Edge computing protection system based on node protection and method thereof

Abstract

The application relates to the technical field of edge computing security protection, and discloses an edge computing protection system based on node protection and a method thereof. The system comprises a trusted microkernel security pedestal layer, a cognitive immune evolution engine layer, a multi-dimensional evolution game trust network layer, a resource-aware lightweight quorum cheating layer, a differential privacy federated threat intelligence network layer, and an infection graph-driven elastic self-healing topology layer. The system realizes adaptive attack detection through antibody population evolution driven by a clone selection algorithm, realizes distributed node trust management without centralization through multi-dimensional time decay trust score and evolution game equilibrium analysis, realizes active cheating defense through adaptive resource declaration of shadow nodes and network attribute rotation, realizes cross-node threat intelligence sharing under privacy protection through differential privacy gradient weighted aggregation, and realizes preventive blocking and topology immune remodeling for horizontal attacks through infection graph risk diffusion probability calculation.

Description

Technical Field

[0001] This invention relates to the field of edge computing security protection technology, and more particularly to an edge computing protection system and method based on node protection. Background Technology

[0002] Edge computing moves computing resources and data processing capabilities from cloud centers to network edge nodes, enabling three types of highly latency-sensitive services—Industrial IoT sensor data aggregation, real-time traffic signal control, and visual quality inspection of production lines—to be processed at the location closest to the data source. However, edge nodes have inherent characteristics such as dispersed deployment, large number, limited hardware resources, unattended operation, and difficulty in guaranteeing physical security. These characteristics make edge nodes a priority target for attackers. Existing security protection solutions for edge computing scenarios mainly follow three technical paths: the first is an intrusion detection system based on a fixed signature rule base, which performs feature matching on network traffic arriving at the node through pre-set attack signatures; the second is a centralized trust authentication system based on public key infrastructure, where a central server uniformly maintains the identity certificates of edge nodes and centrally authorizes interaction requests between nodes; and the third is a deception defense technology based on static honeypot deployment, which pre-configures fixed decoy nodes in the network to capture attacker behavior.

[0003] However, all three existing technical solutions mentioned above have fundamental technical flaws and cannot meet the actual security protection needs of edge computing networks: Intrusion detection systems based on fixed signature rule bases can only identify known attack features already included in the rule base, lacking the ability to detect zero-day exploits and new attack payloads. Furthermore, rule base updates rely on a centralized distribution mechanism, which cannot be updated in a timely manner in network partition scenarios, leaving edge nodes exposed to undetectable attack threats during periods without updates; Centralized trust authentication systems based on public key infrastructure have a single point of failure risk in large-scale distributed edge computing networks. If the central authentication server is attacked or a network partition occurs, the entire network relies on this server. The edge nodes of the server cannot complete legitimate authentication, and there is a fundamental contradiction between the communication latency introduced by centralized authentication and the real-time requirements of edge computing. The deception defense scheme based on static honeypots deploys decoy nodes in a fixed network location. Attackers can identify the network fingerprint characteristics of the honeypot through early reconnaissance and actively avoid it. Moreover, static honeypots cannot dynamically adjust their decoy characteristics according to the attacker's real-time detection behavior, causing the deception defense effect to rapidly decline as reconnaissance technology improves. The above three types of technologies lack a coordination mechanism. When a single node is damaged, it cannot trigger proactive preventive measures against neighboring nodes. Attackers can carry out lateral penetration along the trust relationship chain between nodes without being blocked. Summary of the Invention

[0004] The purpose of this invention is to provide an edge computing protection system and method based on node protection, so as to solve the problems existing in the prior art.

[0005] This invention provides an edge computing protection system and method based on node protection, comprising a trusted microkernel security base layer, a cognitive immune evolutionary engine layer, a multidimensional evolutionary game trust network layer, a resource-aware lightweight mimicry deception layer, a differential privacy federated threat intelligence network layer, and an infection graph-driven elastic self-healing topology layer that are interconnected. The trusted microkernel security base layer includes a secure microkernel running in the trusted execution environment of the edge node processor. The code integrity metric hash value of the secure microkernel is written to the processor platform configuration register. The secure microkernel generates an identity anchor value by concatenating the edge node processor serial number and the unique identifier of the security chip and performing a hash operation. Encapsulated in a sealed storage area within a trusted execution environment, the secure microkernel includes a system call whitelist composed of system call types initiated by business processes under standard operating conditions. System call requests not on the whitelist are intercepted by the secure microkernel. The cognitive immune evolutionary engine layer includes an innate immune subsystem that stores a set of known attack feature rules in a multi-pattern matching data structure, and a subsystem that maintains... An adaptive immune subsystem consisting of an antibody population, each antibody being composed of... The antibody population is trained to tolerate immune conditions by fusing network session traffic statistics, system call frequency sequence features, resource usage distribution features, and service response latency percentile features collected during normal node operation. The multidimensional evolutionary game-theoretic trust network layer includes node identity anchor values. The verification module for the elliptic curve-based non-interactive zero-knowledge proof protocol, and the behavioral trust component. Capability and Trust Weight Identity and Trust Weight and environmental trust weight The trust scoring module quantifies the comprehensive trust score of each node across four dimensions; the resource-aware lightweight mimicry deception layer includes shadow nodes deployed as container processes running in independent network namespaces on each edge node, and a network attribute rotation module that uses a hash chain structure cryptographically secure pseudo-random sequence generator to update the service listening port set, transport layer protocol version identifier, and transport layer secure handshake parameters of the real node at fixed time intervals. The service port set, service version information, and hardware resource declaration parameters of the shadow nodes are configured with fake parameters that are higher than the permission level of the real node; the differential privacy federated threat intelligence network layer includes the execution of locally trained gradient vectors on nodes. The privacy protection module includes norm pruning and Gaussian noise injection, and a secure aggregation module that performs weighted aggregation of the processed gradient vector in the encrypted domain using the ratio of each node's trust score to the upper limit of the trust score as the reputation weight. The infection graph-driven elastic self-healing topology layer includes an infection graph constructed with each edge node as the vertex and the network connections between nodes as directed weighted edges, and a fast migration module that transmits a sealed snapshot of the service status of the damaged node to the alternative node via an encrypted channel and completes the service takeover within 500 milliseconds. The edge weights of the directed weighted edges are determined by the weighted average of three indicators: normalized bandwidth, latency stability, and historical connectivity success rate of the corresponding network path.

