A cloud platform AI intelligent agent data entry and exit security control method

Abstract

This application discloses a cloud platform AI intelligent agent data entry and exit security control method, including: setting up an E2B sandbox service cluster in a domestic data center; fragmenting the data to be entered and exited within the E2B sandbox to generate a fragmentation mapping table; dividing the fragmented data into two groups; performing progressive encryption on the two groups of data to generate two sets of ciphertext fragment sets; deploying a chaos platform based on the ciphertext fragment sets; and verifying the encrypted data through simulated attacks and fault injection by the chaos platform; adjusting the parameters of the progressive encryption based on the verification results of the chaos platform to achieve security control of entry and exit data. Through continuous verification and optimization by the chaos engineering platform, it has stronger continuous security effectiveness. The progressive encryption and obfuscation process forms a deep defense chain through multi-layer encryption and obfuscation, significantly improving data security and ensuring the security, integrity, and availability of data during cross-border flow.

Description

Technical Field

[0001] This invention relates to the field of immigration data security technology, and in particular to a cloud platform AI agent data immigration security control method. Background Technology

[0002] In the era of globalization and information technology, cross-border data flows have become a vital force driving economic development and promoting international cooperation. With the rapid development of technologies such as cloud computing, big data, and artificial intelligence, cloud platform AI agents are increasingly widely used across various fields, and the amount of data they process is exploding. This data not only contains personal privacy information but also involves sensitive content such as corporate trade secrets and national security. Therefore, ensuring the security, integrity, and availability of data during cross-border flows has become a crucial issue that urgently needs to be addressed.

[0003] Traditionally, the secure control of cross-border data flows has relied primarily on encryption technologies and access control policies. Encryption technologies protect data confidentiality by encrypting the data, preventing unauthorized users from accessing its plaintext content. However, with the continuous improvement of computing power and the deepening of cryptographic research, traditional encryption algorithms face the risk of being cracked. The development of quantum computing technology, in particular, poses a serious threat to existing encryption systems. Access control policies, on the other hand, prevent unauthorized users from accessing sensitive data by restricting user access permissions. However, access control policies are often insufficient to handle malicious actions by internal personnel or infiltration attacks by external attackers.

[0004] Against this backdrop, the security control of cross-border data flow for cloud platform AI agents faces unprecedented challenges. On the one hand, cloud platform AI agents process massive amounts of data of diverse types, including structured, semi-structured, and unstructured data, posing significant difficulties for secure data processing. On the other hand, cloud platform AI agents are typically deployed in distributed environments, with data flowing and shared across multiple nodes, increasing the risk of data theft or tampering. Furthermore, cross-border data flows involve differences in laws, regulations, and regulatory requirements across different countries and regions, further complicating data security control. Summary of the Invention

[0005] This application provides a cloud platform AI agent data entry and exit security control method. Through continuous verification and optimization by a chaos engineering platform, the security system can dynamically adapt to the evolution of unknown attack methods, and has stronger continuous security effectiveness. The progressive encryption and obfuscation process forms a deep defense chain through multi-layer encryption and obfuscation, which significantly improves data security and builds a dynamic, elastic, and measurable cloud platform AI agent data entry and exit security control system to ensure the security, integrity and availability of data in the process of cross-border flow.

[0006] This application provides a method for controlling the entry and exit security of AI intelligent agent data on a cloud platform, including: S101, set up an E2B sandbox service cluster in the domestic data center, fragment the data to be imported and exported in the E2B sandbox, generate a fragment mapping table, divide the fragmented data into two groups, and perform progressive encryption on the two groups of data to generate two sets of ciphertext fragment sets. The progressive encryption includes translation obfuscation encryption and inversion obfuscation encryption. S102, deploys a chaos platform based on a set of ciphertext shards. The chaos platform verifies encrypted data by simulating attacks and injecting faults. S103 adjusts the parameters of progressive encryption based on the verification results of the chaos platform, thereby achieving secure control over entry and exit data.

[0007] Preferably, the method for progressive encryption of the two sets of data is as follows: For the first set of data, a shift-and-obfuscation encryption is performed. For this, a byte-level cyclic shift operation is performed on each byte in the first set of data fragments. The byte-level cyclic shift refers to cyclically shifting the 8 bits of each byte in the fragment, treating the byte as a loop with its ends connected, rotating it left or right by a certain number of bits, and filling in the bits shifted out from one end from the other. The step size and direction of the shift are used as dynamic parameters, which are dynamically configured and managed through the CMChaos platform. For the second set of data, an inversion-and-obfuscation encryption is performed, completely reversing the bit order of each byte. For the fragmented data after the bit order inversion operation, a byte-level cyclic shift operation is performed again, this time using a different step size and direction than the first shift operation.

[0008] Preferably, the safety control method further includes: S201: Based on the fragmented data, a chain-like initial seed is generated using a random algorithm. Multiple encryption channels are set for each fragment of data, and these multiple encryption channels constitute a parallel encryption network. S202: Based on the initial chain seed and the first data fragment, the hash function is used to calculate the current chain state value, the first data fragment is input into the parallel encryption network, and the ciphertext results of the parallel encryption network are synthesized to generate the first obfuscated fragment; S203, calculate the second chain state seed based on the current chain state value and the first obfuscated fragment, and execute step S202 to generate the second obfuscated fragment; process the remaining data fragments in sequence according to steps S201 and S202 to obtain multiple obfuscated fragments; S204: Construct an obfuscation chain based on multiple obfuscated shards using association rules, input the obfuscation chain into the chaos platform, execute step S103 to obtain the adjusted obfuscation chain, and transmit the adjusted obfuscation chain to the overseas receiving end through a secure network channel.

