A hardware sandboxing connection device and method

By employing a physically isolated dual-channel architecture and a stateless hardware AI sandbox, the problems of fragile data connector isolation, easy tampering during auditing, and insufficient AI feature transformation capabilities are solved. This enables secure and irreversible data flow with distinctive features, ensuring data security and integrity.

CN123116366APending Publication Date: 2026-11-17百望股份有限公司
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202611223249.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-13
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing data connectors suffer from vulnerabilities in isolation, are susceptible to attack, have soft auditing capabilities that are easily tampered with, and lack AI-based feature transformation capabilities. Furthermore, existing trusted execution environments are vulnerable to side-channel attacks and plaintext data residue.

Method used

It adopts a physically isolated dual-channel architecture, a hardware stateless AI sandbox, physical-level forced erasure, and underlying hardware audit evidence storage. The physically isolated dual-channel architecture achieves complete separation of data flow and control flow. The hardware stateless AI sandbox is used for AI intelligent reasoning to generate dimensionality-reduced AI feature vectors. Physical-level forced erasure ensures that data does not leave the domain. Combined with the underlying hardware audit module, it generates tamper-proof evidence storage.

Benefits of technology

It enables secure, characteristic-based data circulation, eliminates the risk of unauthorized access, ensures data security and integrity, avoids plaintext data residue, achieves massive data compression and irreversibility, and meets the legal requirements for evidence preservation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116366A_ABST
    Figure CN123116366A_ABST
Patent Text Reader

Abstract

This invention belongs to the field of infrastructure technology, specifically relating to a hardware sandboxing connection device and method. Addressing the problems of existing technologies such as fragile software isolation susceptible to unauthorized access, easily tampered audit logs, and lack of AI feature transformation capabilities, this invention constructs a dual-channel architecture for control and data channels through physically isolated first and second network interfaces. A stateless hardware AI sandbox is deployed in the data channel, utilizing a GPU to bypass the operating system protocol stack and directly read plaintext data from heterogeneous shared memory to perform AI inference, generating a reduced-dimensional AI feature vector, which is then encrypted and signed by a national cryptographic card. After a single inference operation, a physical-level forced erasure controller performs a hardware-level zeroing and overwriting of the sandbox memory. The underlying hardware audit and evidence storage module directly generates a timestamped physical log and outputs it with a signature via a security chip. This invention achieves the effect of ensuring that raw data does not leave the domain and that feature-based secure flow is achieved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the technical field of data circulation, privacy computing and artificial intelligence underlying infrastructure, and specifically relates to a hardware sandboxing connection device and method. Background Technology

[0002] In the current development of the data element market, building a Trusted Data Space (TDS) is the core path to achieving "data usable but invisible." However, existing data connectors mostly rely on operating system software containers (such as Docker) and software firewalls for isolation, which has the following technical shortcomings:

[0003] (1) Vulnerable to isolation and vulnerable to privilege escalation: Software isolation cannot prevent internal attacks with the highest privileges (root) on the host machine. Plaintext data in memory is easily stolen by side-channel attacks or physical memory dumps. Traditional software container isolation relies on the namespace and cgroup mechanisms of the operating system kernel. Once an attacker obtains root privileges on the host machine, they can bypass all software isolation boundaries and directly access plaintext data in the container.

[0004] (2) Auditing is easily tampered with: Relying on the operating system (OS) to record call logs, administrators can easily tamper with the contents of log files, and cannot provide legally valid hardware-level anti-counterfeiting evidence. Traditional audit logs are stored on disk in the form of files, and personnel with system administrator privileges can modify, delete or forge log files to cover up traces of illegal operations.

[0005] (3) Lack of AI feature transformation capability: Traditional gateways only perform "simple encryption and decryption and channel transfer" of data, and cannot perform real-time AI intelligent extraction of high-value unstructured data such as invoices and medical records, and cannot output high-value-added artificial intelligence feature elements. Data demanders still obtain the raw data or its simple encrypted form, and the data value is not fully released.

[0006] Furthermore, prior art document CN120910920A discloses a large-scale model corpus data interaction system, interaction method, terminal, and medium based on a Trusted Execution Environment (TEE), including a data provider, a data requester, and a trusted data space service platform. Through TEE hardware isolation and end-to-end encryption, it ensures that the corpus data exists only in plaintext form within the TEE throughout its entire lifecycle, from access to destruction. However, this prior art document still has the following shortcomings:

[0007] (1) The comparison document uses a Trusted Execution Environment (TEE) for hardware-level isolation. However, the TEE is essentially an isolated execution environment inside the processor. Its memory space shares the physical memory bus with the main processor, which poses a risk of side-channel attacks. Moreover, the computing tasks within the TEE are still subject to the scheduling and management of the operating system, and cannot completely avoid attacks at the operating system level. This invention adopts a physically isolated dual-network port dual-channel architecture, which completely separates the control flow and data flow at the physical level, fundamentally eliminating the risk of unauthorized reading caused by the mixing of control signaling and data packets.

[0008] (2) In the data processing flow of the comparison file, the corpus data exists in plaintext form within the TEE, but the plaintext data after decryption by the TEE is still stored in the memory area allocated by the TEE. This memory area shares the physical medium with the main memory. After the TEE calculation task is completed, the plaintext data may still remain in the memory until it is reclaimed by the operating system. This invention introduces a hardware-level forced physical erasure mechanism that is strongly bound to the life cycle of the AI ​​inference task. After a single inference is completed, the sandbox memory is immediately cleared and overwritten by hardware to ensure the "absolute nothingness" inside the device when power is off or idle.

[0009] (3) The audit logs of the comparison documents are generated by the Trusted Data Space Service Platform, which essentially still relies on the software-level logging mechanism. Although the immutability is enhanced by blockchain evidence storage, the log generation process may still be affected by operating system interference. The underlying hardware audit evidence storage module of this invention is physically integrated inside the national cryptographic card and operates as an independent security hardware module. It bypasses the host operating system to directly generate physical logs and provides the most legally effective exemption evidence storage.

[0010] (4) The data delivery carrier for the comparison document is a "data capsule", the core of which is a large model corpus encrypted with the SM4 algorithm. The data requester still obtains the corpus data itself. This invention deploys a hardware stateless AI sandbox in the data channel and uses heterogeneous accelerated computing units such as GPUs to perform AI intelligent reasoning on the original plaintext data, outputting a dimension-reduced AI feature vector, realizing the qualitative change of the data in the moment of leaving the domain, and fundamentally avoiding the direct outflow of the original data. Summary of the Invention

[0011] To achieve the above and other related objectives, the present invention provides a hardware sandboxing connection device and method, comprising the following steps:

[0012] To achieve the above and other related objectives, this invention provides a hardware sandboxing connection device and method. Addressing the deficiencies in the prior art, it achieves the technical effect of ensuring that the original data never leaves the domain and that characteristic-based secure flow is achieved through a dual-channel architecture with physical isolation between control flow and data flow, a stateless AI sandbox, physical-level forced erasure, and underlying hardware audit and evidence storage.

[0013] To solve the above problems, the technical solution adopted by the present invention is as follows:

[0014] A hardware sandboxing connection device and method includes: a dual-channel isolation architect, a dynamic policy gateway, a hardware stateless AI sandbox module, a physical-level forced erase controller, and a dense output device.

[0015] A method for the characteristic circulation of data elements includes the following steps:

[0016] Step 1: Receive raw plaintext data input from the data subject through a physically isolated first network interface to build a data channel; Receive call requests and identity credentials initiated by external data users through a physically isolated second network interface to build a control channel; The second network interface has an embedded underlying hardware offloading mechanism to divert externally received control commands to the control channel, thus physically isolating the control channel from the data channel.

[0017] Step 2: The dynamic policy gateway deployed on the control channel receives and verifies the externally transmitted identity credentials. After verifying that the credentials match the local access policy, it sends a hardware unlock command and a one-way source data input scheduling command to the data channel.

[0018] Step 3: Upon receiving the hardware unlock command, the stateless AI sandbox deployed on the data channel is activated. The stateless AI sandbox physically integrates a general-purpose processor, a heterogeneous accelerated computing unit, and a national cryptographic card. The heterogeneous accelerated computing unit bypasses the operating system protocol stack and directly reads the input plaintext data from the heterogeneous shared memory, loads the AI ​​model for intelligent inference calculation, and generates a dimensionality-reduced AI feature vector. The national cryptographic card performs hardware encryption and digital signature on the AI ​​feature vector.

[0019] Step 4: The encrypted AI feature vector is unidirectionally output to the external data user through the underlying hardware transmission link of the second network interface;

[0020] Step 5: At the instant that a single AI inference task is completed and the dense feature result is fully output through the second network interface, the physical level forced erase controller forcibly takes over and sends a hardware level clear and overwrite signal to the address space of the underlying volatile physical memory of the hardware stateless AI sandbox, and performs a physical level erase operation of overwrite-reset, so that the device returns to stateless standby.

