Data flow tracking and resource release admission method and system based on trusted sandbox

Abstract

The application discloses a data flow tracking and resource release access method and system based on a trusted sandbox, relates to the technical field of data security and trusted computing, and binds data flow tracking and a sandbox life cycle state machine, automatically triggers fingerprint checking and flow tracking recording in an access state, a running state and a pre-destruction state, forms a logical flow tracking map associated with a sandbox measurement state, initiates residual data automatic detection by a trusted management center before the sandbox is destroyed, adopts a multi-stage matching algorithm to comprehensively score, blocks resource release if the detection does not pass, forms a resource release access closed loop, and associates time window and process-level cross verification of network outlet abnormal flow, flow tracking map and vTCM trusted measurement log, and realizes accurate positioning and tracing of leakage behavior. The application cooperatively works through three mechanisms of state-driven tracking, resource release access detection and trusted measurement linkage tracing, and can be used for sensitive data full life cycle security management in a trusted sandbox environment.

Description

Technical Field

[0001] This invention relates to the fields of data security and trusted computing technology, and more specifically to the fields of data flow tracking and resource release access methods and systems based on trusted sandboxes. Background Technology

[0002] With the rapid development of big data and artificial intelligence technologies, the demand for cross-institutional and cross-domain data fusion computing is growing. How to achieve data value mining while ensuring data security and privacy has become a crucial issue in the field of data security. Trusted Execution Environments (TEEs) and sandbox isolation technologies provide important technical support for this endeavor.

[0003] The trusted sandbox environment provides a controlled execution space for sensitive data processing through hardware roots of trust, metric chains, and resource isolation mechanisms. Within the domestic trusted computing technology system, the TCM (Trusted Cryptography Module) chip, based on Chinese cryptographic algorithms, provides cryptographic capabilities such as SM2 signatures, SM3 hashes, and SM4 encryption, constructing a trusted metric chain from hardware to software. TIPU is a trusted processing board integrating TCM roots of trust, network / storage offloading capabilities, and sandbox support capabilities. It can participate in sandbox measurement, data mounting, network egress monitoring, and trusted auditing at the board level, providing a hardware-level security foundation for secure data sharing.

[0004] In the field of data flow tracking, existing technologies primarily record data access behavior through operating system audit logs, file system monitoring, or database auditing. Regarding network anomaly detection, anomaly detection technologies based on traffic baseline analysis, deep packet inspection (DPI), and machine learning methods have been widely researched and applied. For data destruction, mature solutions exist for secure deletion and multiple overwrite techniques.

[0005] Despite significant advancements in data security technology, the trusted sandbox scenario still exhibits three notable shortcomings: Defect 1: Data flow tracking is disconnected from sandbox operation status. Data flow tracing in existing sandbox environments typically employs a separate auditing module, which can only record discrete operation event logs (such as "a process read a file at a certain time" or "a process sent data to a certain IP address"). While these logs contain timestamps and operation entity information, they cannot correlate data flow behavior with the lifecycle state of the sandbox operation (such as whether the data is in the access, processing, or pre-destruction stage). When the sandbox is running, data may be copied, transformed, and merged between different processes. Existing technologies lack a flow tracing mechanism that is automatically triggered based on the sandbox state. As a result, after a data breach occurs, security auditors find it difficult to trace back to the specific operational stage where the breach occurred and the corresponding sandbox trust metric state, leading to severely insufficient efficiency and accuracy in tracing the source.

[0006] Deficiency 2: Lack of residual data detection and access control in sandbox destruction: When a sandbox job completes or expires and requires resource release, data copies may remain in multiple storage areas accessible to the sandbox, including but not limited to: memory snapshots (such as core dump files generated when a process crashes), disk swap partitions (data written by the operating system during memory swapping), temporary directories (such as / tmp, / var / tmp), various log files (application logs, system logs), and mounted data directories. Existing sandbox systems typically only perform simple file deletion commands (such as rm -rf) for data destruction, lacking automated residual detection processes before destruction, and lacking an access control mechanism that "blocks resource release if detection fails." This means that even if copies of sensitive data remain in the sandbox memory and storage, the system may directly release resources, reallocating storage blocks containing residual data to other tenants, posing a serious risk of data leakage.

[0007] Deficiency 3: Lack of linkage and tracing capabilities between leakage behavior and trust measurement status: While trusted sandboxes block conventional file copying and network transmission channels through access control policies, attackers can still use covert channels (such as DNS tunneling, ICMP payload carrying, and covert data transmission via HTTP / S protocol) to smuggle data out. In existing solutions, network egress detection and the trust measurement within the sandbox are typically two separate systems: the network detection system can only detect abnormal traffic but cannot determine which processes within the sandbox have accessed sensitive data or the trust measurement status of these processes; the sandbox auditing system records process behavior but lacks the ability to correlate it with network-side data. This separation makes it difficult to accurately pinpoint which process, during which time window, and through what method the data was leaked, even when abnormal outbound traffic is detected, resulting in extremely high tracing costs. Summary of the Invention

[0008] The purpose of this invention is to address the problems in existing technologies such as the disconnect between data flow and operational status, the lack of residue detection and access control before destruction, and the difficulty in accurately locating leakage behavior. This invention provides a data flow tracking and resource release access method and system based on a trusted sandbox.

