection path perturbation engine is integrated. The structure of the graph convolutional neural network is dynamically perturbed every 60 seconds ± 10 seconds. The dynamic perturbation includes graph topology perturbation, activation function replacement and dynamic adjustment of model pruning rate. Graph topology perturbation is to adjust the edge weight of the adjacency matrix to within ± 30% or change the node connection relationship and the node connection number change rate is 15%-25%. Activation function replacement is to replace ReLU with Swish, LeakyReLU or GELU with a probability of 15%-25%. The dynamic adjustment of model pruning rate is to calculate the pruning rate based on the input complexity index and the pruning rate fluctuates within the range of 5%-30%. (3) Joint detection output: the graph convolutional neural network outputs the feature correlation degree, the artificial immune negative selection algorithm outputs the abnormal matching degree, and the confidence score is generated by weighted fusion. The weight ratio of feature correlation degree to abnormal matching degree is 6:

4. The confidence score ranges from 0 to 1, and the larger the value, the higher the probability of normality.

3. The method according to claim 2, characterized in that, The self-template generation method of the artificial immune negative selection algorithm is as follows: 1000 feature vectors are collected when the device is running normally, and 100 self-templates are generated by cloning selection. Each self-template corresponds to a normal behavior pattern. When the Hamming distance between the detected feature vector and all self-templates is greater than 0.6, it is judged as abnormal.

4. The method according to claim 1, characterized in that, The graded elastic response in step S3 specifically includes: A five-level response strategy based on credibility scores, where the thresholds for the five levels can be dynamically adjusted: When the credibility score is greater than or equal to 0.9, the system enters trust mode, operates normally, and does not trigger any additional operations. When the credibility score is greater than or equal to 0.7 and less than 0.9, it enters the relaxed monitoring mode, and enables light logging and behavior tracking. Light logging saves the user operation sequence every 5 minutes, and behavior tracking marks abnormal process IDs. When the credibility score is greater than or equal to 0.4 and less than 0.7, the system enters the observation mode, starts deep monitoring, records user operation behavior and system resource call information, and generates a behavior pattern template through spectral clustering. The system resource call information includes CPU utilization and memory usage. The number of clusters in the spectral clustering is 3. When the credibility score is greater than or equal to 0.2 and less than 0.4, the system enters the inducement mode, constructing a pseudo-environment to mislead attackers into performing invalid operations. When the credibility score is less than 0.2, the system enters isolation mode and enforces security operations, which include service port migration, control interface isolation, or system function degradation.

5. The method according to claim 4, characterized in that, The construction of pseudo-state environments in induced patterns specifically includes: (1) Fake interface simulation: generate fake service interfaces in a format that conforms to the specifications. The fake service interfaces include an HTTP API with the path / decoy / [device ID] / status. The fake service interfaces respond to the attacker's query commands but return fake data. (2) State masquerading: A virtual state matrix is ​​created in the device memory. The virtual state matrix has a dimension of 16×16 and a data type of float32. The simulated value of the virtual state matrix is ​​updated every 500 milliseconds. The fluctuation range of the simulated value is temperature ±0.5℃ and humidity ±3%RH. (3) Path redirection: A sandbox environment is built using Linux namespaces to redirect the attacker's control commands to the / dev / null virtual device, while recording the semantic features of the control commands.

6. The method according to claim 1, characterized in that, Step S4, the adaptive model update, specifically includes: (1) Cloud model update: The federated learning mechanism is adopted. The device uploads the gradient of the local model to the cloud. The cloud aggregates the gradients of multiple devices to generate a new version of the model. The device downloads the new version of the model and performs differential fine-tuning. (2) Offline model update: When the device is offline, load the latest 5 rounds of perturbation model version and behavior record data cached locally. The perturbation model version adopts ONNX format and the size of a single model does not exceed 3MB. The behavior record data adopts ProtoBuf compression and is retained for 72 hours. Incremental update is achieved through gradient accumulation. (3) Reinforcement learning optimization: Deep Q-learning is used to optimize the detection structure perturbation strategy. The state is defined as the current perturbation parameters, detection accuracy, and resource consumption. The action is defined as adjusting the pruning rate by ±10% or replacing the activation function. The reward function is... Through experience replay and stable learning with the target network, the cache size of the experience replay is 2000, and the target network is updated every 24 hours.