[0009] Preferably, the safety control method further includes: S301: Deploy an E2B sandbox service cluster at the IoT gateway to build an edge decoy gateway. Generate a fake data stream based on the edge decoy gateway and embed the monitoring module into the IoT protocol gateway. Use the monitoring module to monitor IoT attack characteristics and match the policies in the resource consumption policy library with policy priority according to the attack characteristics. S302, constructs an edge attack chain by embedding decryption tasks in a forged data stream; if the attacker parses the forged data, the edge attack chain is triggered, and the attacker's dwell time is calculated based on the edge attack chain; S303 uploads the matched resource consumption strategy and the attacker's dwell time to the Chaos Platform. The Chaos Platform aggregates data from multiple edge nodes to generate an attacker topology map, and generates security control strategies based on the attacker topology map.

[0010] Preferably, the safety control method further includes: S401, Generate real data groups and virtual data groups based on the fragmented data, construct an attack chain based on the virtual data groups, and determine the receiver's environment based on the attack chain. If the environment is secure, execute step S402; otherwise, if the environment is insecure, trigger sponge defense for adjustment. S402 transmits the real data group and the virtual data group in a dual-path manner. The receiver performs cross-validation on the received data. If the verification results are consistent, an alarm is triggered. If the verification results are inconsistent, the virtual data group is decrypted through the attack chain. The decrypted virtual data group is compared with the real data group. If only the trap field is inconsistent, cross-validation is performed and communication continues. Otherwise, if other than the trap field is inconsistent, a transmission error or attack is detected, and communication is blocked.

[0011] Preferably, the steps for generating real data groups and virtual data groups based on fragmented data are as follows: the fragmented data is obfuscated using the methods in steps S201 to S204 to generate real data groups. The real data groups contain the real information of the original data fragments and exist in an encrypted and obfuscated form. The fragmented data of the real data groups before encryption and obfuscation is copied. The trap fields of the fragmented data are randomly selected and replaced with fake values ​​that conform to the format. The same encryption and obfuscation process as that for the real data groups is applied to the replaced data to generate virtual data groups.

[0012] Preferably, the resource consumption strategy library includes a botnet trap strategy, a data pollution loop strategy, and a protocol handshake swamp strategy. The botnet trap strategy refers to the decoy gateway creating fake devices and accepting control when a large number of devices are detected attempting to connect using the same vulnerability. The attacker's control commands are introduced into an infinite loop within a simulated botnet without real devices, consuming server resources. The data pollution loop strategy refers to the decoy gateway generating high-value data streams that conform to a format but have invalid or pre-set fake data, continuously sending them to the attacker and consuming their downlink bandwidth and storage space, in response to attacks attempting to steal specific types of sensor data. The protocol handshake swamp strategy refers to using a chaos platform to orchestrate complex, non-standard protocol handshake processes or simulating a large number of devices abnormally reconnecting, consuming the attacker's scanning resources and connection pool, in response to connection attempts using IoT protocols.

[0013] One or more technical solutions provided in this application have at least the following technical effects or advantages: Through continuous verification and optimization of the chaos engineering platform, the security system can dynamically adapt to the evolution of unknown attack methods and has stronger continuous security effectiveness. The progressive encryption and obfuscation process forms a deep defense chain through multi-layer encryption and obfuscation, which significantly improves the security of data and builds a dynamic, elastic, and measurable cloud platform AI intelligent agent data entry and exit security control system to ensure the security, integrity and availability of data in the process of cross-border flow. By associating obfuscation chains, the "all or nothing" security objective is achieved. If an attacker loses a ciphertext fragment, the entire subsequent part of the chain cannot be decrypted, greatly improving data security. Parallel encryption networks provide multiple layers of protection at the algorithm level for each fragment. Even if a single encryption or obfuscation algorithm is cracked, other parallel paths can still protect the data, significantly improving the system's fault tolerance and effectively defending against selective cracking attacks and single point of failure risks, ensuring the security and confidentiality of data during entry and exit processes. By consuming the attacker's bandwidth through the strategies in the resource consumption strategy library, calculating the attacker's dwell time based on the edge attack chain, and using the dwell time to delay the progress of the attack chain, the defense position is pushed from the cloud to the edge side. Interference and consumption are implemented at the beginning of the data life cycle (device access, data uplink stage) to protect the cleanliness of the data source and build a cloud-edge integrated, proactive IoT data security protection. By employing a virtual-real interweaving and attack chain verification mechanism, the security of the receiving environment is verified in real time, dynamically, and cryptographically, effectively defending against attacks that completely replace data packets. This elevates data entry and exit security from simple communication protection to a dynamic trust management system covering the entire chain from initiation to transmission to reception. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating a cloud platform AI intelligent agent data entry and exit security control method according to the present invention; Figure 2 A schematic diagram illustrating the process of constructing the obfuscation chain for this invention; Figure 3 This is a flowchart illustrating the process of generating a security control strategy based on an attack topology graph according to the present invention. Figure 4 This is a schematic diagram illustrating the process of dual-path transmission and cross-verification in this invention. Detailed Implementation

[0015] To facilitate understanding of the present invention, a more complete description of this application will be given below with reference to the accompanying drawings, which illustrate preferred embodiments of the invention. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to enable a more thorough and complete understanding of the disclosure of the present invention.

[0016] It should be noted that the terms "vertical," "horizontal," "up," "down," "left," "right," and similar expressions used in this article are for illustrative purposes only and do not represent the only possible implementation.