[0021] A method for distributing hardware sandboxed connection devices includes the following steps:

[0022] Step 1: Receive raw plaintext data input from the data subject through a physically isolated first network interface to build a data channel; Receive call requests and identity credentials initiated by external data users through a physically isolated second network interface to build a control channel; The second network interface has an embedded underlying hardware offloading mechanism to divert externally received control commands to the control channel, thus physically isolating the control channel from the data channel.

[0023] Step 2: The dynamic policy gateway deployed on the control channel receives the call request sent by the external user through the second network interface. The call request includes identity credentials.

[0024] Step 3: The dynamic policy gateway verifies the identity credential and performs a full match between the identity credential and the local access policy;

[0025] Step 4: After successful verification, the dynamic policy gateway sends a one-way input scheduling instruction for source data to the data source scheduling system.

[0026] Step 5: The dynamic policy gateway sends a hardware unlock command to the stateless AI sandbox deployed on the data channel.

[0027] Step Six: The data source scheduling system unidirectionally inputs raw plaintext data to the stateless AI sandbox via the first network interface; the raw plaintext data is written into the shared memory input segment within the stateless AI sandbox; the stateless AI sandbox physically integrates a general-purpose processor, a heterogeneous accelerated computing unit, and a national cryptographic card; the heterogeneous accelerated computing unit bypasses the operating system protocol stack and directly reads the raw plaintext data from the heterogeneous shared memory, loads the AI ​​model for intelligent inference calculation, generates a dimensionality-reduced AI feature vector, and writes the feature results into the shared memory output segment; the national cryptographic card performs hardware encryption and digital signature on the AI ​​feature vector to obtain the encrypted AI feature vector;

[0028] Step 7: The encrypted AI feature vector is unidirectionally output to the external data user via the underlying hardware transmission link of the second network interface;

[0029] Step 8: At the instant that a single AI inference task is completed and the encrypted AI feature vector is fully output through the second network interface, the hardware stateless AI sandbox triggers a pulse to the physical level forced erase controller to indicate that the encrypted feature result has been fully output; the physical level forced erase controller forcibly takes over and sends a hardware level clear and overwrite signal to the address space of the underlying volatile physical memory of the hardware stateless AI sandbox to perform a physical level erase operation of overwrite-reset;

[0030] Step 9: After the physical level forced erase controller completes the overwrite reset, it sends an overwrite reset completion signal to the dynamic policy gateway, so that the device returns to stateless standby. Attached Figure Description

[0031] Figure 1 This is the overall system architecture and dual-channel isolation topology diagram of the present invention.

[0032] Figure 2 This is a physical architecture diagram of the internal modules of the hardware sandbox connection device of the present invention. Detailed Implementation

[0033] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0034] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the illustrations only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0035] Example 1

[0036] like Figures 1 to 2 As shown, this embodiment corresponds to the system described in claim 1, and provides a hardware sandboxing connection device and method, which is applied to the secure and characteristic-based circulation scenario of medical image data.

[0037] The system includes: a dual-channel isolation architect, a dynamic policy gateway, a hardware stateless AI sandbox module, a physical-level forced erasure controller, and a dense output device.

[0038] The dual-channel isolation architecture comprises physically isolated first and second network interfaces. The first network interface (internal network port) is directly connected to the hospital's PACS imaging system via fiber optic cable, while the second network interface (external network port) is connected to an external AI diagnostic service provider via Ethernet cable. Both network ports are connected to independent physical switch ports, ensuring complete physical isolation between data and control flows. The second network interface incorporates a low-level hardware flow classifier, specifically implemented as follows: the network interface card (NIC) chip of the second network interface has a built-in hardware flow classifier that performs hardware-level flow classification based on the destination port number and protocol type of the data packets. Data packets with port numbers 443 (HTTPS control signaling) or 8080 (API call requests) are marked as control flows by the hardware flow classifier and directed to the control channel; other types of data packets are discarded or return an error. This flow classification operation is directly executed by the NIC chip's firmware logic, bypassing the operating system's network protocol stack, ensuring that the routing of control commands is completed at the hardware level.

[0039] The dynamic policy gateway is deployed on the control channel and physically connected to the second network interface via a PCIe interface. When an external AI diagnostic service provider sends a call request through the second network interface, the request carries three identity credentials: an API token, a smart contract address, and a device hardware fingerprint. The dynamic policy gateway compares the received API token with a locally stored list of authorized tokens to verify the token's validity and scope of permissions; it queries the smart contract status through the blockchain node to verify whether the contract address matches the authorized data usage contract; and it reads the external user's device hardware fingerprint and compares it with pre-registered authorized device fingerprints. Only when all three authentication factors pass verification does the dynamic policy gateway issue a hardware unlock signal and a source data one-way input scheduling signal to the data channel.

[0040] The stateless AI sandbox module is deployed in the data channel, physically integrating a general-purpose processor, a heterogeneous accelerated computing unit, and a national cryptographic card. The general-purpose processor uses an ARM Cortex-A78 architecture with a 2.4GHz clock speed and is responsible for the overall control and scheduling of the sandbox. The heterogeneous accelerated computing unit uses an NVIDIA A100 Tensor Core GPU and is equipped with 80GB of HBM2e high-bandwidth memory. The national cryptographic card uses a PCI-E cryptographic card conforming to the GM / T 0018 standard and supports SM2 / SM3 / SM4 national cryptographic algorithms. After the hospital PACS system pushes DICOM format CT image data to the connected device through the first network interface, the data is directly written to the GPU's HBM2e high-bandwidth memory, bypassing CPU memory or operating system page cache. The GPU directly reads the image data from HBM2e using CUDADirect RDMA technology and loads a pre-trained 3D U-Net medical image segmentation model for intelligent inference calculations. The input CT image data is a 512×512×256 voxel three-dimensional image, with an original data size of approximately 128MB. After the GPU executes 3D U-Net inference, it outputs a dimensionality-reduced AI feature vector with 2048 dimensions and a data size of approximately 8KB. Upon completion of inference, a national cryptographic card performs hardware encryption on the 2048-dimensional AI feature vector using SM4-CBC mode and generates an SM3 digest signature.

[0041] The encrypted output device unidirectionally outputs the encrypted AI feature vector to an external AI diagnostic service provider via the underlying hardware transmission link of the second network interface. This underlying hardware transmission link bypasses the operating system's network protocol stack, with the network card chip's DMA engine directly reading data from the cryptographic card's output buffer and encapsulating it into an Ethernet frame for transmission. Transmitting an 8.03KB encrypted feature vector in a Gigabit Ethernet environment takes approximately 0.07ms.

[0042] The physical-level forced erase controller is physically connected to the GPU's HBM2e high-bandwidth memory via an independent hardware control bus. The erase controller has a built-in independent hardware clock and status monitoring circuitry, continuously monitoring the transmit buffer status of the second network interface. When it detects that the last byte of the encrypted feature vector has left the transmit buffer (confirmed by the TX_DONE hardware signal of the network card chip), the erase controller immediately triggers the erase process: writing an all-zero sequence to the HBM2e address space 0x00000000 to 0x13FFFFFF (80GB full address space); writing a random number sequence to the same address space; and sending a memory reset signal to the GPU. This erase operation is triggered by the erase controller's independent hardware clock, independent of operating system scheduling, and takes approximately 12 seconds to execute. After erasure, the device returns to stateless standby, awaiting the next call request.

[0043] The above technical solution achieves the following effects: Through a physically isolated dual-channel architecture, control flow and data flow are completely separated at the physical level, eliminating the risk of unauthorized access caused by the mixing of control signaling and data packets in traditional software gateways; by bypassing the OS protocol stack with the GPU, plaintext data is read directly from HBM2e, eliminating the operating system's access path to plaintext data; hardware encryption and digital signature of AI feature vectors are performed using a national cryptographic card, ensuring the security and integrity of outbound data; physical-level forced erasure of the controller's hardware-level zeroing and overwriting ensures that no plaintext data remains in the sandbox memory after a single inference operation; in medical image data circulation scenarios, the original CT image (128MB) is converted into a 2048-dimensional AI feature vector (8KB), compressing the data volume by approximately 16,000 times, and the AI ​​feature vector is irreversibly restored to the original image, fundamentally achieving "original data never leaving the domain."

[0044] Example 2

[0045] This embodiment corresponds to the method described in claim 7, providing a data element feature circulation method, which is applied to the secure feature circulation scenario of medical image data.

[0046] Step 1: The system receives raw plaintext data from the hospital's internal PACS imaging system via a physically isolated first network interface, establishing a data channel. It then receives call requests and credentials from external AI diagnostic service providers via a physically isolated second network interface, establishing a control channel. The second network interface incorporates a low-level hardware traffic splitting mechanism, which performs hardware-level traffic splitting based on the destination port number and protocol type of the data packets, directing control commands to the control channel and thus physically isolating the control channel from the data channel.

[0047] Step 2: The dynamic policy gateway deployed on the control channel receives and verifies the externally input identity credentials. The dynamic policy gateway compares the API Token with the local list of authorized tokens to verify the token's validity period and scope of permissions; it queries the smart contract status through the blockchain node to verify if the contract address matches; and it reads the hardware fingerprint of the external user device and compares it with the pre-registered authorized device fingerprint. Only after all verifications pass will the gateway issue a hardware unlock command and a one-way input scheduling command for source data to the data channel.