[0006] The above technical solution, through the organic combination and synergistic linkage of six functional layers, produces the following overall technical effect: the trusted microkernel security base layer anchors identity values. Encapsulated in a sealed storage area within the processor's trusted execution environment, and employing a dynamic system call whitelist mechanism to minimize the range of callable system interfaces during periods of escalating threat levels, an unbreakable trust chain starting point is established at the hardware layer. The cognitive immune evolutionary engine layer eliminates false positives for normal business traffic through immune tolerance training based on self-antigen maps, and continuously enhances the attack detection capability of the antibody population with each attack event through clonal selection and affinity maturation iteration, achieving adaptive evolution of the detection mechanism. The multidimensional evolutionary game-theoretic trust network layer establishes a distributed dynamic trust management mechanism that can withstand long-term, latent trust hijacking attacks without the need for a central certification authority, using an elliptic curve-based zero-knowledge proof protocol and a multidimensional trust scoring formula including a time decay term. Resource awareness... The lightweight mimicry deception layer achieves proactive deception of attackers and proactive countermeasures against fingerprint scanning reconnaissance simultaneously through adaptive resource declaration adjustment of shadow nodes and cryptographically secure network attribute rotation. The differential privacy federated threat intelligence network layer achieves collaborative sharing of cross-node threat intelligence while protecting the local data privacy of each node through gradient-weighted aggregation with node trust scores as reputation weights, and eliminates the pollution of the global model by malicious gradients of compromised nodes through a weight suppression mechanism. The infection graph-driven elastic self-healing topology layer preventively blocks lateral movement attack paths through iterative calculation of attack risk diffusion probability, and completely eliminates the attacker's knowledge advantage of the original network structure through topology immune reshaping after the damaged node is repaired and re-enters, so that the overall protection capability of the system does not decrease but increases after being attacked.

[0007] Furthermore, the trust scoring module calculates the nodes using the following formula. At any moment Overall trust score

[0008] : ,

[0009] in, , , The weight coefficients of the trust components in each dimension are given, and satisfy the following conditions: , The decay rate coefficient, This is the start time of the current evaluation period. For the node at time The occurrence of the first The change in rating for each trust event is represented by a positive value for a positive event and a negative value for a suspicious event. For when The time value is Indicator functions, Deadline The total number of trust events; the multidimensional evolutionary game trust network layer searches for an evolutionary stable equilibrium point in the repeated game payoff matrix composed of two strategies: honest cooperation and opportunistic deviation, using simulated annealing algorithm, and sets the trust score threshold corresponding to the equilibrium point as the node service interaction admission threshold. The overall trust score is lower than The node's service request was rejected.

[0010] Furthermore, the adaptive immune subsystem extracts attack feature vectors as antigens when it detects anomalous behavior. Calculate the value of each antibody in the antibody population using the following formula. With antigen Affinity between : ,in, The distance between the binarized antibody feature weight vector and the antigen feature vector is the Hamming distance. The length of the antibody feature vector; after the adaptive immune subsystem sorts the antibody population in descending order of affinity, the affinity ranking is calculated using the following formula. Number of antibody clones amplified at the position : ,in, For the range of values ​​within to Cloning scaling factor between Antibody The affinity is ranked in descending order; the standard deviation of the mutation amplitude of the clone is inversely proportional to the affinity of the parent antibody. Mutants with higher affinity than the parent are retained in the next generation population, while mutants with lower affinity than the parent are eliminated.

[0011] Furthermore, the privacy protection module will use the original gradient vector generated by node n during local training. of Norm clipping to sensitivity threshold After that, generate the privacy protection gradient according to the following formula. : ,in, For differential privacy budgeting, This is the upper bound of the probability of failure of the differential privacy mechanism. The global aggregated gradient is an identity matrix with the same dimension as the gradient vector; the security aggregation module calculates the global aggregated gradient using the following formula. : ,in, For nodes Reputation weight, For nodes The overall trust score at each gradient upload time. This is the upper limit for trust scores. This represents the total number of nodes participating in this round of aggregation.

[0012] Furthermore, the infection graph-driven resilient self-healing topology layer is used when the node trust score is below the isolation threshold. At that time, taking the damaged node as the source node of the attack, the probability of attack risk propagation for each node in the infection graph at subsequent time steps is calculated iteratively using the following formula.

[0013] : ,

[0014] in, The infection graph contains pointers to nodes. The set of neighbor nodes of a directed edge. This represents the single-step propagation rate coefficient of the attack between adjacent nodes. For the node To the node The weight of the directed edge. For neighboring nodes At any moment The probability of attack risk, For nodes Vulnerability factors under the current trust score; the infection graph-driven resilient self-healing topology layer will cause the probability of attack risk propagation to exceed the warning threshold. All directed edge weights between the infected node and the damaged node in the infection graph are reset to zero, and the corresponding communication links are cut off.

[0015] Furthermore, when the overall trust score of the communication source is lower than the suspicious threshold, the resource-aware lightweight mimicry deception layer inserts a silent redirection entry into the network layer address translation rule table, directing subsequent connection requests from that communication source to the shadow node. The shadow node parses the type and request field content of the received probe packets, identifies the resource attribute category currently being probed, and adjusts its externally declared storage capacity, number of processor cores, and network bandwidth parameters to false values ​​higher than those of the real node, corresponding to the resource attribute category. It then forwards all received interaction packets and attack operation sequences to the cognitive immune evolution engine layer in real time as antigen sample input and sends back false response data consistent with the false parameter declaration to the communication source.

[0016] Furthermore, the resource-aware lightweight mimicry deception layer allows two service processes implemented in two different programming languages ​​to run simultaneously on edge nodes undertaking high-reliability services. The two service processes independently perform calculations on the same input data and submit their respective output results to the byte-level comparison module. When the byte-level comparison module detects any inconsistency in the output results of the two service processes, it sends an emergency scan request to the cognitive immune evolution engine layer and suspends external service output.

[0017] An edge computing protection method based on node protection includes the following steps: initializing a secure microkernel in the trusted execution environment of the edge node processor; writing the hash value of the secure microkernel code integrity metric into the processor platform configuration register; and generating an identity anchor value by concatenating the processor serial number and the unique identifier of the security chip and performing a hash operation. It is sealed in a sealed storage area, constructing a system call whitelist composed of system call types initiated by business processes under standard operating conditions, and processing all inbound and outbound network traffic through an encrypted proxy channel that negotiates forward confidentiality session keys based on the elliptic curve Diffie-Hellman protocol; a self-antigen map is constructed by fusing network session traffic statistics, system call frequency sequence characteristics, resource usage distribution characteristics, and service response latency percentile characteristics collected under normal business conditions of the node, and is applied to a scale of The initial antibody population completes immune tolerance training. When abnormal behavior is detected, affinity calculation, differential clonal amplification, and affinity maturation iteration are performed on the antibody population. The parameters of the generated high-affinity immune memory antibodies are then processed... After norm pruning and Gaussian noise injection, the reputation weight is calculated using the ratio of each node's comprehensive trust score to its upper limit. This weighted aggregation across nodes is then performed via a secure multi-party computation protocol to update the global threat detection model. Finally, a non-interactive zero-knowledge proof protocol based on elliptic curves is used to anchor node identities. Verification was conducted by continuously quantifying the comprehensive trust score of each node based on four dimensions: behavioral trust component, capability trust component, identity trust component, and environmental trust component. Simulated annealing algorithm was used to search for the evolutionary stable equilibrium point of the payoff matrix in a repeated game consisting of two strategies: honest cooperation and opportunistic deviation. The critical value of the trust score at the equilibrium point was then used as the basis for the evaluation. Node service interaction access criteria; when the node's overall trust score is below the isolation threshold. At that time, taking the damaged node as the attack source node, the attack risk propagation probability of each node in the infection graph is iteratively calculated, and if the risk probability exceeds the warning threshold, the attack risk propagation probability is determined. The node cuts off all communication links with the damaged node and transmits a sealed snapshot of the damaged node's business status to the backup node via an encrypted channel, completing the business takeover within 500 milliseconds.