[0017] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0018] Example 1: Cloud platform AI agents face core security risks in cross-border data flows: data transmission may be stolen or tampered with, and unreliable cross-border links may lead to business interruptions. This solution constructs a three-pronged defense system of isolated execution, obfuscation and overlay, and proactive verification. Combining the strong isolation environment of the E2B security sandbox with the proactive fault injection capability of the CMChaos chaos engineering platform, it realizes a measurable and evolvable data security processing channel to ensure the security, integrity and availability of cross-border data flows.

[0019] Figure 1 This is a flowchart illustrating a cloud platform AI intelligent agent data entry and exit security control method according to an embodiment of the present invention, including: S101, set up an E2B sandbox service cluster in the domestic data center, fragment the data to be imported and exported in the E2B sandbox, generate a fragment mapping table, divide the fragmented data into two groups, and perform progressive encryption on the two groups of data to generate two sets of ciphertext fragment sets. The progressive encryption includes translation obfuscation encryption and inversion obfuscation encryption. Specifically, an E2B sandbox service cluster is deployed in a domestic data center. This cluster includes management nodes and compute nodes. The management nodes include an API gateway and a metadata database. The API gateway receives and processes various requests, forwarding and responding to them. The metadata database stores the cluster's metadata. The compute nodes allocate tasks to various micro-virtual machine engines. The E2B sandbox service cluster adopts a distributed architecture, distributing nodes across different physical or virtual devices to avoid single points of failure. The cluster is set up in an independent network security domain, allowing only the API port to communicate with the AI ​​agent scheduling service, ensuring that only authorized and legitimate services can interact with the cluster. Furthermore, the outbound channel is independently configured, clearly defining the flow and scope of outbound data. Within the E2B sandbox, each AI task is assigned a completely new micro-virtual machine for execution, preventing interference between tasks. When data needs to cross borders, the original data is segmented according to its content hash, the hash value of the original data is calculated, and the data block is divided into multiple fragments of similar size based on the hash value. The fragmented data is then organized into a fragment mapping table, which includes the fragment index, fragment size, and checksum. The fragment index is used to uniquely identify the position of each fragment in the original data block, the fragment size represents the size of the fragment, and the checksum is used to verify the integrity of the fragments during data transmission and processing. The fragment mapping table is stored domestically. Data fragments are randomly assigned to two groups. Progressive encryption is applied to both groups: the first group undergoes shift-and-obfuscation encryption, and the second group undergoes inversion-and-obfuscation encryption. For shift-and-obfuscation encryption, a byte-level cyclic shift operation is performed on each byte in the first group's data fragment. This byte-level cyclic shift refers to cyclically shifting 8 bits of each byte in the fragment, treating the byte as a loop with its ends connected, rotating it left or right by K bits, and filling in the bits shifted out at one end from the other. Specifically, the bits of each byte are cyclically shifted left or right, with the step size and direction of the shift serving as dynamic parameters. These dynamic parameters are dynamically configured and managed through the CMChaos platform. For example, consider a data fragment containing 4 bytes: byte A is 01001101, byte B is 10110010, byte C is 11001011, and byte D is 00111100. The shift step size K is set to 2. The operation involves a circular left shift. For byte A (01001101), starting from its most significant bit, the first two "01" bits are shifted to the left. Since it's a circular left shift, these shifted "01" bits are re-filled into the least significant bit of the byte. The remaining "001101" bits are then shifted two positions to the left, resulting in a 2-bit circular left shift of 00110101. The same method is used for bytes B, C, and D, ensuring that each byte in the original data fragment has undergone a 2-bit circular left shift. The byte-level cyclic shift operation is performed on the fragments after the byte-level cyclic shift operation. A strong encryption algorithm (SM4 algorithm) is used to encrypt the shifted fragments, generating a set of ciphertext fragments. For inversion and obfuscation encryption, a bit-order inversion operation is performed on each byte in the second set of data fragments, completely reversing the bit order of each byte. For example, if the original bit order of a byte is 01100010, it becomes 01000110 after inversion. The fragments after the bit-order inversion operation are then subjected to another byte-level cyclic shift operation. This time, the step size and direction are different from the first shift operation, and are also dynamic parameters managed through the CMChaos platform. This inversion and shift operation further increases the complexity of the data, making it more difficult for attackers to crack. The fragments after the inversion and byte-level cyclic shift operations are then encrypted using the AES-GCM encryption algorithm, generating another set of ciphertext fragments.

[0020] S102, deploys a chaos platform based on a set of ciphertext shards. The chaos platform verifies encrypted data by simulating attacks and injecting faults. Furthermore, the chaos platform is the CMChaos platform. The CMChaos chaos engineering platform is an active fault injection and drill platform used to launch simulated attacks and fault injections on data processing flows and data transmission links in a controlled environment to verify and improve the effectiveness and resilience of the entire security system. The CMChaos platform is deployed on a separate server, physically or logically isolated from the system processing ciphertext fragments, to avoid the simulated attacks causing unexpected impacts on normal business systems. The two sets of obtained ciphertext fragment sets are imported into the CMChaos platform through a data interface, and the fragment number, generation time, and dataset identifier of the ciphertext fragment sets are stored in the CMChaos platform. The chaos platform is used to simulate attacker penetration, simulate cross-border link anomalies, and simulate red team / blue team exercises. For simulating attacker penetration, fault scripts are designed and injected into the key management service in the sandbox environment of the chaos platform. The fault scripts can simulate the behavior of attackers attempting to steal some obfuscation parameters or keys, for example: by injecting malicious code, attempting to access the memory space of the key management service and obtain key fragments stored therein; or by using... Using system vulnerabilities, attempts were made to crack the obfuscation parameters used in the encryption algorithm. Using tools from the chaos platform, simulations were conducted to attempt to reconstruct the original data from intercepted ciphertext fragments using the cracking algorithm. During the cracking process, information such as the number of attacker attempts, the cracking methods used, and the time taken was recorded. The difficulty of data reconstruction was verified through simulated attacker penetration. For simulating cross-border link anomalies, various link anomalies and attacks were injected into the data outbound link using the chaos platform. Examples include: injecting high-latency faults to simulate increased latency due to network congestion or long-distance transmission; and injecting random packet loss faults to simulate network instability or attackers intentionally discarding data packets. After injecting link faults, it was observed whether the verification mechanism could detect the damage or loss of ciphertext fragments during transmission in a timely manner. The verification mechanism was verified through simulated cross-border link anomalies. For simulating red-blue team exercises, red-blue team exercises were organized, clarifying the roles and tasks of the blue and red teams. The blue team was responsible for executing the complete data processing and encryption outbound process, including data collection, encryption, fragmentation, and transmission. The red team used the CMChaos platform to launch multi-dimensional, combined attacks simultaneously, such as key theft plus link DDoS. The attack simulates complex attack scenarios in the real world. During the red team vs. blue team exercise, it verifies whether the system can dynamically rotate keys in a timely manner when it detects key leakage, and re-encrypt the ciphertext fragments with new keys. When the main transmission link is attacked and cannot be used normally, it verifies whether the system can automatically switch to the backup transmission link to ensure normal data transmission.