[0048] Step 3: The stateless AI sandbox deployed in the data channel starts upon receiving the hardware unlock command. The stateless AI sandbox physically integrates an ARM Cortex-A78 general-purpose processor, an NVIDIA A100 Tensor Core GPU, and a national cryptographic card conforming to the GM / T 0018 standard. The GPU bypasses the operating system protocol stack and directly reads the input DICOM format CT image data from the HBM2e high-bandwidth memory, loads the pre-trained 3D U-Net model for intelligent inference, and outputs a 2048-dimensional AI feature vector. The national cryptographic card performs SM4-CBC hardware encryption and SM3 digital signature on the AI ​​feature vector.

[0049] Step 4: The encrypted AI feature vector is sent unidirectionally to the external AI diagnostic service provider through the underlying hardware of the second network interface.

[0050] Step 5: At the instant that a single AI inference task is completed and the dense state feature results are fully output, the physical level forced erase controller sends a hardware level clear and overwrite signal to the GPU's HBM2e address space to perform physical level erase operations of writing all-zero sequences, writing random number sequences, and resetting memory, so that the device returns to stateless standby.

[0051] The above technical solution achieves the following effects: Through a physically isolated dual-channel architecture, control flow and data flow are completely separated at the physical level, eliminating the risk of unauthorized access caused by the mixing of control signaling and data packets in traditional software gateways; by bypassing the OS protocol stack with the GPU, plaintext data is read directly from HBM2e, eliminating the operating system's access path to plaintext data; hardware encryption and digital signature of AI feature vectors are performed using a national cryptographic card, ensuring the security and integrity of outbound data; physical-level forced erasure of the controller's hardware-level zeroing and overwriting ensures that no plaintext data remains in the sandbox memory after a single inference operation; in medical image data circulation scenarios, the original CT image (128MB) is converted into a 2048-dimensional AI feature vector (8KB), compressing the data volume by approximately 16,000 times, and the AI ​​feature vector is irreversibly restored to the original image, fundamentally achieving "original data never leaving the domain."

[0052] Example 3

[0053] This embodiment corresponds to the system described in claim 2. Based on the system described in claim 1, the dynamic policy gateway is further defined to adopt a multi-factor authentication mechanism and is applied to a high-security-level characteristic circulation scenario for financial bill data.

[0054] The system includes: a dual-channel isolation architect, a dynamic policy gateway, a hardware stateless AI sandbox module, a physical-level forced erasure controller, and a secret output device. The dynamic policy gateway employs a multi-factor authentication mechanism when verifying identity credentials. This mechanism includes at least two of the following: API Token verification, smart contract address matching, and hardware fingerprint binding. A hardware unlock command is only issued after all authentication factors have passed verification.

[0055] The dual-channel isolation architecture includes a physically isolated first network interface and a second network interface. The first network interface (internal network port) is connected to the bank's internal electronic invoicing system via fiber optic cable, while the second network interface (external network port) is connected to an external financial and tax analysis service provider via network cable.

[0056] The dynamic policy gateway is deployed on the control channel and physically connected to the second network interface via a PCIe interface. When an external financial and tax analysis service provider sends a call request through the second network interface, the request carries three identity credentials: an API token, a smart contract address, and a device hardware fingerprint. The dynamic policy gateway uses a multi-factor authentication mechanism for identity verification.

[0057] (1) API Token Verification: The gateway compares the received API Token with the locally stored list of authorization tokens to verify the token's validity period and scope of authority. In this embodiment, the API Token adopts the JWT format, which includes standard fields such as issuer (iss), subject (sub), validity period (exp), and scope of authority (scope), as well as a custom invoice type authorization field (invoice_types). The dynamic policy gateway uses the public key of the bank's authorization center to verify the signature validity of the Token and checks whether the Token is within its validity period and whether the scope of authority includes the requested invoice type.

[0058] (2) Smart Contract Address Matching: The gateway queries the smart contract status through an Ethereum blockchain node (a local full node to ensure that the query results cannot be tampered with). The smart contract address carried in the call request is 0x742d35Cc6634C0532925a3b844Bc9e7595f0bEb (example address). The dynamic policy gateway verifies whether the contract address is in the authorized contract whitelist and checks whether the remaining number of uses (uses_remaining) in the contract is greater than 0 and whether the current time is within the contract validity period (valid_until).

[0059] (3) Hardware fingerprint binding: The gateway reads the CPU serial number (obtained via Intel PTT technology) and motherboard serial number (read via ACPI table) of the external user server and compares them with the hardware fingerprint collected in advance during the authorization and registration phase. The comparison adopts a fuzzy matching strategy, which allows changes in the fingerprint of a single component due to hardware replacement, but requires that at least 80% of the fingerprint features match.

[0060] The dynamic policy gateway will only issue a hardware unlock command to the data channel after all three authentication factors have passed verification. If any factor fails verification, the gateway records the reason for the failure and returns an error response through the second network interface.

[0061] The stateless AI sandbox module is deployed in the data channel, physically integrating a general-purpose processor, heterogeneous accelerated computing units, and a national cryptographic card. The original plaintext data of bank electronic invoices is directly written to the GPU's HBM2e high-bandwidth memory via the first network interface. The GPU loads a pre-trained graph neural network (GNN) model to extract invoice risk features, outputting a 512-dimensional invoice risk feature vector. This feature vector contains semantic information such as the invoice's authenticity probability, the issuer's credit score, and transaction anomaly degree, and can be irreversibly restored to the original invoice data. The national cryptographic card performs hardware encryption of the 512-dimensional feature vector in SM4-GCM mode and generates an SM3 signature.

[0062] The dense output device outputs the dense feature vector unidirectionally to an external financial and tax analysis service provider via a second network interface.

[0063] After the physical-level forced erase controller completes the output, it performs a standard erase mode (all-zero sequence + random number sequence + memory reset) on the GPU's HBM2e, which takes about 12 seconds, and restores the device to stateless standby.

[0064] The above technical solution achieves the following effects: By introducing a multi-factor authentication mechanism, the system's identity verification security is enhanced from the single-factor "know what" (API Token) to a three-dimensional authentication system of "know what + possess what + what it is" (Token + Contract + Hardware). Even if an attacker steals the API Token, they cannot pass the verification without an authorized smart contract and legitimate hardware device. In high-security scenarios involving financial bill data, the multi-factor authentication mechanism reduces the risk of unauthorized access to less than 0.01% of that of single-factor authentication. Furthermore, since the entire verification process of multi-factor authentication is completed at the hardware level (smart contract queries are performed through a local full node, and hardware fingerprints are read directly via PCIe), the verification process does not rely on external networks or operating system services, avoiding the risk of authentication bypass due to network interruptions or system compromise. The 512-dimensional bill risk feature vector extracted by the GNN model compresses the original invoice data (average 50KB) by approximately 100 times while retaining key risk semantic information. External financial and tax analysis service providers can perform cross-enterprise joint bill risk modeling based on the feature vector without obtaining the original invoice plaintext, achieving the technical effect of ensuring the original data never leaves the domain.

[0065] Example 4

[0066] This embodiment corresponds to the system described in claim 3. Based on the system described in claim 1, it further defines the type of AI model loaded when the heterogeneous accelerated computing unit performs AI inference and the irreversibility of the AI ​​feature vector, and is applied to the joint feature-based circulation scenario of multimodal commodity data on e-commerce platforms.

[0067] The system includes: a dual-channel isolation architect, a dynamic policy gateway, a hardware stateless AI sandbox module, a physical-level forced erasure controller, and a dense output device. When the heterogeneous accelerated computing unit performs AI inference, the loaded AI model includes at least one of a graph neural network node aggregation model, a large-model semantic extraction model, and a multimodal feature fusion model; the dimension of the dimensionality-reduced AI feature vector is lower than the feature dimension of the original plaintext data, and the AI ​​feature vector can be irreversibly restored to the original plaintext data.

[0068] The dual-channel isolation architecture includes a physically isolated first network interface and a second network interface. The first network interface (internal network port) is connected to the e-commerce platform's multimodal product database via fiber optic cable, and the second network interface (external network port) is connected to an external recommendation algorithm service provider via network cable.

[0069] The dynamic policy gateway is deployed on the control channel to verify the identity credentials of external recommendation algorithm service providers and issue hardware unlocking commands.

[0070] The stateless AI sandbox module is deployed in the data channel, physically integrating a general-purpose processor, a heterogeneous accelerated computing unit, and a national cryptographic card. The heterogeneous accelerated computing unit uses an NVIDIA H100 Tensor Core GPU and is equipped with 96GB of HBM3 high-bandwidth memory. Multimodal product data from the e-commerce platform (including product images, product description text, and structured product attribute data) is directly written to the GPU's HBM3 memory via the first network interface. The GPU loads the multimodal feature fusion model for intelligent inference computation. The multimodal feature fusion model includes the following sub-models:

[0071] (1) Image feature extraction: The Vision Transformer (ViT-L / 16) model is used to extract features from the product images. The input image size is 224×224×3, and the output image feature vector has a dimension of 1024.