[0009] To achieve the above objectives, the present invention specifically adopts the following technical solution: The first aspect of this invention provides a data flow tracking and resource release access method based on a trusted sandbox, comprising the following steps: By binding data flow tracking with the sandbox lifecycle state machine, the recording of data flow events within the sandbox is automatically triggered in the sandbox's access state, running state, and pre-destruction state, forming a logical flow graph associated with the sandbox state. Before the sandbox is destroyed, the Trusted Management Center initiates residual data detection and determines whether sandbox resources can be released based on the detection results, thus forming a closed loop for resource release access. The network egress behavior is continuously monitored in the sandbox running state. When a network anomaly is detected, the abnormal traffic information is linked with the logical flow graph and vTCM trust metric status for analysis to locate the source of the leak.

[0010] In one implementation, the formation of the logical flow graph associated with the sandbox state specifically includes: In the access state, calculate the primary and / or secondary fingerprints of the data entering the sandbox, record the data element information and the current vTCM metric value as the fingerprint verification benchmark, and initialize the flow graph; In runtime, data flow events are collected through kernel-level probes. Each event records a timestamp, operation process, and data flow direction. The flow map is updated incrementally, and fingerprint verification is performed on data copies that remain authorized. The mapping relationship before and after the authorized data conversion is recorded. Before destruction, freeze the business process and stop recording new business process flow events.

[0011] In one implementation, the residual data detection initiated by the trusted management center specifically includes: The Trusted Management Center sends a residual detection command to the sandbox, and the sandbox performs signature verification and integrity verification on the command using a key or certificate protected by vTCM. After the verification is passed, the sandbox freezes the business process and obtains an accessible memory snapshot. Multi-level residual matching scans are then performed on the memory snapshot, disk swap file, temporary directory, log directory, core dump file, and mounted data directory. The multi-level residual matching scan includes: a first-level scan based on seed fragment reverse matching, a second-level scan based on feature matching, and a third-level scan based on semantic matching, and a comprehensive risk score is calculated according to a weighted scoring rule.

[0012] In one implementation, determining whether to allow the release of sandbox resources based on the detection results specifically includes: When the comprehensive risk score reaches the high-risk threshold, resource release or reuse is blocked, triggering a forced overwrite / erase or isolation / retention process. When the overall risk score is medium risk, the decision shall be made by the safety administrator after review. When the overall risk score is low, the release of sandbox resources is permitted through destruction approval.

[0013] In one implementation, the linkage analysis to locate the leakage source specifically includes: Centered on the timestamp of network anomaly detection, data flow events are retrieved from the logical flow graph within a preset time window; Extract a list of processes that have accessed sensitive data from the search results, and filter out irrelevant processes by combining process lifecycle, parent-child process relationship, file handle and network connection records; The runtime of the filtered suspicious processes is compared with the vTCM metric log. If unauthorized code is loaded or the metric value changes abnormally during the process's runtime, a source tracing report is output.

[0014] Two aspects of the present invention provide a data flow tracking and resource release access system based on a trusted sandbox, including a data flow tracking module, a data residue detection and resource release access module, a leakage alarm and trusted measurement linkage tracing module, and an audit report generation module; The data flow tracking module is used to bind to the sandbox lifecycle state machine to collect data fingerprints, behavioral events, process contexts, and vTCM metric states in the access state, running state, and pre-destruction state, forming a logical flow graph. The Data Residual Detection and Resource Release Admission Module is used to receive detection instructions from the Trusted Management Center before destruction, perform residual data scanning, and determine whether sandbox resources can be released based on the scanning results. The leakage alarm and trusted measurement linkage tracing module is used to monitor network egress traffic. When an anomaly is detected, it cross-verifies the abnormal traffic timestamp, the process list in the flow graph, and the vTCM measurement log to locate the leakage source. The audit report generation module is used to summarize circulation tracking records, residual detection reports and source tracing results, and generate an audit report that is stored after being signed with national cryptographic standards.

[0015] In one embodiment, the data residue detection and resource release access module includes an instruction signature verification unit, a process freezing unit, a multi-level scanning unit, and an access determination unit: The instruction verification unit is used to verify the signature and integrity of the detection instructions from the Trusted Management Center using keys or certificates protected by vTCM. The process freezing unit is used to freeze business processes within the sandbox through the operating system's process control mechanism, while allowing the residual detection agent and audit agent processes to continue running. The multi-level scanning unit is used to perform multi-level residual matching scans on sandbox-accessible memory snapshots, disk swap files, temporary directories, log directories, core dump files, and mounted data directories; The access determination unit is used to perform resource release access determination based on the comprehensive risk score from multi-level scanning.

[0016] In one implementation, the leakage alarm and trust measurement linkage tracing module includes: The network monitoring unit is used to perform payload feature matching for plaintext protocols and metadata feature analysis for encrypted protocols. The time window backtracking unit is used to retrieve data flow events within a preset time window in the flow graph, centered on the abnormal timestamp; The process filtering unit is used to extract a list of processes that have accessed sensitive data from the search results, and filter irrelevant processes by combining process lifecycle and parent-child process relationship; It also includes a metric cross-validation unit, which compares the runtime of a suspicious process with the vTCM metric log to identify abnormal changes in metric values ​​and output a source tracing report.

[0017] In one embodiment, the data flow tracking module includes: The fingerprint generation unit is used to generate a first-level fingerprint from structured data using a cryptographic hash algorithm and to generate a second-level feature vector from unstructured data using a feature extraction model. The flow graph construction unit is used to infer logical flow paths and incrementally update the graph based on file handles, process contexts, data fingerprints, and time windows. And a metric state reading unit, used to read the current vTCM trusted metric value and associate it with the flow event record during each state transition.