[0018] The above method also includes the following steps: After the damaged node completes the security firmware re-flash and integrity verification, it re-verifies its identity anchor value through the zero-knowledge proof protocol and applies to join the system. In the infection graph, a new set of neighbor nodes with a difference of no less than 60% from its original set of neighbor nodes is assigned to the damaged node. The topological location of the damaged node is migrated to a new area outside its original location in the infection graph. The node's comprehensive trust score is reset to the lowest baseline value, and the trust accumulation cycle is restarted.

[0019] The beneficial effects of this invention are:

[0020] 1. This invention utilizes the cloning selection algorithm and affinity maturation iteration mechanism of the cognitive immune evolution engine layer to continuously and adaptively enhance the attack detection capability of edge nodes with each attack. Furthermore, through the differential privacy federated threat intelligence network layer, the immune evolution results of a single node are extended to the entire edge computing network in a differential privacy-preserving manner, achieving a herd immunity co-evolution effect. This fundamentally reverses the asymmetric dilemma in the field of network security, where attackers continuously accumulate experience while defenders continuously deplete it. Instead, each successful attack by an attacker on any node becomes a trigger condition for the leap in the immune capability of the entire edge network.

[0021] 2. Through repeated game-theoretic evolutionary stable equilibrium analysis of the network layer and a multidimensional trust scoring formula including a time decay term, a distributed dynamic trust management system capable of resisting long-term latent trust hijacking attacks is established without the need for a central certification authority. This eliminates the single-point failure risk and centralized authentication delay of the public key infrastructure system. Combined with the iterative calculation of the attack risk diffusion probability by the infection graph-driven elastic self-healing topology layer, preventive communication link cutoff is implemented for high-risk neighbor nodes before the damaged node is confirmed, effectively preventing single-point damage security events from being transformed into cascading damage across the entire network. After the damaged node is repaired and re-enters, topology immune reshaping renders all attack path plans built by the attacker based on the original topology structure ineffective.

[0022] 3. This invention utilizes a resource-aware lightweight mimicry deception layer with an adaptive deception mechanism for shadow nodes, a network attribute cryptographic security rotation mechanism, and a functional heterogeneous redundancy mechanism implemented in two programming languages. This simultaneously achieves proactive deception of attackers, proactive countermeasures against fingerprint scanning reconnaissance, and real-time detection of runtime tampering of critical services. The shadow nodes synchronously feed all attack interaction messages and operation sequence data back to the cognitive immune evolution engine layer as antigen samples, forming a positive feedback loop that synergistically enhances proactive deception and immune evolution. Every bit of time and technical resource consumed by the attacker in the false environment is directly converted into the accumulation of immune memory in the system's antibody population, making the deception process itself a continuous source of enhancement for the system's protective capabilities. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in this 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 only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a diagram of the overall system architecture of the present invention;

[0025] Figure 2 This is a diagram of the overall system architecture of the present invention;

[0026] Figure 3 This is a diagram of the node dynamic trust scoring mechanism of the present invention;

[0027] Figure 4 This invention relates to an infection graph risk diffusion and topological self-healing graph. Detailed Implementation

[0028] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0029] It should be noted that the use of terms such as "an embodiment," "an embodiment," "an exemplary embodiment," and "some embodiments" in the specification indicates that the described embodiment may include a specific feature, structure, or characteristic, but not every embodiment necessarily includes that specific feature, structure, or characteristic. Furthermore, when a specific feature, structure, or characteristic is described in connection with an embodiment, implementing such a feature, structure, or characteristic in conjunction with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the art.

[0030] Generally, terms can be understood at least partly from their use in context. For example, depending at least partly on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in a singular sense, or a combination of features, structures, or characteristics in a plural sense. Additionally, the term "based on" can be understood not necessarily to convey an exclusive set of factors, but rather, alternatively, depending at least partly on the context, to allow for the presence of other factors that are not necessarily explicitly described.

[0031] See Figures 1 to 4 As shown

[0032] The edge computing protection system based on node protection of this invention includes six functional layers: a trusted microkernel security base layer, a cognitive immune evolutionary engine layer, a multidimensional evolutionary game trust network layer, a resource-aware lightweight mimicry deception layer, a differential privacy federated threat intelligence network layer, and an infection graph-driven elastic self-healing topology layer. Each layer interacts and coordinates with the others through secure internal interfaces, together forming a deep protection system from hardware trust root to network topology self-healing.

[0033] I. Trusted Microkernel Security Base Layer

[0034] A secure microkernel is initialized in the trusted execution environment of the edge node processor. The integrity metric hash value of the microkernel code is written sequentially into the processor platform configuration register, establishing the starting point of an immutable trust chain from the hardware layer to the software layer. Any runtime tampering with the microkernel code can be detected and intercepted by the benchmark metric value stored in the platform configuration register.

[0035] During initial startup, the microkernel reads the node processor serial number and the security chip's unique identifier, concatenates the two, and then uses a secure hash algorithm to generate the node's unique identity anchor value. It is then sealed in a sealed storage area of ​​the trusted execution environment, so that no process running outside the trusted execution environment can access or modify the identity anchor value.

[0036] The microkernel scans all system call types initiated by each business process on the edge node under standard operating conditions, constructs a system call whitelist from the scan results, and loads it into the kernel layer interception module. All system call requests reaching the kernel layer must be compared with the whitelist one by one before execution. System call requests not in the whitelist are immediately intercepted, and the interception records are written to the security log and synchronized to the cognitive immune evolution engine layer in real time.

[0037] The microkernel continuously polls the threat level status value from the cognitive immune evolution engine layer through a dedicated interface. When the threat level rises to high, it automatically replaces the system call whitelist with a simplified version that contains only a minimal subset of system calls required by core services, while closing all unnecessary external communication interfaces. When the threat level returns to normal, it automatically restores the whitelist to the full version and reopens the communication interfaces. The entire switching process is driven by the microkernel's internal state machine and requires no external manual intervention.

[0038] Within a trusted execution environment, the microkernel establishes an encrypted proxy channel for all network traffic entering and leaving the edge nodes. During the communication initiation phase, it uses an elliptic curve Diffie-Hellman protocol to negotiate a forward confidentiality session key with the peer. All plaintext data entering the proxy channel is encrypted in counter mode using advanced encryption standards before it can be transmitted over the physical network, ensuring that even if the physical link is intercepted, the communication content cannot be recovered.

[0039] II. Cognitive Immune Evolutionary Engine Layer

[0040] When the innate immune subsystem starts up, it loads the pre-set set of known attack feature rules into the deep packet inspection engine in a multi-pattern matching data structure. The engine extracts five types of feature data in parallel for each data packet arriving at the node: source and destination address, source and destination port, protocol type, and key payload fields. It completes the comparison with the full set of rules in microseconds. Once any rule is hit, it immediately executes the interception and sends an alarm to the upper layer.