[0021] S103 adjusts the parameters of progressive encryption based on the verification results of the chaos platform, thereby achieving secure control over entry and exit data.

[0022] Furthermore, machine learning algorithms are used to analyze the verification data of the chaos platform to identify the range of obfuscation parameter values ​​that can effectively resist attacks under different attack scenarios. For example, cluster analysis is used to classify similar attack scenarios and corresponding obfuscation parameter combinations, and then the best-performing parameter combination in each category is analyzed to determine the optimal obfuscation parameter combination. Based on the identified optimal obfuscation parameter combination, the obfuscation algorithm parameters in progressive encryption are dynamically adjusted. For example, if a certain obfuscation parameter is found to perform well in resisting a specific type of attack, its weight in the encryption process can be increased or its value range can be adjusted. In this way, the progressive encryption parameters are adjusted according to the verification results of the chaos platform to achieve effective security control of inbound and outbound data.

[0023] The technical solutions described in the above embodiments of this application have at least the following technical effects or advantages: Through continuous verification and optimization by the CMChaos chaos engineering platform, the security system can dynamically adapt to the evolution of unknown attack methods, and has stronger continuous security effectiveness. The progressive encryption and obfuscation process forms a deep defense chain through multi-layer encryption and obfuscation, which significantly improves the security of data and builds a dynamic, elastic, and measurable cloud platform AI intelligent agent data entry and exit security control system to ensure the security, integrity and availability of data in the process of cross-border flow.

[0024] Example 2: Example 1 is a deep evolution of the progressive encryption and obfuscation process. However, Example 1 has two potential vulnerabilities: data fragments are independent and may be selectively cracked, and the single processing path may constitute a single point of failure. Example 2 introduces cryptographic associations between fragments and multi-path parallel heterogeneous processing, weaving the originally isolated data fragments into a cryptographically linked whole, effectively defending against selective cracking attacks and achieving enhanced chain security, such as... Figure 2 As shown.

[0025] S201: Based on the fragmented data, a chain-like initial seed is generated using a random algorithm. Multiple encryption channels are set for each fragment of data, and these multiple encryption channels constitute a parallel encryption network. Specifically, a 128-bit chained initial seed is generated from the fragmented data using a hardware random number generator (TRNG). A unique identifier is added to the header of each data fragment. Multiple encryption channels are set for each data fragment. In this embodiment, the multiple encryption channels refer to three encryption channels. The first encryption channel uses a byte circular shift and SM4 encryption algorithm for obfuscation. The second encryption channel uses a bit order reversal, byte circular shift and AES-GCM encryption algorithm for obfuscation. The third encryption channel uses a byte block permutation and ChaCha20 encryption algorithm for obfuscation. The three encryption channels constitute a parallel encryption network. Each data fragment is simultaneously input into the three encryption channels, and each encryption channel independently processes different copies of the same fragment.

[0026] S202: Based on the initial chain seed and the first data fragment, the hash function is used to calculate the current chain state value, the first data fragment is input into the parallel encryption network, and the ciphertext results of the parallel encryption network are synthesized to generate the first obfuscated fragment; Furthermore, the formula for calculating the current chain state value using a hash function based on the initial chain seed and the first data shard is as follows: ,in, This is the current chain state value, representing the dynamic chain state identifier after the first data shard has been processed. It is used to bind the processing results of previous shards to ensure chain dependency. For hash message authentication code functions, This serves as the initial seed for the chain encryption, acting as the starting point and root key for the entire chain encryption process. The string concatenation operation is used to join two data blocks (strings or byte streams) sequentially into a new data block. This is a hash function used to generate a fixed-length digest for the first data fragment. The first data fragment, i.e., the first fragment after the original data has been divided, is input into a parallel encryption network. In the parallel encryption network, the first channel performs a byte cyclic shift on the first data fragment, then encrypts it with the national standard SM4, outputting an intermediate result T1. The second channel reverses the bit order of the first data fragment, then performs a byte cyclic shift, and encrypts it with AES-GCM, outputting an intermediate result T2. The third channel performs byte block permutation on the first data fragment, then encrypts it with ChaCha20-Poly1305, outputting an intermediate result T3. The shift step size of the three encryption channels is calculated from the preset step size and the current chain state value. ,in, The translation step size for performing a byte loop translation, Based on step size, The i-th byte of the current chain state value is used to split the global information of the chain state into local bytes, which are then used to generate a dynamic step size specific to each shard. For modulo 256 operations, the three intermediate results are merged to generate the first obfuscated fragment, using the following formula: ,in, For the first obfuscated fragment, , and These are the intermediate results output from the three encrypted channels. This is an XOR operation.