[0018] In one implementation, the audit report generation module includes a signature storage unit; The signature and evidence storage unit is used to digitally sign the audit report using the SM2 national cryptographic algorithm and store the signed report in a secure database or blockchain evidence storage platform.

[0019] Working principle: In terms of system architecture, this solution constructs a system architecture that includes four core modules: a data flow tracking module, a data residue detection and resource release access module before destruction, a leakage alarm and trust measurement linkage tracing module, and an audit report generation module. Each module works collaboratively driven by the sandbox lifecycle status.

[0020] In terms of methodology, this invention uses the four lifecycle states of the TIPU Trusted Sandbox (access state, running state, pre-destruction state, and destruction state) as the driving force, and performs corresponding data collection, detection, and analysis operations at each state transition node to form a complete security closed loop.

[0021] Regarding key algorithms, this invention designs a data fingerprint generation algorithm (first / secondary fingerprint), a multi-level matching and comprehensive scoring algorithm for data residues (seed fragment reverse matching + feature matching + semantic matching), and a leakage tracing algorithm based on time windows and metric cross-validation.

[0022] This solution provides a data flow tracking mechanism based on the sandbox lifecycle state machine. During the state transitions between the access state, running state, and pre-destruction state, fingerprint verification and flow records are automatically triggered, so that each flow record is associated with the current state and trust metric of the sandbox, thus solving the problem of flow tracking being disconnected from the job status.

[0023] This solution provides a pre-destruction residual detection process initiated by an external trusted management center and transmitted through the vTCM protection channel. It employs a multi-level matching algorithm to scan all storage areas accessible by the sandbox, and the detection results serve as the sole criterion for resource release access, thus addressing the security vulnerabilities of no detection and no access control before destruction.

[0024] This solution provides an analysis mechanism that links network anomalies, process behavior, and vTCM metric status. By using time window backtracking, process lifecycle filtering, and metric status cross-validation, it achieves precise process-level location of the leak source, solving the problem of difficulty in leak tracing.

[0025] The beneficial effects of this invention are as follows: 1. State-Driven End-to-End Flow Tracing: This invention binds data flow tracing to the sandbox lifecycle state machine, recording the current sandbox state and vTCM metric value for each flow event, thus solving the problem of disconnect between flow tracing and job state in existing technologies. Different data collection and fingerprint verification strategies are automatically triggered in the access state, running state, and pre-destruction state, achieving full lifecycle tracing from data entering the sandbox to resource release. In the event of a data breach, security auditors can accurately trace back to which job stage in the sandbox the breach occurred, what the sandbox's trust metric status was at the time, and which processes had accessed the sensitive data, significantly improving tracing efficiency and accuracy. Compared to the discrete audit logs in existing solutions, the flow graph of this invention provides a structured view of data lineage, making the data flow path readily apparent.

[0026] 2. Closed-Loop Security Effect of Resource Release Access Control: This invention constructs a closed-loop mechanism for residual detection: "Detection → Scoring → Judgment → Disposal." An external trusted management center initiates detection commands through the vTCM protection channel, ensuring the authority of the detection process and the security of command transmission. Residual detection is forcibly executed before sandbox resource release; if data residuals are detected, resource release or reuse is blocked, fundamentally eliminating the risk of cross-tenant data leakage due to data residuals. Compared to existing technologies that only execute simple deletion commands, this invention's resource release access mechanism elevates the security level of data destruction from "best-effort" to "mandatory protection." Simultaneously, the multi-level residual matching algorithm balances the comprehensiveness and accuracy of detection: the first-level scan ensures zero false negatives for accurate residuals, while the second and third-level scans cover derived residuals after format conversion and semantic transformation. The comprehensive scoring mechanism makes the detection results quantifiable and comparable.

[0027] 3. Precise Source Tracing Effect Through Trusted Measurement Linkage: This invention links network anomaly detection, sandbox internal flow graph, and vTCM trusted measurement status, realizing a closed-loop source tracing link of "network anomaly → process location → measurement verification." Compared to the existing solutions where network detection and sandbox auditing operate in isolation and source tracing relies on manual investigation, this invention uses three automated analysis steps—time window backtracking, process lifecycle filtering, and measurement status cross-verification—to improve source tracing accuracy from "IP address level" to "process level" and shorten source tracing time from "hours" to "minutes." In particular, by using vTCM measurement logs as a key verification dimension, it can identify whether there are unauthorized code loading, configuration changes, or other abnormal behaviors during process execution, providing strong evidence for leakage events. This allows security administrators to determine leakage responsibility based on objective measurement evidence rather than subjective inference.

[0028] 4. Synergistic Effect of Module Linkage: The three core modules of this invention do not operate independently, but form a synergistic organic whole. The residual detection module uses the original data fingerprint registered in the access state as the residual detection benchmark; the leakage alarm and trust measurement linkage tracing module uses the data flow relationship and trust measurement status in the logical flow graph to narrow down the scope of investigation of suspicious processes and data copies; the audit report generation module uniformly summarizes and reliably stores the flow tracking records, residual detection results and tracing results, forming a full lifecycle security closed loop from data access, operation processing, pre-destruction detection to resource release. Attached Figure Description

[0029] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0030] Figure 1 This is a system architecture diagram of a data flow tracking and resource release access system based on a trusted sandbox.