[0041] The adaptive immune subsystem initializes its size with random weights during the initialization phase. A lightweight anomaly detection model set was used as the initial antibody population, with each antibody consisting of a set of... The weight vector is composed of floating-point features. Each dimension of the weight vector corresponds to an observable feature component of node traffic or behavior. The vector as a whole represents the antibody's tendency to identify specific abnormal traffic patterns.

[0042] During normal business operation, the system continuously collects network session traffic statistics, process system call frequency sequence characteristics, processor and memory resource usage distribution characteristics, and service response latency percentile characteristics of nodes every five minutes. These four types of features are fused into a high-dimensional self-feature vector, and the self-antigen map of the node is iteratively updated. The antibody population is then trained to maintain immune tolerance through online learning, so that all antibodies in the population suppress the output response of normal features within the coverage of the self-antigen map, thereby fundamentally eliminating false alarms for normal business traffic.

[0043] When the innate immune subsystem issues an alarm or the node resource anomaly detection module triggers an alarm, the system extracts the attack feature vector from the current abnormal data stream as the antigen to be identified. For each antibody in the antibody population, the affinity is calculated according to the following formula. Rate it:

[0044]

[0045] in, Indicates antibody With antigen The affinity value between them ranges from 0 to 1. The closer the value is to 1, the stronger the antibody's ability to recognize this type of attack mode. The distance between the antibody feature weight vector and the antigen feature vector is calculated after binarization. It reflects the degree of structural difference between the two vectors in the attack feature space. The smaller the distance, the higher the degree of matching. This represents the total dimensional length of the antibody feature vector, and is used as a normalization denominator to prevent the affinity values ​​from losing comparability as the feature dimension changes.

[0046] The system sorts the antibody populations in descending order of affinity and performs differential cloning amplification on the top-ranked antibodies according to the following formula for calculating the number of clones:

[0047]

[0048] in, Indicating a ranking based on affinity The number of clones amplified by the antibody at the higher position is the greater the number of clones obtained. This represents the cloning scaling factor, used to control the overall scale and intensity of each round of cloning operations, with a value range of [value missing]. arrive ; This indicates the total size and number of the current antibody population; Indicates antibody The rank numbers are arranged in descending order of affinity, with number 1 having the highest affinity and yielding the most clones. This is a floor function that guarantees the number of clones is a non-negative integer.

[0049] The system introduces Gaussian random mutations with a mean of zero and a standard deviation inversely proportional to the antibody affinity into the weight vector of each clone copy. Antibodies with high affinity have smaller mutation amplitudes to retain accurate identification features for known attack patterns, while antibodies with low affinity have larger mutation amplitudes to explore a wider feature space. The system recalculates the affinity of all mutated copies, retains variants with higher affinity than the parent generation for inclusion in the next generation population, and eliminates variants with lower affinity than the parent generation, completing a round of affinity maturation iteration and forming high-affinity immune memory antibodies against the current attack.

[0050] The system packages the weight vector parameters of the top ten antibodies with the highest affinity in this round and submits them to the differential privacy federated threat intelligence network layer. Simultaneously, it receives high-quality global antibody parameters from the entire edge network from this layer, supplementing the local population with these external antibody parameters and eliminating an equal number of antibodies with the lowest affinity in the local population to maintain a constant total population size. This enables cross-node herd immunity co-evolution.

[0051] III. Multidimensional Evolutionary Game Trust Network Layer

[0052] Upon joining the system, each edge node generates a one-time proof commitment within the trusted execution environment, proving to its neighboring nodes that it holds a legitimate identity anchor value through an elliptic curve-based non-interactive zero-knowledge proof protocol. The proof process does not reveal any algebraic structure information of the identity anchor value to the outside world. After the neighboring nodes pass the local verification, they are given an initial trust component, thus completing the initial identity trust authentication without the participation of a central institution.

[0053] The system continuously quantifies and evaluates the four trust dimensions of each node at a fixed collection period, including behavioral trust components. By the node in the past The capability trust component is calculated by taking the weighted average of the agreement commitment fulfillment rate and service agreement compliance records within each collection period. The identity trust component is calculated by multiplying the normalized reciprocal of the median service response latency by the accuracy of the data processing result. The environmental trust component is calculated from the ratio of successful zero-knowledge proof verifications to the total number of verifications within this data collection period. It is calculated by multiplying the reciprocal of the variance of the reported coordinates of the node's physical location by the normalized reciprocal of the number of network connection interruptions.

[0054] The system calculates the time for each node based on the following dynamic trust scoring formula. Trust rating:

[0055]

[0056] in, Represents a node At any moment The comprehensive trust score, ranging from 0 to 1, is the core basis for the system to judge whether a node is trustworthy. , , These represent the behavioral trust component, capability trust component, identity trust component, and environment trust component obtained during the current collection cycle settlement, respectively. The values ​​of each component are normalized to between 0 and 1. , , The weighting coefficients for each dimension component must satisfy the following conditions: The system administrator configures the security priority based on the business scenario. For time decay term, The decay rate constant is At the start of the current evaluation period, the decay term causes the historically accumulated trust score to continuously decrease over time, thus preventing attackers from using high trust accumulation to launch sudden attacks after long-term lurking. For the node at time The occurrence of the first The change in rating caused by a trust event is represented by a positive value for a positive event and a negative value for a suspicious event. The absolute value is determined by the severity level of the event. For an indicator function, if and only if The value is set to 1 to ensure that only the impact of historical events that occurred before the current moment is accumulated in the current score; Deadline The total number of trust events recorded by this node.

[0057] The system models the multi-round interactions between nodes as a finitely repeated game, using honest cooperation and opportunistic deviation as the pure strategies available to both sides. After constructing the complete payoff matrix, the simulated annealing algorithm iteratively searches for the system's evolutionary stable equilibrium point in the global policy space. The trust score threshold corresponding to the equilibrium point is set as the service interaction admission threshold between nodes. Trust score is lower than All service requests from a node will be automatically rejected until its trust score recovers to above the threshold.

[0058] IV. Resource-Aware Lightweight Mimicry Deception Layer

[0059] The system utilizes container isolation technology on each edge node, deploying two shadow node processes for each node in an independent network namespace. The shadow nodes share the host machine's physical network card but use independent virtual IP addresses. They also configure their service port set, service version information, and system hardware resource declarations with false parameters that are higher than the privilege level of the real nodes, so that attackers performing reconnaissance scans will identify the shadow nodes as targets with greater attack value than the real nodes.

[0060] The system uses a cryptographically secure pseudo-random sequence generator with a hash chain structure to generate a new set of network attribute parameters for real nodes every 300 seconds. This set covers the service listening port number set, transport layer protocol version identifier, and transport layer secure handshake extension parameters. During the 30-second buffer period before the new parameter set takes effect, the change notification is synchronized to all communication peers holding legitimate sessions through the encrypted channel of the microkernel security base layer, ensuring that legitimate business communication is not interrupted during the attribute rotation period.