[0072] (2) Text feature extraction: The BERT-large model is used to extract features from the product description text. The maximum length of the input text is 512 tokens, and the dimension of the output text feature vector is 1024.

[0073] (3) Structured feature encoding: The Embedding layer is used to encode the structured data of product attributes (category, brand, price range, etc.), and the output structured feature vector has a dimension of 512.

[0074] (4) Multimodal feature fusion: The above three feature vectors are concatenated into a 2560-dimensional joint feature vector, and then dimensionality reduction and fusion are performed through two fully connected layers (2560→1024→512) to finally output a 512-dimensional multimodal fusion feature vector.

[0075] The dimensionality of the reduced AI feature vector (512 dimensions) is significantly lower than the feature dimension of the original plaintext data (1024 dimensions for image features + 1024 dimensions for text features + 512 dimensions for structured features = 2560 dimensions), and the AI ​​feature vector cannot be irreversibly restored to the original product image, product description text, or product attribute structured data. Specifically, the original product image pixel values, text token sequences, or structured attribute fields cannot be recovered from the 512-dimensional multimodal fusion feature vector through any known inverse transformation or generative model.

[0076] The national cryptographic card performs hardware encryption (providing authentication encryption function) on the 512-dimensional multimodal fusion feature vector in SM4-GCM mode and generates an SM3 digest signature. The encrypted output device unidirectionally outputs the encrypted feature vector to an external recommendation algorithm service provider through a second network interface. After the output is completed, the physical-level forced erasure controller performs hardware-level zeroing and overwriting of the sandbox memory.

[0077] The effects of the above technical solution are as follows: Through the multimodal feature fusion model, the system can extract a unified low-dimensional feature representation from three types of heterogeneous data: product images, text descriptions, and structured attributes, realizing cross-modal data value extraction; through the semantic extraction capabilities of pre-trained large models such as ViT and BERT, the output 512-dimensional feature vector contains rich product semantic information, which can be directly used for similarity calculation and candidate recall in recommendation algorithms; since the feature vector cannot be irreversibly restored to the original data, external recommendation algorithm service providers only obtain the "digital fingerprint" of the product rather than the original product information. Even if the feature vector is leaked, it is impossible to deduce the specific image, description, or attributes of the product, fundamentally ensuring the data security of the data subject; in the multimodal product data circulation scenario, the original multimodal data (approximately 2MB per product on average) is transformed into a 512-dimensional feature vector (approximately 2KB), compressing the data volume by about 1000 times, while retaining the semantic similarity information of the product, achieving a balance between data value release and privacy protection, and achieving the technical effect that the original data never leaves the domain.

[0078] Example 5

[0079] This embodiment corresponds to the system described in claim 4. Based on the system described in claim 1, it further defines the specific erasure operation method executed by the physical-level forced erasure controller, and is applied to high-security government data exchange scenarios.

[0080] The system includes: a dual-channel isolation architect, a dynamic policy gateway, a hardware stateless AI sandbox module, a physical-level forced erase controller, and a dense output device. The physical-level forced erase controller performs physical-level erase operations including: writing an all-zero sequence to the address space of the underlying volatile physical memory, writing a random number sequence, and executing a memory reset signal, at least one of the following: the execution time of the erase operation is independent of the operating system scheduling and is triggered by the independent hardware clock of the physical-level forced erase controller.

[0081] The dual-channel isolation architecture includes a physically isolated first network interface and a second network interface. The first network interface (internal network port) is connected to the government department's citizen registration database via fiber optic cable, and the second network interface (external network port) is connected to an external statistical research institution via network cable.

[0082] The dynamic policy gateway is deployed on the control channel to verify the identity credentials of external statistical research institutions and issue hardware unlocking commands.

[0083] The stateless AI sandbox module is deployed in the data channel, physically integrating a general-purpose processor, heterogeneous accelerated computing units, and a national cryptographic card. The original plaintext data of citizen household registration is directly written to the GPU's HBM3 high-bandwidth memory via the first network interface. The GPU loads a pre-trained demographic feature extraction model and outputs a 256-dimensional demographic feature vector. This feature vector contains aggregated statistical information such as age distribution features, geographical distribution features, and education level features, and can be irreversibly restored to the original household registration data of a single citizen. The national cryptographic card performs hardware encryption of the 256-dimensional feature vector in SM4-GCM mode and generates an SM3 signature.

[0084] The dense-state outputter outputs the dense-state feature vector unidirectionally to an external statistical research institution through a second network interface.

[0085] The physical-level forced erase controller is physically connected to the GPU's HBM3 high-bandwidth memory via an independent hardware control bus. At the instant a single AI inference task is completed and the dense-state feature results are fully output, the physical-level forced erase controller performs a physical-level erase operation in deep erase mode, specifically including the following steps:

[0086] (1) First overwrite stage: Write a sequence of all zeros (0x00) to the full address space (0x00000000 to 0x17FFFFFF, 96GB) of the GPU’s HBM3 high-bandwidth memory. This stage lasts about 6 seconds and is performed by the erase controller’s DMA engine at maximum bandwidth.

[0087] (2) Second overwrite stage: A random number sequence is written to the same address space. The random numbers are generated by the hardware random number generator built into the erase controller, conforming to the CTR-DRBG algorithm of the NIST SP 800-90B standard, and the seed is provided by a physical entropy source (ring oscillator jitter). This stage lasts about 6 seconds.

[0088] (3) Third overwrite stage: Write the reverse all-one sequence (0xFF) to the same address space. This stage lasts about 6 seconds.

[0089] (4) Fourth overwrite stage: Write a random number sequence to the same address space again, with a different random number seed than in the second overwrite stage. This stage lasts about 6 seconds.

[0090] (5) Memory reset phase: A memory reset signal is sent to the GPU to reinitialize the address mapping of HBM3 and clear the internal cache and address translation table of the memory controller. This phase lasts about 0.5 seconds.

[0091] The execution time of the erase operation is independent of the operating system scheduling and is triggered by the independent hardware clock of the physical-level forced erase controller. The independent hardware clock is a 32.768kHz real-time clock crystal oscillator, which is completely isolated from the system master clock. Even if the system master clock is interfered with or stopped by an attacker, the erase controller can still perform the erase operation according to the preset timing.

[0092] After erasing is complete, the device returns to standby mode.

[0093] The above technical solution achieves the following results: By employing a deep erase mode (all zeros + random numbers + all ones + random numbers + memory reset), the system can effectively resist advanced data recovery techniques based on magnetic remanence and cold boot attacks. Research shows that after four overwrites (especially those involving random numbers), even if an attacker physically disassembles the memory chip and reads the residual charge at low temperatures, the data recovery success rate is less than 10^-6. Triggering the erase operation with an independent hardware clock ensures the deterministic timing and unbypassable nature of the erase process; even if an attacker gains the highest operating system privileges, they cannot prevent or delay the erase operation by modifying the system clock. In high-security government data exchange scenarios, the deep erase mode complies with the "Information Security Technology: Guidelines for De-identification of Personal Information" (GB / T). The highest level of requirements for sensitive data destruction in 37964-2019 provides data subjects with physical-level data destruction protection that meets national standards; through the output of 256-dimensional demographic feature vectors, external statistical research institutions can conduct population trend analysis based on aggregated features without obtaining the original household registration information of individual citizens, thus achieving a balance between data value utilization and personal privacy protection, and achieving the technical effect of ensuring that the original data never leaves the domain.

[0094] Example 6

[0095] This embodiment corresponds to the system described in claim 5. Based on the system described in claim 1, it further adds an underlying hardware audit and evidence storage module, which is applied to data element trading platforms that require compliance auditing.

[0096] The system includes: a dual-channel isolation architect, a dynamic policy gateway, a hardware stateless AI sandbox module, a physical-level forced erasure controller, a cryptographic output device, and a low-level hardware audit and evidence storage module. The low-level hardware audit and evidence storage module is physically integrated within the national cryptographic card and operates as an independent security hardware module. It is used for gating and unlocking control channels, single inference operations on data channels, and physical erasure operations, bypassing the host operating system. The security chip directly generates physical transaction logs with high-precision timestamps, which are then signed with the hardware-level private key and output to the monitoring node.

[0097] The dual-channel isolation architecture includes a physically isolated first network interface and a second network interface. The first network interface (internal network port) is connected to the internal network of the data subject via fiber optic cable, and the second network interface (external network port) is connected to the external user via network cable.

[0098] The dynamic policy gateway is deployed on the control channel to verify the identity credentials of external users and issue hardware unlock commands.

[0099] The hardware-based stateless AI sandbox module is deployed in the data channel, physically integrating a general-purpose processor, heterogeneous accelerated computing units, and a national cryptographic card. Raw plaintext data is directly written to the GPU's HBM3 via the first network interface. The GPU executes AI inference to generate dimensionality-reduced AI feature vectors, and the national cryptographic card performs hardware encryption and digital signature on the AI ​​feature vectors.

[0100] The dense-state outputter outputs the dense-state feature vector unidirectionally through the second network interface.

[0101] The physical-level forced erase controller performs a hardware-level zeroing and overwriting of the sandbox memory after the output is complete.