[0031] Figure 2 This is a flowchart of the sandbox lifecycle status and tracking process.

[0032] Figure 3 This is a logic diagram for residual detection and resource release access.

[0033] Figure 4 This is a traceability diagram linking leak alerts and trust metrics. Detailed Implementation

[0034] To make the technical problems, technical solutions, and technical effects of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0035] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0036] like Figures 1 to 4As shown, one aspect of the present invention provides a data flow tracking and resource release access system based on a trusted sandbox, including a data flow tracking module, a data residue detection and resource release access module, a leakage alarm and trusted measurement linkage tracing module, and an audit report generation module; Module 1: Data Flow Tracking Module; This module is bound to the sandbox lifecycle state machine to collect data fingerprints, behavioral events, process context, and vTCM metric status in the access state, running state, and pre-destruction state, respectively, forming a logical flow graph. Flow tracing can be implemented using an eBPF (Extended Berkeley Packet Filter) probe or a kernel audit module to reduce the impact on sandbox service performance. The flow graph is a graphical representation of the logical flow path inferred based on file handles, process context, data fingerprints, and time windows. Nodes represent the state of data at a specific processing position, and edges represent the data flow direction and operation type.

[0037] Module 2: Data Residual Detection and Resource Release Access Module Before Destruction; This module is activated when the sandbox enters the pre-destruction state, and an external trusted management center sends a residual detection command to the TIPU sandbox. The sandbox verifies the signature and integrity of the command using a vTCM-protected key or certificate, ensuring that the command origin is trustworthy and has not been tampered with. After successful verification, the sandbox freezes the business processes within the sandbox through the operating system process control mechanism (while preserving the residual detection agent, audit agent, or management channel processes), obtains an accessible memory snapshot, and performs a comprehensive scan of the memory snapshot, disk swap files, temporary directories ( / tmp, / var / tmp), system log directory, coredump files, and mounted data directories. The scan employs a multi-level matching algorithm: seed fragment reverse matching, feature matching, and semantic matching combined scoring, generating a residual detection report that is returned to the trusted management center. The trusted management center performs resource release access determination based on the comprehensive risk score in the report.

[0038] Module 3: Leakage Alarm and Trust Measurement Linkage and Source Tracing Module This module continuously monitors the outbound traffic of the virtual network interface card (NIC) on the TIPU board during sandbox operation. Network anomaly detection performs payload feature matching for plaintext protocols (such as HTTP) and metadata feature analysis (including but not limited to TLS handshake features, certificate information, traffic direction, packet length distribution, and time series features) for encrypted protocols such as HTTPS. When an anomaly is detected, the module cross-validates the abnormal traffic timestamp, the process list in the flow graph, and the vTCM metric log, and then executes: ① Time window backtracking centered on abnormal timestamps; ② Extraction of the list of processes that have accessed sensitive data and filtering of process lifecycles; ③ By comparing and verifying the running time of suspicious processes with vTCM measurement logs, the suspicious processes, loaded files, configuration items, or execution paths related to abnormal measurement status changes can be located.

[0039] Module 4, Audit Report Generation Module: This module aggregates circulation tracking records, residual detection reports, and leakage tracing results to generate an audit report containing complete circulation records, residual detection results, and tracing results. The audit report is stored after being signed with SM2 national cryptographic standards. Optional storage methods include secure databases or consortium blockchain storage, ensuring the report's authenticity, completeness, and non-repudiation.

[0040] Another aspect of the present invention provides a data flow tracking and resource release access method based on a trusted sandbox. This method is implemented based on the TIPU trusted sandbox platform and includes the following steps: By binding data flow tracking with the sandbox lifecycle state machine, the recording of data flow events within the sandbox is automatically triggered in the sandbox's access state, running state, and pre-destruction state, forming a logical flow graph associated with the sandbox state. Before the sandbox is destroyed, the Trusted Management Center initiates residual data detection and determines whether sandbox resources can be released based on the detection results, thus forming a closed loop for resource release access. The network egress behavior is continuously monitored in the sandbox running state. When a network anomaly is detected, the abnormal traffic information is linked with the logical flow graph and vTCM trust metric status for analysis to locate the source of the leak.

[0041] This solution is driven by the lifecycle states of the TIPU Trusted Sandbox, dividing data processing into four states. The transitions between these states are triggered by job scheduling events, user commands, or commands from the Trusted Management Center. State 1, Access State (Data enters the sandbox); Data providers access the TIPU Trusted Sandbox via upload, mounting, authorization mapping, or secure channels. The system performs the following operations in the access state: (1) Calculate data fingerprint: For structured data (such as database tables, CSV files, etc.), the cryptographic hash algorithm SM3 is used to generate a 256-bit primary fingerprint (precise identifier); for unstructured data (such as images, audio and video, text documents, etc.), in addition to generating the SM3 primary fingerprint, a feature extraction model (such as ResNet, ViT, BERT, etc., which can be selected according to the business scenario) is used to generate a fixed-dimensional secondary feature vector (fuzzy identifier) ​​for subsequent fuzzy matching.

[0042] (2) Record data element information: data type, data volume, sensitivity level, data party authorization status and authorization scope.

[0043] (3) Read the current vTCM trusted measurement value, including but not limited to PCR (Platform Configuration Register) value, measurement log summary, etc., as a trusted benchmark for subsequent comparison.