[0061] When the multidimensional evolutionary game trust network layer marks the trust score of a certain communication source as below the suspicious threshold, the system inserts a silent redirection entry into the address translation rule table of the network layer, redirecting all subsequent connection requests from that source to the corresponding shadow node. The shadow node parses the probe message type and request field content sent by the source in real time, identifies the resource attribute category that the attacker is probing, and dynamically adjusts its externally declared storage capacity, number of processor cores, and network bandwidth parameters, so that the shadow node always presents itself as a high-value target in the dimension currently being probed by the attacker.

[0062] The shadow node forwards all received interaction messages, attack command sequences, and operation behavior records to the cognitive immune evolution engine layer in real time as new antigen sample input. At the same time, it sends back false response data that matches the false resource declaration to the attack source, inducing the attacker to continuously consume time and technical resources in the shadow node environment, thus extending the effective time window for the system to collect the attacker's tactical and technical characteristics.

[0063] For edge nodes undertaking high-reliability services, the system uses two service processes implemented in different programming languages ​​to run the same business logic simultaneously. The two processes independently calculate the same input data and submit the output results to the byte-level comparison module. When the comparison results are consistent, they are output to the outside world in a unified manner. Once the comparison finds any byte inconsistency, it is determined that at least one process has been tampered with. The system immediately suspends the output of external services and sends an emergency scan request to the cognitive immune evolution engine layer, triggering a deep behavior detection process for the two processes.

[0064] V. Differential Privacy Federal Threat Intelligence Network Layer

[0065] At the start of each federated learning iteration, each edge node uses the attack sample data collected from the cognitive immune evolutionary engine layer in this iteration to perform several steps of stochastic gradient descent training on the lightweight deep neural network threat detection model locally, and calculates the parameter update gradient vector generated by this round of local training. and will of Norm clipping to a preset sensitivity threshold Within.

[0066] The system injects noise into the pruned gradient vector according to the following differential privacy Gaussian mechanism formula to obtain a privacy-preserving gradient that can be securely uploaded. :

[0067]

[0068] in, Represents a node The privacy-preserving gradient vector obtained after differential privacy processing can be used by any... Attackers cannot exceed and Additional information from the defined upper bound of privacy disclosure is used to infer the content of local training data; Represents a node The original gradient vector calculated during this round of local training; This represents a multidimensional Gaussian distribution, where the first parameter in parentheses is the zero-mean vector and the second parameter is the covariance matrix. Represents the gradient function Sensitivity, i.e., the gradient vector when a single sample in the local training dataset is replaced. The maximum possible change in the norm; This represents the differential privacy budget; the smaller the value, the stronger the privacy protection. This represents the upper bound of the failure probability of the differential privacy mechanism, typically taken as... Magnitude; This represents the identity matrix with the same dimension as the gradient vector, ensuring that the injected noise is independent and has the same variance across all gradient dimensions.

[0069] The node will be encrypted Current trust score of this node The data is then uploaded to the aggregation node via a secure channel. The aggregation node performs the following reputation-weighted aggregation calculation within the encrypted domain using a secure multi-party computation protocol:

[0070]

[0071] in, This represents the global gradient vector generated in this round of federated learning; the aggregation node uses... After updating the global threat detection model parameters, the new model is distributed to all participating nodes in an encrypted manner; Represents a node The reputation weight of this round of gradient upload is determined by the ratio of the node's current trust score to the system's maximum trust score. The higher the trust score of a node, the greater its gradient weight. The compromised node's malicious gradient weight will approach zero due to the sharp drop in its trust score, thus achieving robust defense against gradient poisoning attacks without the need for central auditing. Represents a node At the gradient upload time Comprehensive trust score; This indicates the maximum upper limit of the system's trust score; This indicates the total number of nodes participating in this round of aggregation.

[0072] VI. Infection Graph-Driven Elastic Self-Healing Topology Layer

[0073] The system abstracts all nodes in the edge computing network as vertices of a directed weighted graph, and the physical or logical network connections between nodes as directed weighted edges, with edge weights... The edge weight matrix of the infection graph is determined by the weighted average of three indicators: normalized bandwidth, latency stability, and historical connectivity success rate of the corresponding path, and is continuously updated based on real-time changes in network topology.

[0074] When a certain node Trust score below isolation threshold At that time, the system takes this node as the attack source and its current risk probability as the initial value, and iteratively calculates the risk probability of each target node in the infection graph being laterally infected at subsequent times according to the following attack risk diffusion formula:

[0075]

[0076] in, Represents a node At any moment The probability of being attacked and infected laterally, ranging from 0 to 1. The larger the value, the more urgent the threat of attack spread to the node. This indicates that there are pointers to nodes in the infection graph. The set of all neighbor nodes of a directed edge, i.e., the set of nodes that an attack might occur via a directed connection. The set of all source nodes that propagate; This represents the single-step propagation rate coefficient of an attack between adjacent nodes, configured by the system administrator based on statistical analysis results of historical lateral movement events; Indicates from node To the node The weight of the directed edge indicates that the larger the value, the closer the network connection between the two nodes, and the higher the probability that the attack will spread through this path. Representing neighboring nodes At any moment The attack risk probability represents the likelihood that the neighboring node has been infected and is spreading the attack outwards. Represents a node The vulnerability factor at the current trust score level. The lower the trust score, the larger the factor, indicating that the node is more likely to have been compromised and become a springboard for the spread of attacks. This represents the upper limit of the system's trust score.

[0077] The system will detect when the probability of risk spread exceeds the warning threshold. The set of nodes is marked as a candidate for preventive isolation. For each node in the candidate set, the weights of all directed edges between it and the confirmed damaged nodes are immediately reset to zero in the edge weight matrix of the infection graph, and the corresponding communication links are forcibly cut off at the network layer. This preventive blocking is completed before the attacker initiates a lateral movement to reach the candidate node.

[0078] The system will take a sealed snapshot of the business status generated at the end of the most recent collection cycle before the damaged node is isolated, and transmit it to the prepared replacement node that has completed the readiness test through the encrypted channel of the microkernel security base layer. The replacement node will deseal the sealed snapshot in its trusted execution environment and restore it to the memory running state, and take over all business output of the damaged node within the target switching time limit of 500 milliseconds.

[0079] After the damaged node completes the security firmware re-flash and integrity verification, it reapplies to join the system. The system assigns it a new set of neighbors in the infection graph with a difference of no less than 60% from its original set of neighbor nodes, and migrates its topological location to a new area outside its original location. This renders any attack path planning based on the original topological knowledge invalid. At the same time, the node's trust score is reset to the minimum baseline value and it is required to go through a full trust accumulation cycle again to complete the topological immune reshaping and restore the overall protection capability of the system.

[0080] Example 1

[0081] Supply chain attack protection in industrial IoT scenarios

[0082] A smart factory has deployed an industrial IoT edge computing protection network consisting of twenty edge nodes. Each node is connected to the core aggregation switch via an industrial Ethernet network in a star topology. Each node is equipped with a quad-core processor, eight gigabytes of memory, and a dedicated security chip that supports a trusted execution environment. The types of business undertaken by the nodes cover three core tasks: real-time aggregation of sensor data, issuance of control commands for processing equipment, and visual quality inspection of the production line.