[0102] The underlying hardware audit and evidence storage module is physically integrated within the national cryptographic card and operates as an independent security hardware module. This module includes the following hardware components:

[0103] (1) Independent security chip: It adopts a security microcontroller that meets the Common Criteria EAL4+ security level, and has an independent ARM Cortex-M33 processor, 256KB security SRAM and 8MB security Flash memory.

[0104] (2) Independent hardware clock: It adopts a temperature-compensated crystal oscillator (TCXO) with a frequency accuracy of ±0.5ppm, providing microsecond-level timestamps.

[0105] (3) Hardware private key storage area: It adopts a physical protection structure that is tamper-proof. The private key is injected through the hardware security module (HSM) during the chip manufacturing stage and cannot be read or exported through any software interface.

[0106] (4) Hardware hash engine: Supports SM3 and SHA-256 hash algorithms, with a hash calculation throughput of 1Gbps.

[0107] The audit process for the underlying hardware audit and evidence storage module is as follows, covering gating and unlocking of the control channel, single inference of the data channel, and physical erasure operations:

[0108] (1) Gated Unlock Audit: When the dynamic policy gateway issues a hardware unlock command, the underlying hardware audit and evidence storage module receives the unlock event notification through an independent hardware signal line (an audit bus physically isolated from both the data channel and the control channel). The security chip records the following information: operation type identifier (0x01), operation timestamp (Unix microsecond-level timestamp), operation subject identifier (SM3 public key hash of the external user, 32 bytes), operation result status (0x00 indicates success), and previous status hash value (SM3 hash value of the previous transaction record, 32 bytes). The security chip uses its hardware private key to perform SM2 signing on the above information, generating a 64-byte signature value. The complete transaction record is written to the secure Flash memory and sent to the supervisory node through an independent network interface (a third network interface used only for audit output).

[0109] (2) AI Inference Audit: When the stateless AI sandbox in hardware starts AI inference, the underlying hardware audit and evidence storage module receives the inference start event notification and records the operation type identifier (0x02), operation timestamp, input data digest (SM3 hash value of the original plaintext data, 32 bytes), and AI model identifier (SM3 hash value of the pre-trained model, 32 bytes). When inference is completed, it records the operation type identifier (0x03), operation timestamp, output data digest (SM3 hash value of the AI ​​feature vector, 32 bytes), and inference time (microseconds). Both of the above pipeline records are written to the secure Flash memory after being signed with SM2 and sent to the supervisory node.

[0110] (3) Physical Erasure Audit: When the physical-level forced erasure controller initiates the erasure operation, the underlying hardware audit and evidence storage module receives the erasure initiation event notification and records the operation type identifier (0x04), operation timestamp, erasure mode identifier, and erasure address range. When erasure is complete, it records the operation type identifier (0x05), operation timestamp, erasure time, and post-state hash value (the SM3 hash value of the memory state after erasure, calculated by reading sampled data from the memory address space). Both of the above log records are signed with SM2, written to the secure Flash memory, and sent to the monitoring node.

[0111] The effects of the above technical solution are as follows: By physically integrating the underlying hardware audit and evidence storage module into the national cryptographic card, the audit function operates completely independently of the main processor and operating system. Even if an attacker gains the highest privileges on the device's operating system, they cannot tamper with, delete, or forge the audit log. Through an independent security chip and hardware private key storage area, the private key for audit signatures cannot be extracted by software, ensuring the non-repudiation of the signature. In data element trading platforms that require compliant auditing, the physical log generated by the underlying hardware audit and evidence storage module can be directly used as legal evidence, meeting the requirements of the Electronic Signature Law for reliable electronic signatures and providing strong technical support for the compliance of data element circulation. At the same time, because audit and evidence storage form a complete closed loop with data circulation and physical erasure, any data access, reasoning calculation, erasure and destruction operation is recorded immutably, achieving the secure circulation of data elements that is traceable and auditable throughout the entire process, and achieving the technical effect that the original data never leaves the domain.

[0112] Example 7

[0113] This embodiment corresponds to the system described in claim 6. Based on the system described in claim 5, it further defines the specific fields and chained storage structure of the physical flow generated by the underlying hardware audit and evidence storage module, and is applied to financial regulatory scenarios that require highly credible audit evidence.

[0114] The system includes: a dual-channel isolation architect, a dynamic policy gateway, a hardware stateless AI sandbox module, a physical-level forced erasure controller, a secret output device, and a low-level hardware audit and evidence storage module. The physical pipeline generated by the low-level hardware audit and evidence storage module includes: operation type identifier, operation timestamp, operation subject identifier, operation result status, and hash values ​​of the preceding and following states. The physical pipeline adopts an immutable chained storage structure, with the hash value of each pipeline record serving as the input parameter for the next pipeline record, forming a hardware-level audit evidence chain.

[0115] The dual-channel isolation architecture includes a physically isolated first network interface and a second network interface. The first network interface (internal network port) is connected to the bank's internal transaction database via fiber optic cable, while the second network interface (external network port) is connected to external financial regulatory agencies via network cable.

[0116] The dynamic policy gateway is deployed on the control channel to verify the identity credentials of external regulatory agencies and issue hardware unlocking commands.

[0117] The stateless AI sandbox module is deployed in the data channel, physically integrating a general-purpose processor, heterogeneous accelerated computing units, and a national cryptographic card. The original plaintext data of bank transactions is directly written to the GPU's HBM3 via the first network interface. The GPU executes AI inference to generate transaction risk feature vectors, and the national cryptographic card performs hardware encryption and digital signature on these feature vectors.

[0118] The dense-state outputter outputs the dense-state feature vector unidirectionally through the second network interface.

[0119] The physical-level forced erase controller performs a hardware-level zeroing and overwriting of the sandbox memory after the output is complete.

[0120] The underlying hardware audit and evidence storage module is physically integrated within the national cryptographic card and operates as an independent security hardware module. The physical transaction log generated by the underlying hardware audit and evidence storage module includes the following fields:

[0121] (1) Operation type identifier: 1 byte, 0x01 indicates gate unlocking, 0x02 indicates AI inference start, 0x03 indicates AI inference complete, 0x04 indicates physical erase start, 0x05 indicates physical erase complete.

[0122] (2) Operation timestamp: 8 bytes, Unix timestamp format, with a precision of microseconds.

[0123] (3) Operation subject identifier: 32 bytes, the public key hash value of the external user.

[0124] (4) Operation result status: 1 byte, 0x00 indicates success, 0x01 indicates failure.

[0125] (5) Before and after state hash value: 32 bytes, SM3 hash value of device state before and after operation.

[0126] The physical pipeline employs an immutable chained storage structure: Let the nth pipeline record be R_n, whose content includes operation type T_n, timestamp S_n, subject identifier A_n, result status C_n, and previous state hash H_{n-1}. The hash value H_n of the nth pipeline record is calculated using the SM3 algorithm:

[0127] H_n = SM3(T_n || S_n || A_n || C_n || H_{n-1})

[0128] The "||" symbol represents the byte string concatenation operation.

[0129] The previous state hash value of the (n+1)th transaction record is H_n, forming a chain structure. If an attacker attempts to tamper with the content of the kth transaction record, H_k will change, causing the previous state hash value verification of the (k+1)th and all subsequent transaction records to fail, and the tampering attempt will be detected immediately.

[0130] The effects of the above technical solution are as follows: By clearly defining the fields of the physical transaction log (operation type identifier, operation timestamp, operation subject identifier, operation result status, and hash values ​​of the previous and subsequent states), each audit record contains complete operation context information, facilitating post-event traceability and responsibility determination; through the chained storage structure, any tampering with historical transaction log records will cause the hash verification of all subsequent records to fail, providing mathematically verifiable audit integrity assurance; in financial regulatory scenarios requiring highly credible audit evidence, the chained storage structure enables the audit evidence chain to have the same data integrity protection capabilities as the blockchain, but without relying on an external blockchain network, reducing system complexity and operating costs; at the same time, since the chained hash calculation is directly executed by the hardware security chip, the calculation process cannot be interfered with by software, ensuring the authenticity and reliability of the hash value, providing financial regulatory agencies with audit evidence of data element circulation with the highest credibility, and achieving the technical effect of ensuring that the original data never leaves the domain.

[0131] Example 8

[0132] This embodiment corresponds to the method described in claim 8. Based on the method described in claim 7, it further specifies that the dynamic policy gateway adopts a multi-factor authentication mechanism and is applied to the high-security-level characteristic circulation scenario of financial bill data.

[0133] The method includes the following steps:

[0134] Step 1: Receive the original plaintext data of electronic invoices from within the bank via a physically isolated first network interface, establishing a data channel; receive call requests and identity credentials from external financial and tax analysis service providers via a physically isolated second network interface, establishing a control channel. The second network interface incorporates a low-level hardware offloading mechanism to redirect externally received control commands to the control channel, thus physically isolating the control channel from the data channel.

[0135] Step 2: The dynamic policy gateway deployed on the control channel receives and verifies the externally input identity credentials. When verifying identity credentials, the dynamic policy gateway employs a multi-factor authentication mechanism, which includes at least two of the following: APIToken verification, smart contract address matching, and hardware fingerprint binding. The hardware unlock command is only issued after all authentication factors have passed verification.