[0044] (4) Initialize the flow graph, create the root node to mark the sandbox state when the data is accessed as "access state", and record the access timestamp and initial metric value.

[0045] State 2, Running State (Data is processed within the sandbox): After the sandbox enters the running state, the system continues to perform the following operations: (1) Collect data flow events through eBPF probe or kernel audit module. Record the timestamp, process PID (ProcessIdentifier), process name, data flow direction (read / write / copy / convert / network send, etc.) and current sandbox metric value for each event.

[0046] (2) Incremental update of the flow graph: Each flow event adds a node (representing a new data processing location) or edge (representing the flow direction and operation type) to the graph. The node contains data fingerprint, process information and measurement status.

[0047] (3) Continuous monitoring of network outbound traffic, performing baseline comparison (comparing with historical normal traffic patterns), plaintext protocol payload feature matching (detecting sensitive data features), and encrypted protocol metadata analysis.

[0048] (4) Periodically perform fingerprint verification on data copies that remain unchanged under authorization (recalculate the SM3 hash and compare it with the access state baseline) to verify data consistency and integrity; record the mapping relationship or derived fingerprint before and after the authorization transformation of data (such as desensitization, format conversion, aggregation calculation) to ensure that the data flow is traceable.

[0049] State 3, Pre-destruction state (operation expires or destruction is actively triggered); When the job expires or the Trusted Management Center issues a destruction command, the sandbox enters the pre-destruction state: (1) Freeze the business flow graph and stop recording new flow events in the business process (the operations generated by the residual detection process itself are recorded separately as detection audit events and stored separately from business flow events).

[0050] (2) The Trusted Management Center sends a residual detection instruction to the sandbox (the instruction includes the detection range and parameters). The sandbox uses the key or certificate protected by vTCM to sign and verify the integrity of the instruction. If the verification fails, it refuses to execute and returns an error to the management center.

[0051] (3) After the verification is passed, the business processes in the sandbox are frozen through the operating system process control mechanism (such as cgroup freezing or SIGSTOP signal), while the residual detection agent, audit agent or management channel processes are kept running to prevent the business processes from continuing to generate new data copies or interfering with the detection results during the detection process. Then, the memory snapshot that can be accessed by the current sandbox is obtained.

[0052] (4) Perform multi-level residual matching scans on the following areas: memory snapshots, disk swap partition files, / tmp and / var / tmp temporary directories, system log directories, coredump files, mounted data directories and other sandbox-accessible storage paths.

[0053] (5) Calculate the comprehensive risk score based on the scanning results, generate a detection report and return it to the Trust Management Center.

[0054] (6) The Trusted Management Center makes a resource release access determination based on the test report.

[0055] State 4, Destruction State (Resource Release): After resource release access is approved: (1) Summarize the full flow records from the access state to the destruction state, residual detection reports, and network monitoring alarms to form a complete security audit dataset.

[0056] (2) Generate an audit report, which includes a snapshot of the flow map, residual detection results and source tracing conclusions, and stores it after being signed by SM2.

[0057] (3) Perform sandbox resource cleanup, including memory release, disk space reclamation, file system unloading, etc., and return the computing resources occupied by the sandbox to the resource pool.

[0058] The key algorithms are detailed below: (I) Data fingerprint generation algorithm: Primary fingerprint (precise identification): A 256-bit digest value is calculated for sensitive data using the SM3 cryptographic hash algorithm. For file-type data, the SM3 hash of the complete file content is calculated; for database table data, the serialized SM3 hash of the entire table data is calculated. The primary fingerprint is used to accurately identify data copies and detect data tampering.

[0059] Secondary fingerprint (fuzzy identifier, only for unstructured data): A fixed-dimensional feature vector is extracted using a feature extraction model. Specifically, for image data, ResNet-50 can be used to extract a 2048-dimensional feature vector; for text data, the BERT model can be used to extract a 768-dimensional feature vector. In actual deployment, the appropriate model architecture can be selected based on the data type of the business. The secondary fingerprint uses cosine similarity for fuzzy matching to identify derived copies that still contain the semantics of the original data after format conversion, compression, partial truncation, etc.

[0060] (II) Data Residual Multi-Level Matching and Resource Release Admission Scoring Algorithm: This scheme designs a three-level progressive residual matching algorithm to control computational overhead while ensuring detection accuracy: Level 1 Scan (Seed Fragment Reverse Matching): The original sensitive data samples registered in the access state, or their preset sampling segments, are divided into a fixed-length set of seed segments of N=512 bytes (the actual length can be configured according to data type and detection strategy). The efficient Boyer-Moore-Horspool string matching algorithm is used to perform location scanning in all scan areas accessible within the sandbox. The SM3 hash is recalculated for each matched contiguous region and compared with the original first-level fingerprint. If they match, it is considered an exact residual match, with a first-level score S1=1.0; otherwise, S1=0. The first-level scan has a zero false alarm rate and can accurately identify unmodified original data copy residuals.

[0061] Secondary scanning (fuzzy feature matching): For suspicious regions discovered by the primary scan (i.e., regions that match seed fragments but have inconsistent hash comparisons), their secondary feature vectors are extracted. Cosine similarity is calculated between these vectors and the reference fingerprint (secondary feature vectors registered in the access state), and normalized to obtain a secondary score S2 (range 0~1). The closer the S2 value is to 1, the higher the semantic similarity between the region and the original data, and the more likely it is to be a derived residual that has undergone format conversion or partial modification.