7. The method according to claim 6, characterized in that, The local cache model patch generation rule is as follows: after each round of model update, calculate the difference between the current model and the previous model. When the difference exceeds 0.1, store the difference as a patch file. The patch file adopts Delta format and the size does not exceed 500KB. When the network is disconnected, load the three most recent patch files for incremental update.

8. The method according to claim 1, characterized in that, The brain-like adaptation feedback mechanism in step S4 specifically includes: (1) Short-term memory simulation: The behavior sequence of the device in the past 24 hours is stored in the short-term memory unit. The behavior sequence includes user operation behavior, system resource call information and detection results; (2) Abnormal reactivation: Every 30 minutes, the abnormal reactivation unit reviews the behavior sequences that are not confirmed as attacks and calculates the behavior offset rate. The behavior offset rate is the difference in Euclidean distance between the current behavior vector and the historical behavior map. (3) Policy adaptation: The detection threshold and response policy are dynamically updated based on the behavior offset rate. The update formula is as follows: , where γ is the adjustment rate, and the value of γ ranges from 0.1 to 0.3, θ is the policy threshold, and BDR is the behavior offset rate.

9. The method according to claim 1, characterized in that, Step S5, the identification of semantically abnormal instructions, specifically includes: (1) Semantic feature extraction: Based on the RoBERTa-base pre-trained model, the semantic embedding vector of the control command is extracted, and the semantic-behavior joint feature is generated by combining the context behavior sequence. The context behavior sequence is the first 5 commands of the control command. The hidden layer dimension of the RoBERTa-base pre-trained model is 768. (2) Anomaly detection: use isolated forest to calculate semantic deviation. When the semantic deviation exceeds 0.7, it is judged as a malicious instruction with correct format but abnormal semantics. The number of trees in the isolated forest is 100. (3) Feature recording: Generate a unified temporal feature vector of malicious instructions and store the unified temporal feature vector in the local training dataset. The local training dataset uses an SQLite database. The acceleration of neuromorphic computation in step S6 specifically includes: (1) Hardware foundation: Sparse synaptic paths are constructed using HfO2-based 1T1R resistive switching memory arrays. The number of cells in the resistive switching memory array is 8×8 and the thickness of the resistive switching layer is 5nm. (2) Event-driven mechanism: when the rate of change of the input feature value exceeds 10%, the corresponding resistive random access memory unit is activated for calculation; (3) Energy consumption optimization: the working voltage is limited to 0.8V-1.2V by pulse width modulation.

10. A mimicry intrusion detection system for home appliances, used to implement the mimicry intrusion detection method according to any one of claims 1-9, characterized in that, include: The system trusted boot unit is used to perform integrity verification on the functional modules of the detection system during the device startup phase. After the verification is passed, the pre-trained detection model and strategy library are loaded. The multimodal feature acquisition module is used to acquire multimodal raw data and generate a unified temporal feature vector through a modality alignment mechanism; The dynamic perturbation detection module is used to perform perturbation joint detection on the unified temporal feature vector based on the fusion deep learning model and the bio-inspired algorithm, and output a confidence score. The graded resilience response module is used to trigger the graded resilience response strategy based on the credibility score and execute the corresponding security operation. The perturbation-based model update module is used to update the detection model and strategy through a distributed model update mechanism, supporting local incremental model updates in offline mode. A neuromorphic adaptive feedback module is used to combine reinforcement learning to optimize detection parameters and response strategies; The semantic pseudo-instruction recognition module is used to identify malicious control instructions that are correctly formatted but semantically abnormal, and to record their feature vectors. The neuromorphic security enhancement module is used to accelerate detection computation and optimize energy consumption based on the event-driven mechanism of neuromorphic hardware.

Images