[0027] S203, calculate the second chain state seed based on the current chain state value and the first obfuscated fragment, and execute step S202 to generate the second obfuscated fragment; process the remaining data fragments in sequence according to steps S201 and S202 to obtain multiple obfuscated fragments; Specifically, the formula for calculating the second chain state seed based on the current chain state value and the first obfuscated shard is as follows: ,in, This is the seed for the second chain state; SHA_3_512 is the hash function. This is the current chain state value. For the first obfuscated fragment, This indicates a string concatenation operation. The second chain state value is calculated using the formula for calculating chain state values ​​based on the second chain state seed. The second data fragment is input into the parallel encryption network. The translation step size of the byte cyclic translation is calculated based on the preset step size and the second chain state value. The ciphertext results of the parallel encryption network are synthesized to generate the second obfuscated fragment. The remaining data fragments are processed sequentially according to steps S201 and S202 to obtain multiple obfuscated fragments.

[0028] S204: Based on multiple obfuscated fragments, an obfuscated chain is constructed using association rules. This obfuscated chain is input to the chaos platform, and step S103 is executed to obtain an adjusted obfuscated chain. The adjusted obfuscated chain is then transmitted to the overseas receiving end via a secure network channel. Further, multiple obfuscated fragments are used to construct an obfuscated chain using association rules. These association rules refer to the process of deriving a unique identifier from the initial key using a cryptographic hash function, dividing it into sub-identifiers, and embedding them into the header of each fragment to form an initial obfuscated chain. This obfuscated chain is then input to the chaos platform, and step S103 is executed to obtain an adjusted obfuscated chain. The adjusted obfuscated chain is then transmitted to the overseas receiving end via a secure network channel.

[0029] The technical solutions described in the above embodiments of this application have at least the following technical effects or advantages: By associating obfuscation chains, the "all or nothing" security objective is achieved. If an attacker loses a ciphertext fragment, the entire subsequent part of the chain cannot be decrypted, which greatly improves the security of the data. The parallel encryption network provides multiple layers of protection at the algorithm level for each fragment. Even if a single encryption or obfuscation algorithm is cracked, other parallel paths can still protect the data, which significantly improves the fault tolerance of the system, effectively defends against selective cracking attacks and single point of failure risks, and ensures the security and confidentiality of data during entry and exit processes.

[0030] Example 3: While Examples 1 and 2 above are passive defense paradigms focused on self-reinforcement, concealing weaknesses, and dynamic evasion, this example employs an active counter-attack paradigm of proactive detection, luring the enemy in, and resource depletion. Due to their terminal vulnerabilities, protocol vulnerabilities, and massive connectivity, IoT devices become the core entry point for attackers to penetrate cloud platforms. By constructing a "cloud-edge integrated, proactive" defense system, penetration attacks are intercepted at the edge before data reaches the cloud platform, attacker resources are depleted, and the cleanliness of the data source is protected. Figure 3 As shown.

[0031] S301: Deploy an E2B sandbox service cluster at the IoT gateway to build an edge decoy gateway. Generate a fake data stream based on the edge decoy gateway and embed the monitoring module into the IoT protocol gateway. Use the monitoring module to monitor IoT attack characteristics and match the policies in the resource consumption policy library with policy priority according to the attack characteristics. Specifically, NVIDIA Jetson AGX Xavier was selected as the IoT gateway. An E2B sandbox service cluster was installed in the IoT gateway. Based on the functional interface of the E2B sandbox service cluster, a configuration tool was used to generate virtual IoT devices such as cameras and sensors. For each device, parameters such as device ID, data reporting frequency consistent with real devices, and data format simulating real devices were set. A data generation module was used to generate a fake data stream in encrypted form, which was mixed with real data at a ratio of 1:3. Finally, the mixed data was transmitted uplink to the cloud or other target locations through the IoT gateway, thus completing the construction of the edge decoy gateway. Based on the actual architecture and protocols used in the IoT system, the protocol gateways that require embedded monitoring modules are determined. The monitoring modules are configured according to these gateways, and they possess connection monitoring, protocol field registration, and data volume monitoring functions. The monitoring modules collect device connection and data transmission information in real time, store the collected data in a local database, and use the Spearman correlation coefficient algorithm to perform real-time analysis on the collected data. This analysis examines connections, protocol fields, and data volume to determine if any anomalies exist. For connection issues, if the device connection frequency exceeds the normal peak by 200%, it is considered an abnormal connection, indicating a brute-force attack or scanning attack. For protocol field issues, if the hash value of the message field does not match the preset normal hash value, it indicates that the protocol field has been tampered with, suggesting a man-in-the-middle attack or malicious code injection. For data reporting, a sudden increase of 500% in single-device data volume or continuous low-frequency reporting (< 0.1 times / second) indicates abnormal data transmission due to data theft or device malfunction. When the attack characteristic analysis algorithm detects anomalies, it generates corresponding alarm information, including the suspicious device IP, protocol type, and key information about the attack behavior. A resource consumption strategy library is constructed using botnet trapping, data pollution loopback, and protocol handshake swamp strategies. The botnet trapping strategy involves a decoy gateway dynamically creating tens of thousands of spoofed devices and gaining control when a large number of devices attempt to connect using the same vulnerability. The attacker's control commands are then routed into an infinite loop within a sandbox-simulated botnet devoid of real devices, consuming server resources. The data pollution loopback strategy targets attacks attempting to steal specific types of sensor data (such as video streams or industrial telemetry data). The decoy gateway generates high-value data streams that conform to a specific format but contain invalid or pre-set fake data, continuously feeding them to the attacker. Sending data through the edge defense module consumes its downlink bandwidth and storage space. The protocol handshake swamp strategy refers to the use of CMChaos to orchestrate complex, non-standard protocol handshake processes or simulate a large number of abnormal reconnections of devices to consume the attacker's scanning resources and connection pool for connection probing of IoT protocols. The strategies are prioritized according to the attack threat level: botnet trap strategy > data pollution loopback strategy > protocol handshake swamp strategy. After receiving the alarm information uploaded by the monitoring module, the edge defense module matches the corresponding strategy in the resource consumption strategy library according to the attack behavior type and the preset strategy priority. It checks the resource consumption of the matched strategy to ensure that the resource consumption of a single strategy is ≤ 30% of the total computing power of the edge device to avoid affecting the normal operation of the defense engine. The matched strategy is executed to consume the attacker's resources according to the execution action defined by the strategy.