[0062] Level 3 scanning (semantic matching): For text fragments, structured fields, log fragments, or parsable content in the scanned area, a predefined set of sensitive data feature rules is used for matching. This rule set includes, but is not limited to: regular expressions for ID card numbers, mobile phone numbers, bank card numbers, financial transaction code patterns, and medical diagnosis code patterns. The more rules matched, the higher the likelihood that the area contains sensitive data semantics. The number of matched rules is normalized to obtain a three-level score S3 (range 0-1).

[0063] Overall rating: The comprehensive risk score is calculated according to the following formula: Score=W1×S1+W2×S2+W3×S3; In a preferred embodiment, the weights are set as W1=0.5, W2=0.3, W3=0.2 (in actual deployment, each weight parameter can be configured according to data sensitivity levels, business scenarios and security policies); This weight configuration reflects the highest credibility of first-level exact matching, while taking into account the supplementary roles of fuzzy matching and semantic matching. In actual deployment, the weights and thresholds can be dynamically configured according to data types, business scenarios and security policies. For example, the weight of W1 can be appropriately increased for highly sensitive data.

[0064] Resource release access judgment:

[0065] (III) Leakage traceability linked analysis algorithm: When network egress monitoring detects abnormal traffic, the following three-step linked analysis is performed: Step 1: Time window backtracking: Centering on the timestamp T when the abnormal traffic is captured, a time window of ΔT=5 minutes is taken before and after the timestamp respectively (it can be dynamically configured according to actual business scenarios), and all data flow events within this window are retrieved in the flow graph. The size of the time window can be adjusted according to business scenarios. For computationally intensive jobs, the window can be appropriately expanded to cover the complete data processing cycle.

[0066] Step 2: Process filtering: Extract the list of PIDs of all processes that have accessed sensitive data within the time window from the retrieval results, and construct a candidate process set. Combining process life cycle information (process start time, end time), parent-child process relationship (pstree), file handle list ( / proc / [pid] / fd) and network connection records ( / proc / [pid] / net), processes that cannot generate external network connection are filtered out to narrow the scope of suspicion.

[0067] Step 3: Cross-validation of measurement status: Align and compare the running time period of the remaining suspicious processes after filtering with the vTCM measurement log on the timeline. If any of the following situations occur during the running of a process, the leakage confidence of the process is significantly increased: ① There is an unauthorized code loading event (for example, the dynamic link library loading fails the measurement verification); ② The key system configuration file is changed but the measurement value is not updated; ③ The PCR measurement value deviates from the expected reference value. For high-confidence processes, further trace back the data copies they operated, the loaded files and the executed command paths.

[0068] The comprehensive three-step output source tracing report includes: the PID and process name of the suspicious process, the data copy identifier of the operation (file path or data table name), the time window of the abnormal event, the corresponding vTCM metric status change record, and the leakage confidence level.

[0069] Example 1 The system boundary conditions and key configuration parameters for this embodiment are as follows:

[0070] Scenario Description: A fintech company has built a joint risk control platform based on the TIPU Trusted Sandbox to serve bank lending business. Data provider A is a credit reporting agency, providing corporate credit data (including basic corporate information, loan records, repayment delinquency status, etc., approximately 500GB); data provider B is an e-commerce platform, providing corporate transaction behavior data (including transaction flow, refund records, merchant ratings, etc., approximately 300GB). The data from both parties is used for joint risk control modeling within the TIPU Trusted Sandbox. After modeling is completed, a risk control scoring model is output. The original data is not allowed to be taken out of the sandbox in any form.

[0071] The implementation process is as follows: (I) Implementation in Access State: Data provider A uploads its credit data to the sandbox via the secure channel of the TIPU board. After entering access mode, the sandbox executes the following: ① The data flow tracking module calculates the SM3 primary fingerprint (precise identification) for each credit data file. At the same time, since the credit data contains a large number of text description fields (such as descriptions of enterprise operating conditions), the BERT model is used to extract 768-dimensional secondary feature vectors.

[0072] ② Record metadata: Data type = structured tabular data, data volume = 500GB, sensitivity level = high (contains personal financial information), authorization scope = limited to risk control modeling use only, data party signature confirmation.

[0073] ③ Read the current vTCM PCR measurement value through the TCM chip and record it as PCR_Baseline, which will serve as the benchmark for subsequent comparisons.

[0074] ④ Initialize the flow graph, create the root node "Credit Data Access", label the status as "Access Status", timestamp as "2026-06-18 10:00:00", and measure as "PCR_Baseline".

[0075] Following the same process, data provider B completes the access registration for its transaction behavior data and initializes another root node.

[0076] (II) Implementation in Operational State: The modeling task has started, and the sandbox has entered runtime. System execution: ① Enable eBPF probes to monitor all file read / write, inter-process communication, and network operations related to credit data and transaction data. Record each data access event, for example: "Process PID=1234 (main modeling process) read the credit data table table_customer at 10:05:23, file handle fd=15, data flow direction=READ, current metric value=PCR_1".

[0077] ② Incremental update of the flow map: Add a sub-node "Modeling process read" under the root node of credit data, and label the operation type as READ, process as PID1234, and time as 10:05:23.

[0078] ③ Network egress monitoring runs continuously, detecting egress traffic from the sandbox virtual network interface card. During normal operation, the sandbox only allows data to be sent to the model result output port (whitelisted port), and the egress traffic model consists of periodic model parameter update packets (approximately 10MB / time).