[0136] The multi-factor authentication mechanism includes three authentication factors:

[0137] (1) API Token Verification: The call request from the external financial and tax analysis service provider carries an API Token issued by the bank's authorization center. The Token adopts the JWT format and includes standard fields such as issuer (iss), subject (sub), validity period (exp), and authorization scope (scope), as well as a custom invoice type authorization field (invoice_types). The dynamic policy gateway uses the public key of the bank's authorization center to verify the signature validity of the Token and check whether the Token is within the validity period and whether the authorization scope includes the requested invoice type.

[0138] (2) Smart Contract Address Matching: The dynamic policy gateway queries the smart contract status through an Ethereum blockchain node (a local full node to ensure that the query results cannot be tampered with). The smart contract address carried in the call request is 0x742d35Cc6634C0532925a3b844Bc9e7595f0bEb (example address). The dynamic policy gateway verifies whether the contract address is in the authorized contract whitelist and checks whether the remaining number of uses (uses_remaining) in the contract is greater than 0 and whether the current time is within the contract validity period (valid_until).

[0139] (3) Hardware fingerprint binding: The dynamic policy gateway reads the CPU serial number (obtained via Intel PTT technology) and motherboard serial number (read via ACPI table) of the external user server and compares them with the hardware fingerprint collected in advance during the authorization and registration phase. The comparison adopts a fuzzy matching strategy, which allows changes in the fingerprint of a single component due to hardware replacement, but requires that at least 80% of the fingerprint features match.

[0140] The dynamic policy gateway will only issue a hardware unlock command to the data channel after all three authentication factors have passed verification. If any factor fails verification, the gateway records the reason for the failure and returns an error response through the second network interface.

[0141] Step 3: The stateless AI sandbox deployed on the data channel starts upon receiving the hardware unlock command. The stateless AI sandbox physically integrates a general-purpose processor, a heterogeneous accelerated computing unit, and a national cryptographic card. The heterogeneous accelerated computing unit uses an NVIDIA H100 Tensor Core GPU and is equipped with 96GB of HBM3 high-bandwidth memory. The original plaintext data of the bank's electronic invoice is directly written to the GPU's HBM3 via the first network interface. The GPU loads a pre-trained graph neural network (GNN) model to extract invoice risk features, outputting a 512-dimensional invoice risk feature vector. This feature vector contains semantic information such as the invoice's authenticity probability, the issuer's credit score, and transaction anomaly degree, and can be irreversibly restored to the original invoice data. The national cryptographic card performs hardware encryption on the 512-dimensional feature vector in SM4-GCM mode and generates an SM3 signature.

[0142] Step 4: The dense feature vector is output unidirectionally to the external financial and tax analysis service provider through the second network interface.

[0143] Step 5: After the physical level forced erase controller completes the output, it performs a standard erase mode (all zero sequence + random number sequence + memory reset) on the GPU's HBM3, which takes about 12 seconds, so that the device returns to stateless standby.

[0144] The above technical solution achieves the following effects: By introducing a multi-factor authentication mechanism, the system's identity verification security is enhanced from the single-factor "know what" (API Token) to a three-dimensional authentication system of "know what + possess what + what it is" (Token + Contract + Hardware). Even if an attacker steals the API Token, they cannot pass the verification without an authorized smart contract and legitimate hardware device. In high-security scenarios involving financial bill data, the multi-factor authentication mechanism reduces the risk of unauthorized access to less than 0.01% of that of single-factor authentication. Furthermore, since the entire verification process of multi-factor authentication is completed at the hardware level (smart contract queries are performed through a local full node, and hardware fingerprints are read directly via PCIe), the verification process does not rely on external networks or operating system services, avoiding the risk of authentication bypass due to network interruptions or system compromise. The 512-dimensional bill risk feature vector extracted by the GNN model compresses the original invoice data (average 50KB) by approximately 100 times while retaining key risk semantic information. External financial and tax analysis service providers can perform cross-enterprise joint bill risk modeling based on the feature vector without obtaining the original invoice plaintext, achieving the technical effect of ensuring the original data never leaves the domain.

[0145] Example 9

[0146] This embodiment corresponds to the method described in claim 9. Based on the method described in claim 7, it further defines the specific method of intelligent reasoning calculation by the AI ​​model and the irreversibility of AI feature vectors, and applies it to the joint feature-based circulation scenario of multimodal commodity data on e-commerce platforms.

[0147] The method includes the following steps:

[0148] Step 1: Receive multimodal product data from the e-commerce platform via a physically isolated first network interface, including product images, product description text, and structured product attribute data, thus constructing a data channel. Receive call requests and identity credentials from external recommendation algorithm service providers via a physically isolated second network interface, thus constructing a control channel. The second network interface incorporates an underlying hardware offloading mechanism to redirect externally received control commands to the control channel, physically isolating the control channel from the data channel.

[0149] Step 2: The dynamic policy gateway deployed on the control channel receives and verifies the externally transmitted identity credentials. After verifying that the credentials match the local access policy, it sends a hardware unlock command and a one-way source data input scheduling command to the data channel.

[0150] Step 3: The stateless AI sandbox deployed on the data channel starts upon receiving the hardware unlock command. The stateless AI sandbox physically integrates a general-purpose processor, a heterogeneous accelerated computing unit, and a national cryptographic card. The heterogeneous accelerated computing unit bypasses the operating system protocol stack and directly reads the input plaintext data from heterogeneous shared memory, loading the AI ​​model for intelligent inference computation. The intelligent inference computation of the AI ​​model includes multimodal feature fusion, specifically implemented as follows:

[0151] (1) Image feature extraction: The Vision Transformer (ViT-L / 16) model is used to extract features from the product images. The input image size is 224×224×3, and the output image feature vector has a dimension of 1024.

[0152] (2) Text feature extraction: The BERT-large model is used to extract features from the product description text. The maximum length of the input text is 512 tokens, and the dimension of the output text feature vector is 1024.

[0153] (3) Structured feature encoding: The Embedding layer is used to encode the structured data of product attributes (category, brand, price range, etc.), and the output structured feature vector has a dimension of 512.

[0154] (4) Multimodal feature fusion: The above three feature vectors are concatenated into a 2560-dimensional joint feature vector, and then dimensionality reduction and fusion are performed through two fully connected layers (2560→1024→512) to finally output a 512-dimensional multimodal fusion feature vector.

[0155] The dimensionality of the reduced AI feature vector (512 dimensions) is lower than the feature dimension of the original plaintext data (image features 1024 dimensions + text features 1024 dimensions + structured features 512 dimensions = 2560 dimensions), and the AI ​​feature vector cannot be irreversibly restored to the original product image, product description text, or product attribute structured data. Specifically, the original product image pixel values, text token sequences, or structured attribute fields cannot be recovered from the 512-dimensional multimodal fusion feature vector through any known inverse transformation or generative model.

[0156] The national cryptographic card performs SM4-GCM mode hardware encryption (providing authentication encryption function) on the 512-dimensional multimodal fusion feature vector and generates an SM3 digest signature.

[0157] Step 4: The dense feature vector is output unidirectionally to the external recommendation algorithm service provider through the second network interface.

[0158] Step 5: After the physical level forced erase controller completes the output, it performs a hardware-level zeroing and overwriting of the sandbox memory.

[0159] The effects of the above technical solution are as follows: Through multimodal feature fusion, the system can extract a unified low-dimensional feature representation from three types of heterogeneous data: product images, text descriptions, and structured attributes, realizing cross-modal data value extraction; through the semantic extraction capabilities of pre-trained large models such as ViT and BERT, the output 512-dimensional feature vector contains rich product semantic information, which can be directly used for similarity calculation and candidate recall in recommendation algorithms; since the feature vector cannot be irreversibly restored to the original data, external recommendation algorithm service providers only obtain the "digital fingerprint" of the product rather than the original product information. Even if the feature vector is leaked, it is impossible to deduce the specific image, description, or attributes of the product, fundamentally ensuring the data security of the data subject; in the multimodal product data circulation scenario, the original multimodal data (approximately 2MB per product on average) is transformed into a 512-dimensional feature vector (approximately 2KB), compressing the data volume by about 1000 times, while retaining the semantic similarity information of the product, achieving a balance between data value release and privacy protection, and achieving the technical effect that the original data never leaves the domain.

[0160] Example 10

[0161] This embodiment corresponds to the method described in claim 10. Based on the method described in claim 7, it further defines the specific erasure operation mode executed by the physical-level forced erasure controller, and is applied to high-security government data exchange scenarios.

[0162] The method includes the following steps:

[0163] Step 1: Receive raw plaintext citizen registration data from government departments through a physically isolated first network interface to establish a data channel; receive call requests and identity credentials from external statistical research institutions through a physically isolated second network interface to establish a control channel. The second network interface has an embedded underlying hardware offloading mechanism that diverts externally received control commands to the control channel, thus physically isolating the control channel from the data channel.

[0164] Step 2: The dynamic policy gateway deployed on the control channel receives and verifies the externally transmitted identity credentials. After verifying that the credentials match the local access policy, it sends a hardware unlock command and a one-way source data input scheduling command to the data channel.