[0032] S302, constructs an edge attack chain by embedding decryption tasks in a forged data stream; if the attacker parses the forged data, the edge attack chain is triggered, and the attacker's dwell time is calculated based on the edge attack chain; Furthermore, an RSA-2048 decryption task is embedded in the forged data stream. This task is embedded in specific fields of the data packets. For example, an identifier field is added to the header or trailer of a custom protocol data packet to indicate the presence of a decryption task. The encrypted ciphertext is included in the payload of the data packet. When an attacker attempts to parse the forged data stream and read the contents of various fields, a edge attack chain is triggered upon encountering a field containing the identifier of a decryption or hash calculation task. If the attacker attempts to modify protocol fields in the forged data stream to tamper with data or bypass security checks, this also triggers an edge attack chain. Once an edge attack chain is triggered, the attacker's device needs to execute the embedded decryption task. For an RSA-2048 decryption task, a single decryption operation may consume ≥500ms of computation time on a typical attacker's device. If the attacker continuously attempts to parse the forged data or modify protocol fields, constantly triggering edge attack chains, the attacker's device will continuously execute decryption. As the number of triggers increases, the CPU of the attacker's device will be consumed. The CPU utilization rate will gradually increase, eventually reaching ≥80%. High CPU utilization will lead to a decrease in device performance, causing issues such as lag and slow response, forcing the attacker to interrupt the attack to restore normal device operation. The edge device will monitor the attacker's dwell time in real time, which is the time interval from the first triggering of the edge attack chain to the current moment. When the attacker's dwell time reaches a certain threshold, the edge device will automatically upgrade the decryption algorithm in the trap from RSA-2048 to ECC-384. The formula for calculating the attacker's initial dwell time is: ,in, This refers to the moment when the attacker first triggers the edge attack chain, i.e., the initial dwell time. At the current moment, when the edge device dynamically adjusts the trap complexity according to the rules, the attacker's dwell time will be affected. The formula for calculating the adjusted dwell time is: ,in, The adjusted stay time, The initial dwell time, The i-th adjustment term represents the change in residence time caused by specific factors. This represents the total number of adjustment items.

[0033] S303 uploads the matched resource consumption strategy and the attacker's dwell time to the Chaos Platform. The Chaos Platform aggregates data from multiple edge nodes to generate an attacker topology map, and generates security control strategies based on the attacker topology map.

[0034] Specifically, the matching resource consumption strategy and the attacker's dwell time are uploaded to the Chaos Platform. Attack data is aggregated by geographical region, regional attack trends are identified, and the dwell time distribution is analyzed. Data from different nodes is associated with the attack ID to construct attack paths. Neo4j is used to store the relationship between attackers, victim nodes, and attack methods. A dynamic topology map is generated using ECharts, where node size represents attack frequency and color represents threat level. If the topology map shows that the attacker moves laterally through ports, a policy to close high-risk ports is generated. If the dwell time exceeds a preset threshold, a policy to isolate infected nodes is triggered.

[0035] The technical solutions in the above-described embodiments of this application have at least the following technical effects or advantages: by consuming the attacker's bandwidth through the strategies in the resource consumption strategy library, calculating the attacker's dwell time according to the edge attack chain, using the dwell time to delay the progress of the attack chain, pushing the defense position from the cloud to the edge side, implementing interference and consumption at the beginning of the data lifecycle (device access, data uplink stage), protecting the cleanliness of the data source, and building a cloud-edge integrated, proactive IoT data security protection.

[0036] Example 4: The above three examples all focus on the security control of raw data awaiting export. This example controls data after export. In this example, the raw data awaiting export is split into a real data group and one or more virtual data groups in a sandbox. The virtual data group consists of replacing some key fields in the real data group with fake data conforming to the business format. Simultaneously, an attack chain containing the recovery key for the virtual data group is generated. The real and virtual data groups undergo cross-verification during overseas transmission. The attack chain is used to test the receiver: only a receiver in a secure state can correctly identify the virtual data group and securely recover the real data using the attack chain. Figure 4 As shown.