[0079] ④ Periodic fingerprint verification: During the modeling process, at intermediate checkpoints, the SM3 hash of the credit data file is recalculated and compared with the access state benchmark to confirm that the data has not been tampered with.

[0080] ⑤ When the modeling program performs desensitization processing on the credit data (replacing the ID number with a hash value) to generate derived data, it records the mapping relationship before and after the conversion, and the derived fingerprint is labeled "derived_from=credit data table".

[0081] (III) Implementation of the destruction process: Once the modeling task is complete, the Trusted Management Center issues a task termination and resource release command, and the sandbox enters the pre-destruction state. ① The data flow tracking module freezes the business flow graph and stops recording new flow events in the business process.

[0082] ② The Trusted Management Center sends a residual detection command to the sandbox. The command contains an SM2 signature (signed by the management center's private key). The sandbox verifies the signature using the management center's public key pre-installed in vTCM and checks the certificate validity to confirm that the command's source is trustworthy and has not been tampered with.

[0083] ③ Freeze all business processes (including the modeling main process, data processing process, etc.) using cgroupfreezer, but allow the residual detection agent process and audit agent process to continue running.

[0084] ④ Obtain a memory snapshot of the current sandbox process (read via / proc / [pid] / mem), and simultaneously scan the disk swap partition ( / dev / sda2), / tmp directory, / var / tmp directory, system log / var / log, configured coredump path ( / var / crash), and mounted data directories ( / mnt / data_a and / mnt / data_b).

[0085] ⑤ Perform multi-level residual matching scan: Level 1 Scan: The first 1MB of the credit data file registered in the access state is divided into 512-byte seed segments (approximately 2000 seeds). The Boyer-Moore-Horspool algorithm is used to locate the seeds in each scan area. A temporary feature engineering intermediate file is found in the / tmp directory. Its first 512 bytes completely match the 15th seed segment. Further calculation of the complete SM3 hash of this file is performed and compared with the level 1 fingerprint of the credit data. The hash matches, indicating an exact residual match, S1=1.0.

[0086] Secondary scan: Extract BERT feature vectors from the same file and calculate the cosine similarity with the secondary features of the credit data. The result is 0.98, and S2 = 0.98.

[0087] Level 3 scan: Semantic rule matching was performed on the file, and 32 "ID number regular expressions" and 18 "phone number regular expressions" were found. After normalization, S3=0.85.

[0088] Overall score: Score = 0.5 × 1.0 + 0.3 × 0.98 + 0.2 × 0.85 = 0.964, which is considered high risk (≥ 0.7).

[0089] ⑥ Generate a residual detection report, sign it with SM2, and return it to the Trusted Management Center. The report details the residual file path / tmp / feature_cache_1234.tmp, the residual data type = credit data derived feature, and the comprehensive score = 0.964.

[0090] ⑦ The Trusted Management Center performs access control: it blocks resource release, triggers a forced overwrite and erase process, uses / dev / urandom to generate random data to overwrite / tmp / feature_cache_1234.tmp three times, and records the administrator's review log.

[0091] (iv) Destruction Implementation: After the residual removal is completed, the Trusted Management Center re-initiates the residual detection command. Once it confirms that there are no residuals, it grants access for resource release. The system summarizes the full flow record, residual detection report, and network monitoring logs, generates an audit report, signs it with SM2, stores it on the consortium blockchain evidence storage platform, and then performs sandbox resource cleanup.

[0092] (V) Experimental Data: Test environment: TIPU board (with built-in TCM 2.0 chip), Intel Xeon E5-2680 v4, 64GB memory, Linux Kernel 5.10.

[0093] Test results: Under typical business loads, the performance overhead of the tracing module is within an acceptable range and does not affect the main business operations of the sandbox. The residual detection and linked tracing mechanism can be used to verify the coverage of residual detection and the integrity of the tracing link.

[0094] Specific verifiable metrics include: differences in job completion time with the tracking module enabled / disabled, the ability of residual detection to identify residual fragments in sandbox-accessible memory snapshots / temporary directories / mounted directories / core dumps, and examples of tracing DNS / ICMP covert channels.

Claims

1. A data flow tracking and resource release admission method based on a trusted sandbox, characterized in that, This method is implemented based on the TIPU Trusted Sandbox platform and includes the following steps: By binding data flow tracking with the sandbox lifecycle state machine, the recording of data flow events within the sandbox is automatically triggered in the sandbox's access state, running state, and pre-destruction state, forming a logical flow graph associated with the sandbox state. Before the sandbox is destroyed, the Trusted Management Center initiates residual data detection and determines whether sandbox resources can be released based on the detection results, thus forming a closed loop for resource release access. The network egress behavior is continuously monitored in the sandbox running state. When a network anomaly is detected, the abnormal traffic information is linked with the logical flow graph and vTCM trust metric status for analysis to locate the source of the leak.

2. The data flow tracking and resource release admission method based on a trusted sandbox according to claim 1, characterized in that, The formation of the logical flow graph associated with the sandbox state specifically includes: In the access state, calculate the primary and / or secondary fingerprints of the data entering the sandbox, record the data element information and the current vTCM metric value as the fingerprint verification benchmark, and initialize the flow graph; In runtime, data flow events are collected through kernel-level probes. Each event records a timestamp, operation process, and data flow direction. The flow map is updated incrementally, and fingerprint verification is performed on data copies that remain authorized. The mapping relationship before and after the authorized data conversion is recorded. Before destruction, freeze the business process and stop recording new business process flow events.