[0083] During the system deployment phase, the secure microkernel is initialized sequentially in the trusted execution environment of twenty nodes. Each node reads its secure chip serial number and processor unique identifier, concatenates them, and then uses a hash operation to generate its own unique identity anchor value. After the identity anchor value is sealed in sealed storage, each node broadcasts its identity declaration to its neighboring nodes through a zero-knowledge proof protocol and completes two-way verification, thereby obtaining an initial trust score for each node. The infection graph is initialized with a star topology, with the core aggregation switch node as the central vertex and each edge node as a leaf vertex. The edge weights are initially assigned based on the industrial Ethernet link test data.

[0084] During the first 48 hours of normal business operation, the system performs immune tolerance training. The cognitive immune evolution engine layer collects traffic statistics feature vectors, system call frequency vectors, resource occupancy distribution vectors, and service latency percentile vectors for each node every five minutes under normal business conditions. After aggregating the data from 576 collection cycles, a self-antigen map of each node is constructed. After completing immune tolerance training on a population with an initial scale of 500 antibodies, the false alarm rate of each node for normal industrial protocol messages is reduced to no more than twice per thousand hours.

[0085] In the 72nd hour, the attacker bypassed the factory boundary firewall by using a forged sensor firmware update packet, and implanted a backdoor process into the operating system kernel of the edge node numbered five. This backdoor process periodically sent small-volume reconnaissance packets to the attacker's control server at 30-second intervals, with each packet containing 82 bytes and destined for the non-standard port number 4321.

[0086] The microkernel interception module of node 5 generates an interception alarm and writes it to the security log when the backdoor process initiates its first socket system call to an unauthorized external address. The innate immune subsystem performs feature matching on the first reconnaissance packet but fails to match because the feature is not in the rule base. After the second reconnaissance packet arrives, the adaptive immune subsystem extracts periodic abnormal outgoing packets as feature vectors from the time interval, destination address clustering features, and payload length distribution of the two consecutive reconnaissance packets, using them as antigens. Vector dimension It has 128 dimensions.

[0087] The system calculates the affinity for each of the five hundred antibodies individually. Feature weight vector and antigen The Hamming distance is The affinity is calculated using the formula as follows: The attack confirmation affinity threshold set by the system is higher than the threshold. The adaptive immune subsystem confirms that the current behavior is malicious outbound communication and immediately triggers the clone selection process.

[0088] The system uses , to perform clonal amplification on the antibody ranked first in affinity , with the calculated clonal number being . It performs Gaussian mutation with a standard deviation of 0.02 on 300 copies, and after affinity maturation, retains the top ten variants with the highest affinity as immune memory antibodies targeting this type of backdoor communication behavior, which are incorporated into the persistent antibody library of Node 5.

[0089] The cognitive immune evolution engine layer reports a suspicious outgoing communication event occurring at Node 5 to the multi-dimensional evolutionary game trust network layer, and the trust network layer records this event as a negative trust event. After substituting into the trust scoring formula, the comprehensive trust score of Node 5 drops sharply from 0.79 to 0.23, which is lower than the isolation threshold set by the system , immediately triggering the node isolation and subsequent linkage protection process.

[0090] The infection graph-driven elastic self-healing topology layer takes Node 5 as the attack source node, sets its initial risk probability to perform the first round of risk propagation iterative calculation on the infection graph. The calculation results show that the risk probability of Node 3 directly connected to Node 5 is 0.73, and the risk probability of Node 9 is 0.67, both exceeding the warning threshold . The system immediately resets the weights of the directed edges connecting Node 3 and Node 9 to Node 5 in the infection graph to zero, and cuts off the corresponding links at the network layer to complete preventive isolation.

[0091] The system transmits the complete state of the sensor data aggregation service saved in the latest sealed snapshot of Node 5 to the standby Node 12 through an encrypted channel. After Node 12 completes state unsealing in a trusted execution environment, it takes over the data reporting links of 23 sensors on the production line and the device control instruction issuing links within 380 milliseconds, without any interruption to production operations.

[0092] The resource-aware lightweight mimicry deception layer synchronously and silently redirects all subsequent communications between the attacker's control server and Node 5 to the shadow node running on Node 5. The shadow node parses the lateral scanning instructions issued by the attacker afterwards, identifies that the attacker is probing targets with database read permissions, then adjusts its declared storage capacity to four times that of the real node and opens the fake database service port 3306, luring the attacker to continuously carry out database injection attacks on the shadow node for 2 hours and 17 minutes. During this period, the shadow node completely records the operation feature vectors of the six attack tools used by the attacker, and synchronizes all of them to the cognitive immune evolution engine layer.

[0093] The differential privacy federal threat intelligence network layer generates high-affinity immune memory antibodies from the feature vectors of six attack tools, respectively. The differential privacy budget injects Gaussian noise into the antibody parameters. After completing one round of federated learning iteration, the global threat detection model parameters are updated and distributed to all twenty nodes. After the antibody population of the remaining nineteen nodes is updated, the detection speed of this series of supply chain attack tools is improved from being undetectable to triggering affinity confirmation on the first reconnaissance packet. The herd immunity capability of the entire factory edge network has significantly improved after this attack event.

[0094] Example 2

[0095] Lateral penetration protection against advanced persistent threats in smart city scenarios

[0096] A provincial-level intelligent traffic management platform has deployed an urban road traffic edge computing protection network consisting of 36 edge nodes. Each node is interconnected in a mesh topology through a metropolitan wireless private network and then connected to the cloud-based traffic command center via a dedicated line. Each node is deployed at a major intersection in the city and undertakes three core businesses: real-time optimization and control of traffic lights, vehicle target detection and recognition, and traffic flow statistics and reporting. The infection graph is initialized with a mesh topology, the average connectivity between nodes is 4.7, and the edge weights are assigned based on the measured link quality data of the metropolitan wireless private network.

[0097] After deployment, the system entered normal operation. After 30 days of continuous normal service, the trust scores of each node stabilized between 0.85 and 0.92 through trust accumulation. The antibody population continued to evolve through daily federated learning iterations. Evolutionary stability equilibrium analysis of the multidimensional evolutionary game trust network layer determined the admission threshold. It converged stably to 0.52.

[0098] On the thirty-first day, the attackers launched a man-in-the-middle attack on the authentication protocol of the metropolitan wireless private network, successfully intercepting a segment of encrypted communication data between node number 17 and its neighboring node number 21. They then used a known vulnerability in the network device firmware to inject a piece of resident code into the user-space process space of node number 21. This resident code passively waited for the attacker's command channel to be activated at random intervals between 200 and 600 seconds, without actively initiating any outgoing communication, deliberately avoiding periodic anomaly detection.

[0099] The resident code remained dormant for the first six hours after injection, without triggering any active communication. Neither the innate nor adaptive immune subsystems of node 21 detected any anomalies. However, the trust component of the multidimensional evolutionary game trust network layer... The data collected showed that the 95th percentile of the service response latency of node 21 after the resident code injection increased from the normal 23 milliseconds to 41 milliseconds, causing the capability trust component to drop from 0.88 to 0.71 and the overall trust score to gradually decrease from 0.91 to 0.83. Although the isolation threshold was not reached, the trust decline warning flag was triggered.