[0037] S401, Generate real data groups and virtual data groups based on the fragmented data, construct an attack chain based on the virtual data groups, and determine the receiver's environment based on the attack chain. If the environment is secure, execute step S402; otherwise, if the environment is insecure, trigger sponge defense for adjustment. Furthermore, the fragmented data is obfuscated using the methods in steps S201 to S204 to generate a real data group. This real data group contains the true information of the original data fragments and exists in an encrypted and obfuscated form. The fragmented data of the real data group before encryption and obfuscation is copied. Trap fields of the fragmented data are randomly selected and replaced with spurious values ​​conforming to a certain format. The same encryption and obfuscation process as the real data group is applied to the replaced data to generate a virtual data group. An attack chain is constructed based on the virtual data group. The attack chain includes an attack part and a decryption part. The virtual data group identifier, trap field location, virtual data group recovery key, and attack data are collected. The virtual data group identifier, trap field location, and virtual data group recovery key are integrated into the decryption part, and the attack data is integrated into the attack part, thus forming the attack chain. Based on the attack chain... The attack component submits the attack data to the receiver's service interface using a payload. If the receiver can intercept the attack data, the environment is considered secure; otherwise, it is considered insecure. The receiver marks the key pair and service instance used in the current session as suspicious and routes them to a reinforced isolation diagnostic sandbox. In the diagnostic sandbox, based on the CMChaos platform, a comprehensive security chaos exercise based on the latest threat intelligence is launched on its own environment. By simulating various real security attack scenarios, existing problems are checked and repairs are made accordingly. The sender suspends sending new real data to the receiver and instead sends a higher proportion and more complex virtual data groups and attack chains. Based on the receiver's diagnostic request or preset strategy, repair probes are delivered through the attack chains to repair and adjust the receiver's problems.

[0038] S402 transmits the real data group and the virtual data group in a dual-path manner. The receiver performs cross-validation on the received data. If the verification results are consistent, an alarm is triggered. If the verification results are inconsistent, the virtual data group is decrypted through the attack chain. The decrypted virtual data group is compared with the real data group. If only the trap field is inconsistent, cross-validation is performed and communication continues. Otherwise, if other than the trap field is inconsistent, a transmission error or attack is detected, and communication is blocked.

[0039] Specifically, real and virtual data groups are sent via different network paths or staggered time windows. The selection of different network paths is optimized based on network topology, bandwidth, and latency to ensure that the two paths are independent and have different transmission characteristics. The staggered time windows, by offsetting the sending times of the two data groups, make it difficult for attackers to simultaneously capture the complete data stream, further enhancing the concealment and security of data transmission. After receiving the data, the receiver buffers the real and virtual data groups separately. The receiver performs the same decryption process on both the buffered real and virtual data groups and compares the two decrypted data groups. If... If the two sets of data are completely identical, it indicates that the attacker replaced both the real and virtual data sets with the same copy, thus bypassing the dual-path transmission security mechanism. The receiver will immediately trigger a packet switching alert, notifying relevant personnel to take appropriate security measures, such as stopping data transmission and conducting a security audit. If the two sets of data are inconsistent, the receiver will further analyze the inconsistent fields to obtain the virtual data set identifier and recovery key from the attack chain. The virtual data set identifier is used to locate the virtual data set in the cache, and the recovery key is used to decrypt and restore the virtual data set within a trusted execution environment. The receiver will then compare the restored data with the real data set. If only the trap fields differ, and Furthermore, this discrepancy is in line with expectations, meaning the transmission process was complete and the data has not been maliciously tampered with by an attacker. Trap fields are special fields pre-set in the data to detect whether it has been tampered with. By verifying the differences in trap fields, the receiver can confirm the integrity and authenticity of the data, thus achieving cross-validation. If inconsistencies occur in non-trap fields, it indicates an error occurred during transmission, such as network packet loss, data corruption, or data tampering by an attacker. The receiver verifies the attack chain signature to ensure the attack chain's source is credible and has not been tampered with. After completing data verification and processing, the receiver generates a verification receipt. The verification receipt contains information such as the attack chain processing results and environment proof digest. It is used to provide feedback to the sender on the security status and processing results of the receiver. The environment proof digest is generated through trusted computing technology to prove the trustworthiness of the receiver's operating environment. The receiver signs the verification receipt to ensure that the source of the receipt is trustworthy and has not been tampered with. Then, the receiver sends the signed verification receipt back to the sender. After receiving the verification receipt, the sender will verify it. If the verification is successful, it means that the receiver's environment is trustworthy and the data transmission and processing meet the security requirements. The sender can continue to carry out subsequent data transmission or business operations. If the verification fails, the sender blocks communication.

[0040] The technical solutions described in the above embodiments of this application have at least the following technical effects or advantages: by using a virtual-real interweaving and attack chain verification mechanism, real-time, dynamic, and cryptographically enforced verification of the security of the receiving environment is achieved, effectively defending against attacks that completely replace data packets, and upgrading data entry and exit security from simple communication protection to a dynamic trust management system covering the entire link from initiation to transmission to reception.

[0041] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for controlling the entry and exit security of AI intelligent agent data on a cloud platform, characterized in that, include: S101, set up an E2B sandbox service cluster in the domestic data center, fragment the data to be imported and exported in the E2B sandbox, generate a fragment mapping table, divide the fragmented data into two groups, and perform progressive encryption on the two groups of data to generate two sets of ciphertext fragment sets. The progressive encryption includes translation obfuscation encryption and inversion obfuscation encryption. S102, deploys a chaos platform based on a set of ciphertext shards. The chaos platform verifies encrypted data by simulating attacks and injecting faults. S103 adjusts the parameters of progressive encryption based on the verification results of the chaos platform, thereby achieving secure control over entry and exit data.

2. The cloud platform AI intelligent agent data entry and exit security control method as described in claim 1, characterized in that, The method for progressive encryption of the two sets of data is as follows: For the first set of data, a shift obfuscation encryption is performed. For the shift obfuscation encryption, a byte-level cyclic shift operation is performed on each byte in the first set of data fragments. The byte-level cyclic shift refers to cyclically shifting 8 bits of each byte in the fragment, treating the byte as a loop with its head and tail connected, rotating it to the left or right by a certain number of bits, and filling the bits shifted out at one end with the bits shifted out at the other end. The step size and direction of the shift are used as dynamic parameters, which are dynamically configured and managed through the CMChaos platform. For the second set of data, an inversion obfuscation encryption is performed, completely reversing the bit order of each byte. For the fragmented data after the bit order inversion operation, a byte-level cyclic shift operation is performed again. The step size and direction used this time are different from the first shift operation.