3. The data flow tracking and resource release admission method based on a trusted sandbox according to claim 1, characterized in that, The residual data detection initiated by the Trusted Management Center specifically includes: The Trusted Management Center sends a residual detection command to the sandbox, and the sandbox performs signature verification and integrity verification on the command using a key or certificate protected by vTCM. After the verification is passed, the sandbox freezes the business process and obtains an accessible memory snapshot. Multi-level residual matching scans are then performed on the memory snapshot, disk swap file, temporary directory, log directory, core dump file, and mounted data directory. The multi-level residual matching scan includes: a first-level scan based on seed fragment reverse matching, a second-level scan based on feature matching, and a third-level scan based on semantic matching, and a comprehensive risk score is calculated according to a weighted scoring rule.

4. The data flow tracking and resource release admission method based on a trusted sandbox according to claim 3, characterized in that, The process of determining whether to allow the release of sandbox resources based on the detection results specifically includes: When the comprehensive risk score reaches the high-risk threshold, resource release or reuse is blocked, triggering a forced overwrite / erase or isolation / retention process. When the overall risk score is medium risk, the decision shall be made by the safety administrator after review. When the overall risk score is low, the release of sandbox resources is permitted through destruction approval.

5. The data flow tracking and resource release admission method based on a trusted sandbox according to claim 1, characterized in that, The linked analysis to locate the leakage source specifically includes: Centered on the timestamp of network anomaly detection, data flow events are retrieved from the logical flow graph within a preset time window; Extract a list of processes that have accessed sensitive data from the search results, and filter out irrelevant processes by combining process lifecycle, parent-child process relationship, file handle and network connection records; The runtime of the filtered suspicious processes is compared with the vTCM metric log. If unauthorized code is loaded or the metric value changes abnormally during the process's runtime, a source tracing report is output.

6. A data flow tracking and resource release access control system based on a trusted sandbox, applicable to the data flow tracking and resource release access control method based on a trusted sandbox as described in any one of claims 1 to 5, characterized in that, It includes a data flow tracking module, a data residue detection and resource release access module, a leakage alarm and trust measurement linkage tracing module, and an audit report generation module; The data flow tracking module is used to bind to the sandbox lifecycle state machine to collect data fingerprints, behavioral events, process contexts, and vTCM metric states in the access state, running state, and pre-destruction state, forming a logical flow graph. The Data Residual Detection and Resource Release Admission Module is used to receive detection instructions from the Trusted Management Center before destruction, perform residual data scanning, and determine whether sandbox resources can be released based on the scanning results. The leakage alarm and trusted measurement linkage tracing module is used to monitor network egress traffic. When an anomaly is detected, it cross-verifies the abnormal traffic timestamp, the process list in the flow graph, and the vTCM measurement log to locate the leakage source. The audit report generation module is used to summarize circulation tracking records, residual detection reports and source tracing results, and generate an audit report that is stored after being signed with national cryptographic standards.

7. The data flow tracking and resource release access system based on a trusted sandbox according to claim 6, characterized in that, The data residue detection and resource release access module includes an instruction signature verification unit, a process freezing unit, a multi-level scanning unit, and an access determination unit. The instruction verification unit is used to verify the signature and integrity of the detection instructions from the Trusted Management Center using keys or certificates protected by vTCM. The process freezing unit is used to freeze business processes within the sandbox through the operating system's process control mechanism, while allowing the residual detection agent and audit agent processes to continue running. The multi-level scanning unit is used to perform multi-level residual matching scans on sandbox-accessible memory snapshots, disk swap files, temporary directories, log directories, core dump files, and mounted data directories; The access determination unit is used to perform resource release access determination based on the comprehensive risk score of multi-level scanning.

8. The data flow tracking and resource release access control system based on a trusted sandbox according to claim 6, characterized in that, The leakage alarm and trust measurement linkage tracing module includes: The network monitoring unit is used to perform payload feature matching for plaintext protocols and metadata feature analysis for encrypted protocols. The time window backtracking unit is used to retrieve data flow events within a preset time window in the flow graph, centered on the abnormal timestamp; The process filtering unit is used to extract a list of processes that have accessed sensitive data from the search results, and filter irrelevant processes by combining process lifecycle and parent-child process relationship; It also includes a metric cross-validation unit, which compares the runtime of a suspicious process with the vTCM metric log to identify abnormal changes in metric values ​​and output a source tracing report.

9. The data flow tracking and resource release access control system based on a trusted sandbox according to claim 6, characterized in that, The data flow tracking module includes: The fingerprint generation unit is used to generate a first-level fingerprint from structured data using a cryptographic hash algorithm and to generate a second-level feature vector from unstructured data using a feature extraction model. The flow graph construction unit is used to infer logical flow paths and incrementally update the graph based on file handles, process contexts, data fingerprints, and time windows. And a metric state reading unit, used to read the current vTCM trusted metric value and associate it with the flow event record during each state transition.

10. The data flow tracking and resource release access control system based on a trusted sandbox according to claim 6, characterized in that, The audit report generation module includes a signature and evidence storage unit; The signature and evidence storage unit is used to digitally sign the audit report using the SM2 national cryptographic algorithm and store the signed report in a secure database or blockchain evidence storage platform.

Images