[0165] Step 3: The stateless AI sandbox deployed on the data channel starts upon receiving the hardware unlock command. The stateless AI sandbox physically integrates a general-purpose processor, a heterogeneous accelerated computing unit, and a national cryptographic card. The heterogeneous accelerated computing unit bypasses the operating system protocol stack and directly reads the original plaintext data of citizen household registration from heterogeneous shared memory, loads the AI ​​model for intelligent inference calculation, and generates a dimensionality-reduced AI feature vector. The national cryptographic card performs hardware encryption and digital signature on the AI ​​feature vector.

[0166] Step 4: The encrypted AI feature vector is unidirectionally output to an external statistical research institution through the underlying hardware transmission link of the second network interface.

[0167] Step 5: At the instant the single AI inference task is completed and the dense feature result is fully output through the second network interface, the physical-level forced erase controller forcibly takes over and issues a hardware-level zero-overwrite signal to the address space of the underlying volatile physical memory of the stateless AI sandbox, performing a write-reset physical-level erase operation. The physical-level erase operation performed by the physical-level forced erase controller includes at least one of the following: writing an all-zero sequence to the address space of the underlying volatile physical memory, writing a random number sequence, and executing a memory reset signal; the execution time of the erase operation does not depend on the operating system scheduling and is triggered by the independent hardware clock of the physical-level forced erase controller.

[0168] Specifically, the physical-level forced erase controller performs physical-level erase operations using a deep erase mode:

[0169] (1) First overwrite stage: Write a sequence of all zeros (0x00) to the full address space (0x00000000 to 0x17FFFFFF, 96GB) of the GPU’s HBM3 high-bandwidth memory. This stage lasts about 6 seconds and is performed by the erase controller’s DMA engine at maximum bandwidth.

[0170] (2) Second overwrite stage: A random number sequence is written to the same address space. The random numbers are generated by the hardware random number generator built into the erase controller, conforming to the CTR-DRBG algorithm of the NIST SP 800-90B standard, and the seed is provided by a physical entropy source (ring oscillator jitter). This stage lasts about 6 seconds.

[0171] (3) Third overwrite stage: Write the reverse all-one sequence (0xFF) to the same address space. This stage lasts about 6 seconds.

[0172] (4) Fourth overwrite stage: Write a random number sequence to the same address space again, with a different random number seed than in the second overwrite stage. This stage lasts about 6 seconds.

[0173] (5) Memory reset phase: A memory reset signal is sent to the GPU to reinitialize the address mapping of HBM3 and clear the internal cache and address translation table of the memory controller. This phase lasts about 0.5 seconds.

[0174] The execution time of the erase operation is independent of the operating system scheduling and is triggered by the independent hardware clock of the physical-level forced erase controller. The independent hardware clock is a 32.768kHz real-time clock crystal oscillator, which is completely isolated from the system master clock. Even if the system master clock is interfered with or stopped by an attacker, the erase controller can still perform the erase operation according to the preset timing.

[0175] After erasing is complete, the device returns to standby mode.

[0176] The above technical solution achieves the following results: By employing a deep erase mode (all zeros + random numbers + all ones + random numbers + memory reset), the system can effectively resist advanced data recovery techniques based on magnetic remanence and cold boot attacks. Research shows that after four overwrites (especially those involving random numbers), even if an attacker physically disassembles the memory chip and reads the residual charge at low temperatures, the data recovery success rate is less than 10^-6. Triggering the erase operation with an independent hardware clock ensures the deterministic timing and unbypassable nature of the erase process; even if an attacker gains the highest operating system privileges, they cannot prevent or delay the erase operation by modifying the system clock. In high-security government data exchange scenarios, the deep erase mode complies with the "Information Security Technology: Guidelines for De-identification of Personal Information" (GB / T). The highest level of requirements for sensitive data destruction in 37964-2019 provides data subjects with physical-level data destruction protection that meets national standards; through the output of 256-dimensional demographic feature vectors, external statistical research institutions can conduct population trend analysis based on aggregated features without obtaining the original household registration information of individual citizens, thus achieving a balance between data value utilization and personal privacy protection, and achieving the technical effect of ensuring that the original data never leaves the domain.

[0177] This method first receives raw plaintext data input from within the data subject through a physically isolated first network interface, establishing a data channel. Simultaneously, it receives call requests and credentials from external users through a physically isolated second network interface, establishing a control channel. The second network interface embeds a low-level hardware offloading mechanism, diverting externally received control commands to the control channel, thus physically isolating the control channel from the data channel. This physical isolation architecture fundamentally blocks the direct data exchange path between the control channel and the data channel, ensuring that external users cannot use control commands to infiltrate the data channel and steal raw plaintext data. It also prevents the leakage of highly sensitive raw plaintext data in the data channel due to network attacks on the control channel. The first and second network interfaces are deployed independently at the hardware level, each with its own independent physical layer transceiver module and link layer processing unit. There is no shared bus or switching matrix, thus completely separating the data channel and control channel at the physical transmission level.

[0178] The dynamic policy gateway deployed in the control channel receives a call request from an external user via a second network interface. This call request includes identity credentials. The dynamic policy gateway verifies the identity credentials, performing a full match between the credentials and the local access policy. Upon successful verification, the dynamic policy gateway issues a one-way input scheduling command for source data to the data source scheduling system, and simultaneously issues a hardware unlock command to the stateless AI sandbox deployed in the data channel. This coordinated issuance of scheduling and unlock commands enables indirect scheduling of the data channel by the control channel, rather than direct intervention in the data processing flow, thus maintaining the physical isolation boundary between the two channels while ensuring scheduling flexibility. The identity credential verification process is completed independently within the control channel, without touching any original plaintext data in the data channel, further reducing the risk of sensitive data being exposed during the authentication process.

[0179] The data source scheduling system unidirectionally inputs raw plaintext data to the stateless AI sandbox via the first network interface. The raw plaintext data is written to the shared memory input segment within the stateless AI sandbox. The stateless AI sandbox physically integrates a general-purpose processor, a heterogeneous accelerated computing unit, and a national cryptographic card. The heterogeneous accelerated computing unit bypasses the operating system protocol stack to directly read the raw plaintext data from the heterogeneous shared memory, loads the AI ​​model for intelligent inference computation, generates a dimensionality-reduced AI feature vector, and writes the feature results to the shared memory output segment. The national cryptographic card performs hardware encryption and digital signature on the AI ​​feature vector to obtain the encrypted AI feature vector. By bypassing the operating system protocol stack's direct memory read mechanism, the audit and caching risks of data access at the operating system level are eliminated, ensuring that the raw plaintext data remains within the trusted hardware execution environment during inference and is not accessed or retained by the operating system or other software processes. The heterogeneous accelerated computing unit communicates with the general-purpose processor via an on-chip high-speed interconnect bus, avoiding data exposure during transmission on the motherboard bus. AI-driven feature-based data transfer breaks the limitations of traditional gateways that only perform encrypted transmission, outputting irreversible intelligent feature vectors and achieving a qualitative change in data characteristics the instant it leaves the domain.

[0180] The encrypted AI feature vector is unidirectionally output to the external data user via the underlying hardware transmission link of the second network interface. This unidirectional output link restricts the data flow at the hardware level, preventing the external data user from injecting data or instructions in reverse through this link. The hardware implementation of the unidirectional output link uses a unidirectional optocoupler or a unidirectional network card to physically cut off the possibility of reverse data transmission, ensuring that data can only flow from the stateless AI sandbox in hardware to the external data user, and cannot flow back in reverse.

[0181] At the instant that a single AI inference task is completed and the encrypted AI feature vector is fully output through the second network interface, the hardware stateless AI sandbox triggers a pulse to the physical-level forced erase controller to indicate that the encrypted feature results have been fully output. The physical-level forced erase controller forcibly takes over and sends a hardware-level clear and overwrite signal to the address space of the underlying volatile physical memory of the hardware stateless AI sandbox, performing a write-reset physical-level erase operation. This forced erase mechanism takes effect immediately after the end of a single inference task, without relying on software-level memory release logic, thereby completely eliminating the risk of residual original plaintext data and intermediate calculation results in volatile physical memory, ensuring that the hardware stateless AI sandbox is in a trustworthy state with no data residue before and after each task execution. The hardware-level clear and overwrite signal acts directly on the physical pins of the memory controller, bypassing the address mapping of the memory management unit, and performs bit-by-bit overwriting on the entire physical address space. The overwriting mode adopts a multi-round overwriting strategy that alternates between random number filling and all-zero filling to combat potential cold start attacks and data recovery methods. By introducing hardware-level forced physical erasure that is strongly tied to the lifecycle of AI inference tasks, the device is completely empty when power is off or idle, ensuring that the original physical data never leaves the domain.

[0182] After the physical-level forced erase controller completes the overwrite reset, it sends an overwrite reset completion signal to the dynamic policy gateway, causing the device to return to stateless standby, awaiting the next call request. The feedback path of the overwrite reset completion signal is independent of the data channel, and is transmitted through a dedicated signaling link within the control channel, ensuring that the confirmation of the erase status does not introduce new data leakage channels.