[0100] On the 31st day, in the 7th hour, the attacker issued a lateral movement command to the resident code via the command channel. The resident code then initiated a scan of the local process list and memory data within node 21. This action triggered multiple unauthorized memory access system call alerts from the microkernel interception module. The cognitive immune evolutionary engine layer extracted the current process behavior feature vector as an antigen. The antibody with the highest affinity in the antibody population affinity calculation had an affinity value of 0.81, exceeding the confirmation threshold. Therefore, the current behavior was determined to be malicious memory reconnaissance, and the incident was immediately reported to the trust network layer. The value is -0.48. After the existing capability trust decreases, the trust score of node 21 drops to 0.35, which is lower than the isolation threshold of 0.40, triggering the complete isolation linkage process.

[0101] The infection graph-driven resilient self-healing topology layer uses node 21 as the attack source. The risk diffusion iterative calculation was performed. In the first round of iterations, the risk probability of node 17, which is directly connected to node 21 through a high-weight edge, was 0.81; the risk probability of node 23 was 0.74; and the risk probability of node 28 was 0.62. All three nodes exceeded the warning threshold of 0.60. The system set all the weights of the directed edges between the three candidate nodes and node 21 to zero in the infection graph and cut off the corresponding wireless communication channels at the network layer, thus completing the preventive blocking before the attacker could move laterally to any adjacent node.

[0102] The system encrypted and transmitted a real-time snapshot of the traffic signal control business status of node 21 to the backup node 34. After node 34 completed the status desealing in the trusted execution environment, it took over the traffic signal control of the intersection under the jurisdiction of node 21 within 420 milliseconds. The traffic signal control of the intersection remained in continuous operation during the handover process, and no abnormal traffic signal status occurred.

[0103] The resource-aware lightweight mimicry deception layer silently redirects all subsequent commands issued by the attacker's command channel to the shadow node of node 21. The shadow node parses the attacker's subsequent lateral movement commands, identifies the attacker's intention to exploit the inter-node trust relationships of the traffic control system to infiltrate the core convergence node, and then adjusts its declared node role to a super control node with cross-regional signal coordination authority. It then returns a false neighboring node topology map to the attacker. This false topology map replaces all paths pointing to the real core node with paths pointing to other shadow nodes, inducing the attacker to continuously perform infiltration operations in the false topology for three hours and forty minutes.

[0104] During the continuous interaction between the attacker and the shadow node, the shadow node fully recorded the attacker's operation sequences of five types of customized attack tools targeting the traffic control protocol. The cognitive immune evolution engine layer transformed these into five high-affinity immune memory antibodies. Taking the fifth node, which had the highest trust score, as an example, the differential privacy federated threat intelligence network layer determined its reputation weight. In this federated aggregation, the reputation weight of node number 21, whose trust score has already decreased due to its resident code, is also affected. This ensures that any malicious gradient contribution from node 21 will affect the aggregation result. The impact of the middle is insufficient After the global model parameters were updated, they were securely distributed to all 35 normal nodes. The first packet detection capability of the 35 nodes for this series of customized traffic control protocol attack tools was updated. The herd immunity level of the entire city's traffic edge protection network achieved an overall leap after this advanced persistent threat event ended.

[0105] After completing firmware re-flashing, integrity verification, and zero-knowledge identity re-authentication, node 21 applied to rejoin the system. The infection graph-driven elastic self-healing topology layer assigned it seven new neighbor nodes, with only one overlap with the original five neighbor nodes. The new topology location is in the edge region of the infection graph rather than the original network center region. Node 21's trust score was reset to the baseline value of 0.40, and it re-entered the trust accumulation cycle. After the topology immune reshaping was completed, the overall protection topology of the system was restored to a fully connected mesh structure of thirty-six nodes. Any subsequent lateral penetration attempts launched by an attacker who has knowledge of the original topology against node 21 could not proceed because all paths were invalid.

[0106] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0107] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An edge computing protection system based on node protection, characterized in that, The system comprises a trusted microkernel security base layer with interconnected communication, a cognitive immune evolutionary engine layer, a multidimensional evolutionary game-theoretic trust network layer, a resource-aware lightweight mimicry deception layer, a differential privacy federated threat intelligence network layer, and an infection graph-driven resilient self-healing topology layer. The trusted microkernel security base layer includes a secure microkernel running in the trusted execution environment of an edge node processor. The secure microkernel's code integrity metric hash value is written to the processor platform configuration register. The secure microkernel generates an identity anchor value by concatenating the edge node processor serial number and the unique identifier of the security chip and performing a hash operation. The system call requests are sealed in a trusted execution environment’s sealed storage area. The secure microkernel has a system call whitelist consisting of system call types initiated by business processes in standard operating conditions. System call requests not in the system call whitelist are intercepted by the secure microkernel. The cognitive immune evolutionary engine layer includes an innate immune subsystem that stores a set of known attack feature rules in a multi-pattern matching data structure, and a subsystem that maintains a set of rules for attack features. An adaptive immune subsystem consisting of an antibody population, each antibody being composed of... The antibody population is constructed using a floating-point feature weight vector. It undergoes immune tolerance training based on a self-antigen map fused from network session traffic statistics, system call frequency sequence features, resource usage distribution features, and service response latency percentile features collected during normal node operation. The multidimensional evolutionary game-theoretic trust network layer includes node identity anchor values. The verification module for the elliptic curve-based non-interactive zero-knowledge proof protocol, and the behavioral trust component. Capability and Trust Weight Identity and Trust Weight and environmental trust weight The trust scoring module quantifies the comprehensive trust score of each node across four dimensions; the resource-aware lightweight mimicry deception layer includes shadow nodes deployed as container processes running in independent network namespaces on each edge node, and a network attribute rotation module that updates the service listening port set, transport layer protocol version identifier, and transport layer secure handshake parameters of the real node at fixed time intervals using a hash chain-structured cryptographically secure pseudo-random sequence generator. The service port set, service version information, and hardware resource declaration parameters of the shadow nodes are configured with fake parameters that are higher than the permission level of the real nodes; the differential privacy federated threat intelligence network layer includes the execution of locally trained gradient vectors on nodes. A privacy protection module that clips norms and injects Gaussian noise, and a secure aggregation module that performs weighted aggregation of the processed gradient vector in the ciphertext domain using the ratio of each node's trust score to the upper limit of the trust score as the reputation weight. The infection graph-driven elastic self-healing topology layer includes an infection graph constructed with each edge node as a vertex and the network connections between nodes as directed weighted edges, and a fast migration module that transmits a sealed snapshot of the service status of the damaged node to the alternative node via an encrypted channel and completes the service takeover within 500 milliseconds. The edge weight of the directed weighted edge is determined by the weighted average of three indicators: normalized bandwidth, latency stability, and historical connectivity success rate of the corresponding network path.