3. The cloud platform AI intelligent agent data entry and exit security control method as described in claim 1, characterized in that, Safety control methods also include: S201: Based on the fragmented data, a chain-like initial seed is generated using a random algorithm. Multiple encryption channels are set for each fragment of data, and these multiple encryption channels constitute a parallel encryption network. S202: Based on the initial chain seed and the first data fragment, the hash function is used to calculate the current chain state value, the first data fragment is input into the parallel encryption network, and the ciphertext results of the parallel encryption network are synthesized to generate the first obfuscated fragment; S203, calculate the second chain state seed based on the current chain state value and the first obfuscated fragment, and execute step S202 to generate the second obfuscated fragment; process the remaining data fragments in sequence according to steps S201 and S202 to obtain multiple obfuscated fragments; S204: Construct an obfuscation chain based on multiple obfuscated shards using association rules, input the obfuscation chain into the chaos platform, execute step S103 to obtain the adjusted obfuscation chain, and transmit the adjusted obfuscation chain to the overseas receiving end through a secure network channel.

4. The cloud platform AI intelligent agent data entry and exit security control method as described in claim 3, characterized in that, The formula for calculating the current chain state value is: ,in, This is the current chain state value, representing the dynamic chain state identifier after the first data shard has been processed. It is used to bind the processing results of the preceding shards. For hash message authentication code functions, This serves as the initial seed for the chain, acting as the starting point for the entire chain encryption process. This represents a string concatenation operation. This is a hash function used to generate a fixed-length digest for the first data fragment. This is the first data shard.

5. A cloud platform AI intelligent agent data entry and exit security control method as described in claim 4, characterized in that, The formula for calculating the second chain state seed is: ,in, This is the seed for the second chain state; SHA_3_512 is the hash function. This is the current chain state value. For the first obfuscated fragment, This indicates a string concatenation operation.

6. The cloud platform AI intelligent agent data entry and exit security control method as described in claim 1, characterized in that, Safety control methods also include: S301: Deploy an E2B sandbox service cluster at the IoT gateway to build an edge decoy gateway. Generate a fake data stream based on the edge decoy gateway and embed the monitoring module into the IoT protocol gateway. Use the monitoring module to monitor IoT attack characteristics and match the policies in the resource consumption policy library with policy priority according to the attack characteristics. S302, constructs an edge attack chain by embedding decryption tasks in a forged data stream; if the attacker parses the forged data, the edge attack chain is triggered, and the attacker's dwell time is calculated based on the edge attack chain; S303 uploads the matched resource consumption strategy and the attacker's dwell time to the Chaos Platform. The Chaos Platform aggregates data from multiple edge nodes to generate an attacker topology map, and generates security control strategies based on the attacker topology map.

7. The cloud platform AI intelligent agent data entry and exit security control method as described in claim 6, characterized in that, The formula for calculating the attacker's dwell time is: ,in, For the duration of stay, The initial dwell time, The i-th adjustment term represents the change in residence time caused by specific factors. This represents the total number of adjustment items.

8. A cloud platform AI intelligent agent data entry and exit security control method as described in claim 1, characterized in that, Safety control methods also include: S401, Generate real data groups and virtual data groups based on the fragmented data, construct an attack chain based on the virtual data groups, and determine the receiver's environment based on the attack chain. If the environment is secure, execute step S402; otherwise, if the environment is insecure, trigger sponge defense for adjustment. S402 transmits the real data group and the virtual data group in a dual-path manner. The receiver performs cross-validation on the received data. If the verification results are consistent, an alarm is triggered. If the verification results are inconsistent, the virtual data group is decrypted through the attack chain. The decrypted virtual data group is compared with the real data group. If only the trap field is inconsistent, cross-validation is performed and communication continues. Otherwise, if other than the trap field is inconsistent, a transmission error or attack is detected, and communication is blocked.

9. A cloud platform AI intelligent agent data entry and exit security control method as described in claim 8, characterized in that, The steps for generating real data groups and virtual data groups based on fragmented data are as follows: the fragmented data is obfuscated using the methods in steps S201 to S204 to generate real data groups. The real data groups contain the real information of the original data fragments and exist in an encrypted and obfuscated form. The fragmented data of the real data groups before encryption and obfuscation is copied. The trap fields of the fragmented data are randomly selected and replaced with fake values ​​that conform to the format. The same encryption and obfuscation process as that for the real data groups is applied to the replaced data to generate virtual data groups.

10. A cloud platform AI intelligent agent data entry and exit security control method as described in claim 6, characterized in that, The resource consumption strategy library includes botnet trap strategy, data pollution loop strategy, and protocol handshake swamp strategy. The botnet trap strategy refers to the decoy gateway creating fake devices and accepting control when a large number of devices are detected attempting to connect using the same vulnerability. The attacker's control commands are introduced into an infinite loop within a sandbox-simulated botnet without real devices, consuming server resources. The data pollution loop strategy refers to the decoy gateway generating a high-value data stream that conforms to a certain format but has invalid or pre-set fake data, continuously sending it to the attacker and consuming their downlink bandwidth and storage space, in response to attacks attempting to steal specific types of sensor data. The protocol handshake swamp strategy refers to the use of a chaotic platform to orchestrate complex, non-standard protocol handshake processes for connection probing of IoT protocols, or to simulate a large number of devices abnormally reconnecting, thereby consuming the attacker's scanning resources and connection pool.

Images