[0183] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

[0184] Throughout this description, numerous specific details, such as examples of components and / or methods, are provided to provide a complete understanding of embodiments of the invention. However, those skilled in the art will recognize that embodiments of the invention may be practiced without one or more of these specific details or by other devices, systems, components, methods, parts, materials, components, etc. In other instances, well-known structures, materials, or operations have not been specifically shown or described in detail to avoid obscuring aspects of embodiments of the invention.

[0185] Throughout this specification, references to "an embodiment," "an embodiment," or "a specific embodiment" mean that a particular feature, structure, or characteristic described in connection with an embodiment is included in at least one embodiment of the invention, but not necessarily in all embodiments. Therefore, the various representations of the phrases "in one embodiment," "in an embodiment," or "in a specific embodiment" in different places throughout the specification do not necessarily refer to the same embodiment. Furthermore, a particular feature, structure, or characteristic of any specific embodiment of the invention can be combined with one or more other embodiments in any suitable manner. It should be understood that other variations and modifications of the invention and the embodiments shown herein may be based on the teachings herein and will be considered part of the spirit and scope of the invention.

[0186] It should also be understood that one or more of the elements shown in the figures may be implemented in a more separate or more integrated manner, or may even be removed because they are inoperable in certain circumstances or provided because they may be useful for a particular application.

[0187] Furthermore, unless otherwise expressly stated, any arrows in the accompanying drawings should be considered illustrative only and not limiting. Additionally, unless otherwise stated, the term "or" as used herein is generally intended to mean "and / or". Where a term is anticipated to provide a separation or combination capability that is unclear, a combination of components or steps will also be considered as indicated.

[0188] As used herein and throughout the claims below, unless otherwise specified, “a” and “the” include the plural references. Similarly, as used herein and throughout the claims below, unless otherwise specified, “in” means “in” and “on”.

[0189] The above description of the embodiments shown in this invention (including the content in the specification summary) is not intended to be exhaustive or to limit the invention to the precise forms disclosed herein. Although specific embodiments and examples of the invention have been described herein for illustrative purposes only, various equivalent modifications are possible within the spirit and scope of the invention, as will be recognized and understood by those skilled in the art. As indicated, these modifications can be made to the invention in accordance with the above description of the embodiments, and such modifications will be within the spirit and scope of the invention.

[0190] This document has generally described the systems and methods in detail to aid in understanding the invention. Furthermore, various specific details have been set forth to provide a general understanding of embodiments of the invention. However, those skilled in the art will recognize that embodiments of the invention can be practiced without one or more specific details, or using other means, systems, accessories, methods, components, materials, parts, etc. In other instances, well-known structures, materials, and / or operations have not been specifically shown or described in detail to avoid obscuring aspects of embodiments of the invention.

[0191] Therefore, although the invention has been described herein with reference to specific embodiments thereof, freedom of modification, various changes and substitutions are also within the scope of the foregoing disclosure, and it should be understood that in some cases, certain features of the invention may be adopted without departing from the scope and spirit of the invention and without corresponding use of other features. Thus, many modifications can be made to adapt a particular environment or material to the essential scope and spirit of the invention. The invention is not intended to be limited to the specific terminology used in the following claims and / or the specific embodiments disclosed as the best mode for carrying out the invention, but the invention will include any and all embodiments and equivalents falling within the scope of the appended claims. Therefore, the scope of the invention will be defined only by the appended claims.

Claims

1. A hardware sandboxing connection device, characterized in that, include: The dual-channel isolation architecture includes a physically isolated first network interface and a second network interface. The first network interface is used to build a data channel to receive raw plaintext data in one direction. The second network interface is used to build a control channel to receive call requests and identity credentials. The second network interface has an embedded underlying hardware diversion mechanism to divert externally received control commands to the control channel, so that the control channel and the data channel are physically isolated. The dynamic policy gateway is deployed on the control channel to receive and verify externally transmitted identity credentials. After verifying that the credentials match the local access policy, it sends hardware unlocking instructions and one-way input scheduling instructions for source data to the data channel. The hardware stateless AI sandbox module is deployed in the data channel and physically integrates a general-purpose processor, a heterogeneous accelerated computing unit, and a national cryptographic card. The heterogeneous accelerated computing unit is used to bypass the operating system protocol stack, directly read raw plaintext data from heterogeneous shared memory, and perform AI inference to generate a dimensionality-reduced AI feature vector. The national cryptographic card is used to perform hardware encryption and digital signature on the AI ​​feature vector. The physical-level forced erase controller is physically connected to the underlying volatile physical memory of the stateless AI sandbox. After a single AI inference task is completed and the dense feature results are fully output, it sends a hardware-level clear and overwrite signal to the address space of the physical memory to perform a physical-level erase operation of overwrite-reset, so that the device returns to stateless standby. The encrypted output device is used to unidirectionally output the encrypted AI feature vector to an external user via the underlying hardware transmission link of the second network interface.

2. The system according to claim 1, characterized in that, When verifying identity credentials, the dynamic policy gateway uses a multi-factor authentication mechanism, which includes at least two of the following: API Token verification, smart contract address matching, and hardware fingerprint binding. The hardware unlock command is only issued after all authentication factors have passed verification.

3. The system according to claim 1, characterized in that, When the heterogeneous accelerated computing unit performs AI inference, the loaded AI model includes at least one of the following: graph neural network node aggregation model, large model semantic extraction model, and multimodal feature fusion model; the dimension of the dimensionality-reduced AI feature vector is lower than the feature dimension of the original plaintext data, and the AI ​​feature vector can be irreversibly restored to the original plaintext data.

4. The system according to claim 1, characterized in that, The physical level erase operation performed by the physical level forced erase controller includes at least one of the following: writing an all-zero sequence to the address space of the underlying volatile physical memory, writing a random number sequence, and executing a memory reset signal; the execution time of the erase operation is independent of the operating system scheduling and is triggered by the independent hardware clock of the physical level forced erase controller.

5. The system according to claim 1, characterized in that, The system also includes a low-level hardware audit and evidence storage module, which is physically integrated inside the national cryptographic card and operates as an independent security hardware module. It is used for gating and unlocking of the control channel, single inference and physical erasure operations of the data channel, bypassing the host operating system. The security chip directly generates physical logs with high-precision timestamps and signs them with the hardware low-level private key before outputting them to the monitoring node.

6. The system according to claim 5, characterized in that, The physical transaction log generated by the underlying hardware audit and evidence storage module includes: operation type identifier, operation timestamp, operation subject identifier, operation result status, and hash values ​​of the previous and subsequent states. The physical transaction log adopts an immutable chain storage structure, and the hash value of each transaction log record is used as the input parameter for the next transaction log record, forming a hardware-level audit evidence chain.

7. A method for distributing hardware sandboxed connection equipment, characterized in that, Includes the following steps: Step 1: Receive raw plaintext data input from the data subject through the physically isolated first network interface to establish a data channel; Receive call requests and identity credentials initiated by external data users through the physically isolated second network interface to establish a control channel. The second network interface has an embedded underlying hardware offloading mechanism that diverts externally received control commands to the control channel, thus physically isolating the control channel from the data channel. Step 2: The dynamic policy gateway deployed on the control channel receives the call request sent by the external user through the second network interface. The call request includes identity credentials. Step 3: The dynamic policy gateway verifies the identity credential and performs a full match between the identity credential and the local access policy; Step 4: After successful verification, the dynamic policy gateway sends a one-way input scheduling instruction for source data to the data source scheduling system. Step 5: The dynamic policy gateway sends a hardware unlock command to the stateless AI sandbox deployed on the data channel. Step Six: The data source scheduling system unidirectionally inputs raw plaintext data to the stateless AI sandbox via the first network interface; the raw plaintext data is written into the shared memory input segment within the stateless AI sandbox; the stateless AI sandbox physically integrates a general-purpose processor, a heterogeneous accelerated computing unit, and a national cryptographic card; the heterogeneous accelerated computing unit bypasses the operating system protocol stack and directly reads the raw plaintext data from the heterogeneous shared memory, loads the AI ​​model for intelligent inference calculation, generates a dimensionality-reduced AI feature vector, and writes the feature results into the shared memory output segment; the national cryptographic card performs hardware encryption and digital signature on the AI ​​feature vector to obtain the encrypted AI feature vector; Step 7: The encrypted AI feature vector is unidirectionally output to the external data user via the underlying hardware transmission link of the second network interface; Step 8: At the instant that a single AI inference task is completed and the encrypted AI feature vector is fully output through the second network interface, the hardware stateless AI sandbox triggers a pulse to the physical level forced erase controller to indicate that the encrypted feature result has been fully output; the physical level forced erase controller forcibly takes over and sends a hardware level clear and overwrite signal to the address space of the underlying volatile physical memory of the hardware stateless AI sandbox to perform a physical level erase operation of overwrite-reset; Step 9: After the physical level forced erase controller completes the overwrite reset, it sends an overwrite reset completion signal to the dynamic policy gateway, so that the device returns to stateless standby.

Citation Information

Patent Citations

  • Large model corpus data interaction system and interaction method based on trusted execution environment, terminal and medium

    CN120910920A