2. The edge computing protection system based on node protection according to claim 1, characterized in that, The trust scoring module calculates nodes according to the following formula. At any moment Overall trust score ： , in, , , The weight coefficients of the trust components in each dimension are given, and satisfy the following conditions: , The decay rate coefficient, This is the start time of the current evaluation period. For the node at time The occurrence of the first The change in rating for each trust event is represented by a positive value for a positive event and a negative value for a suspicious event. For when The time value is Indicator functions, Deadline The total number of trust events; the multidimensional evolutionary game trust network layer searches for an evolutionarily stable equilibrium point in the repeated game payoff matrix composed of two strategies, honest cooperation and opportunistic deviation, using a simulated annealing algorithm, and sets the trust score threshold corresponding to the equilibrium point as the node service interaction admission threshold. The overall trust score is lower than The node's service request was rejected.

3. The edge computing protection system based on node protection according to claim 1, characterized in that, The adaptive immune subsystem extracts attack feature vectors as antigens when it detects abnormal behavior. Calculate the value of each antibody in the antibody population using the following formula. With antigen Affinity between ： , in, The Hamming distance between the binarized antibody feature weight vector and the antigen feature vector is given. The length of the antibody feature vector; after the adaptive immune subsystem sorts the antibody population in descending order of affinity, the affinity ranking is calculated using the following formula. Number of antibody clones amplified at the position : ,in, For the range of values ​​within to Cloning scaling factor between Antibody The affinity is ranked in descending order; the standard deviation of the mutation amplitude of the clone is inversely proportional to the affinity of the parent antibody. Mutants with higher affinity than the parent are retained in the next generation population, while mutants with lower affinity than the parent are eliminated.

4. A protective edge computing system, characterized in that, The privacy protection module will store the original gradient vector generated by the local training of node n. of Norm clipping to sensitivity threshold After that, generate the privacy protection gradient according to the following formula. : ,in, For differential privacy budgeting, This is the upper bound of the probability of failure of the differential privacy mechanism. The identity matrix has the same dimension as the gradient vector; the security aggregation module calculates the global aggregated gradient using the following formula. : ,in, For nodes Reputation weight, For nodes The overall trust score at each gradient upload time point. This is the upper limit for trust scores. This represents the total number of nodes participating in this round of aggregation.

5. The edge computing protection system based on node protection according to claim 1, characterized in that, The infection graph-driven resilient self-healing topology layer operates when the node trust score is below the isolation threshold. At that time, taking the damaged node as the source node of the attack, the probability of attack risk propagation for each node in the infection graph at subsequent time steps is calculated iteratively using the following formula. ： ， in, The infection graph contains pointers to nodes. The set of neighbor nodes of a directed edge. This represents the single-step propagation rate coefficient of the attack between adjacent nodes. For the node To the node The weight of the directed edge. Neighboring nodes At any moment The probability of attack risk, For nodes Vulnerability factors under the current trust score; the infection graph-driven resilient self-healing topology layer will cause the attack risk to spread beyond the warning threshold. All directed edge weights between the infected node and the damaged node in the infection graph are reset to zero, and the corresponding communication links are cut off.

6. The edge computing protection system based on node protection according to claim 1, characterized in that, When the overall trust score of the communication source is lower than the suspicious threshold, the resource-aware lightweight mimicry deception layer inserts a silent redirection entry into the network layer address translation rule table, directing subsequent connection requests from the communication source to the shadow node. The shadow node parses the type and request field content of the received probe packets, identifies the current probed resource attribute category, adjusts its externally declared storage capacity, number of processor cores, and network bandwidth parameters to false values ​​higher than those of the real node corresponding to the resource attribute category, forwards all received interaction packets and attack operation sequences to the cognitive immune evolution engine layer in real time as antigen sample input, and sends back false response data consistent with the false parameter declaration to the communication source.

7. The edge computing protection system based on node protection according to claim 1, characterized in that, The resource-aware lightweight mimicry deception layer runs two service processes implemented in two different programming languages ​​simultaneously on the edge nodes undertaking high-reliability services. The two service processes independently perform calculations on the same input data and submit their respective output results to the byte-level comparison module. When the byte-level comparison module detects any byte inconsistency in the output results of the two service processes, it sends an emergency scan request to the cognitive immune evolution engine layer and suspends external service output.

8. A design method for an edge computing protection system based on node protection according to any one of claims 1-7, characterized in that, Includes the following steps: In the trusted execution environment of the edge node processor, a secure microkernel is initialized. The hash value of the secure microkernel code integrity metric is written to the processor platform configuration register. The processor serial number and the unique identifier of the security chip are concatenated and hashed to generate an identity anchor value. It is sealed in a sealed storage area, and a system call whitelist is constructed consisting of system call types initiated by business processes under standard operating conditions. All inbound and outbound network traffic is processed through an encrypted proxy channel that negotiates forward confidential session keys based on the elliptic curve Diffie-Hellman protocol. A self-antigen map is constructed by fusing network session traffic statistics, system call frequency sequence characteristics, resource usage distribution characteristics, and service response latency percentile characteristics collected under normal node operation conditions. This map is then applied to nodes with a scale of [missing information - likely related to network session traffic statistics, system call frequency sequence characteristics, resource usage distribution characteristics, and service response latency percentile characteristics]. The initial antibody population completes immune tolerance training. When abnormal behavior is detected, affinity calculation, differential clonal amplification, and affinity maturation iteration are performed on the antibody population. The parameters of the generated high-affinity immune memory antibodies are then processed... After norm pruning and Gaussian noise injection, the reputation weight is calculated using the ratio of each node's comprehensive trust score to its upper limit. This weighted aggregation across nodes is then performed via a secure multi-party computation protocol to update the global threat detection model. Finally, a non-interactive zero-knowledge proof protocol based on elliptic curves is used to anchor node identities. Verification was conducted by continuously quantifying the comprehensive trust score of each node based on four dimensions: behavioral trust component, capability trust component, identity trust component, and environmental trust component. Simulated annealing algorithm was used to search for the evolutionary stable equilibrium point of the payoff matrix in a repeated game consisting of two strategies: honest cooperation and opportunistic deviation. The critical value of the trust score at the equilibrium point was then used as the basis for the evaluation. Node service interaction access criteria; when the node's overall trust score is below the isolation threshold. At that time, taking the damaged node as the attack source node, the attack risk propagation probability of each node in the infection graph is iteratively calculated, and if the risk probability exceeds the warning threshold, the attack risk propagation probability is determined. The node cuts off all communication links with the damaged node and transmits a sealed snapshot of the damaged node's business status to the backup node via an encrypted channel, completing the business takeover within 500 milliseconds.

9. The edge computing protection method based on node protection according to claim 8, characterized in that, It also includes the following steps: After the damaged node completes the security firmware re-flash and integrity verification, it re-verifies its identity anchor value through the zero-knowledge proof protocol and applies to join the system. In the infection graph, the damaged node is assigned a new set of neighbor nodes with a difference of no less than 60% from its original set of neighbor nodes. The topological location of the damaged node is migrated to a new area outside its original location in the infection graph, and the node's comprehensive trust score is reset to the lowest baseline value, restarting the trust accumulation cycle.

Images