A model supply chain verifiable traceability method and system

Abstract

The embodiment of the specification discloses a model supply chain verifiable traceability method and system. The method comprises the following steps: generating an anchor point binding proof according to an anchor point abstract; analyzing a pre-abstract carried in each stage proof, and constructing a directed acyclic model supply chain traceability graph; based on a deployment control node, performing closure verification on a to-be-deployed image abstract according to the traceability graph; reading the anchor point abstract in an inference container group, generating a runtime proof load containing a request abstract, a response abstract and the anchor point abstract for an inference session, signing the runtime proof load, and then writing the signed runtime proof load back to an image list identified by the anchor point abstract through a reference application programming interface; pulling a reference set from an image repository according to the anchor point abstract or an inference session credential, reconstructing a traceability graph, verifying a signature chain and an abstract reference chain of the reconstructed traceability graph, and outputting a traceability report. The embodiment of the specification realizes tamper-proof traceability and automatic access blocking of the model supply chain.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence security technology, specifically to a verifiable traceability method and system for a model supply chain. Background Technology

[0002] With the deployment of large language models and multimodal models across various industries, the AI ​​model supply chain has extended to multiple heterogeneous links, including training data, baseline training environments, weight files, quantization and distillation derivative versions, inference service images, and runtime configurations. Security risks may be introduced by all participants in the supply chain. In cloud-native scenarios, model inference services are typically delivered as container images. The image inventory is uniquely identified by content hash values; the content corresponding to the same hash value cannot be tampered with in situ without changing the hash value—this characteristic is called image immutability. However, existing supply chain traceability solutions have the following shortcomings: First, general software bills of materials and software artifact supply chain-level proofs mostly target traditional application dependency packages, failing to bind AI model weights, training data Merkle roots, hyperparameter lists, and inference image hash values ​​as a single verifiable object. In actual deployment, operations personnel typically reference images through image tags, but the hash values ​​pointed to by these tags can be silently changed, making it impossible to guarantee that the model running in the production environment is consistent with the model registered during the training phase, resulting in variable traceability anchors. Secondly, training platforms, model repositories, image repositories, and container orchestration clusters belong to different systems. Existing container image source traceability proofs typically only cover the image building steps, failing to link upwards to the model weight derivative chain (such as training, quantization, and distillation) or downwards to the hash value verification during inference instance runtime. This leads to a broken evidence chain across stages, making it difficult to fully reconstruct the supply chain path after a security incident. Thirdly, the deployment phase lacks automated verification mechanisms. Container security products focus on runtime vulnerability scanning and abnormal behavior detection, while blockchain traceability solutions focus on decentralized data storage. Neither includes model supply chain proofs in the scope of automated verification during the deployment access phase, making it impossible to promptly block replacements when images or weights are replaced. Fourthly, inference outputs lack verifiable binding with model versions. Existing model lifecycle traceability specifications mostly remain at the framework level, lacking engineered implementations that cryptographically bind inference requests and responses to specific model versions and training datasets, making it difficult to meet increasingly stringent AI compliance audit requirements. With the explosive growth of supply chain attacks targeting the AI ​​technology stack, and the continuous upgrading of attack methods such as model weight poisoning and image tag hijacking, these shortcomings further amplify the security risks of AI models in production environments. Summary of the Invention

[0003] This specification provides a verifiable traceability method and system for a model supply chain, the technical solution of which is as follows:

[0004] Firstly, embodiments of this specification provide a verifiable traceability method for a model supply chain, comprising: obtaining a content digest of the inference mirror and recording an anchor point digest; generating an anchor point binding proof based on the anchor point digest, signing the anchor point binding proof, and attaching the signed anchor point binding proof to the mirror list identified by the anchor point digest via an application programming interface; parsing the preceding digests carried in each stage proof, constructing a directed acyclic model supply chain traceability graph with stage digests as nodes and preceding digest reference relationships in each stage proof as edges; forming a signature chain through signing each stage proof and a digest reference chain through preceding digest reference relationships; calculating a graph digest for the traceability graph, writing the graph digest into a root proof, and attaching the root proof to the mirror list identified by the anchor point digest; and, based on a deployment control node, performing a traceability analysis based on the traceability graph. The process involves performing closure verification on the image digest to be deployed. Closure verification includes package consistency verification, weight consistency verification, signature trust verification, and training baseline policy verification. If all verifications pass, the inference image is deployed as an inference container group and subsequent steps continue. Otherwise, the process terminates and a verifiable failure credential is output. Within the inference container group, the anchor digest is read, and a runtime proof payload containing a request digest, response digest, and anchor digest is generated for the inference session. After signing the runtime proof payload, it is written back to the image list identified by the anchor digest via the referenced application programming interface. Based on the anchor digest or inference session credential, a reference set is pulled from the image repository, the tracing graph is reconstructed, and the signature chain and digest reference chain of the reconstructed tracing graph are verified. If both verifications pass, a tracing report is output.

[0005] Secondly, embodiments of this specification provide a verifiable traceability system for a model supply chain, comprising: an anchor registration module, used to obtain the content summary of the inference image and record the anchor summary; generate an anchor binding proof based on the anchor summary, sign the anchor binding proof, and attach the signed anchor binding proof to the image list identified by the anchor summary via an application programming interface; a traceability graph construction module, used to parse the preceding summaries carried in each stage proof, construct a directed acyclic model supply chain traceability graph with stage summaries as nodes and preceding summary reference relationships in each stage proof as edges; each stage proof forms a signature chain through signing and a summary reference chain through preceding summary reference relationships; calculates a graph summary for the traceability graph, writes the graph summary into the root proof, and attaches the root proof to the image list identified by the anchor summary; and a deployment verification module, running on the deployment control node, used to perform closure verification based on the traceability graph and the closure verification includes packaging consistency verification and weight consistency verification. The process involves several modules: verification, signature trust verification, and training baseline policy verification. If all verifications pass, the inference image is deployed as an inference container group, and subsequent steps continue. Otherwise, the process terminates and a verifiable failure credential is output. The runtime proof module, deployed within the inference container group, reads anchor digests within the inference container group and generates a runtime proof payload containing request, response, and anchor digests for the inference session. After signing the runtime proof payload, it is written back to the image list identified by the anchor digest via a referenced application programming interface (API). The audit verification module pulls a reference set from the image repository based on the anchor digest or inference session credentials, reconstructs the traceability graph, and verifies the signature chain and digest reference chain of the reconstructed traceability graph. If both verifications pass, a traceability report is output. The anchor registration module, traceability graph construction module, deployment verification module, runtime proof module, and audit verification module all use anchor digests as associated identifiers and communicate with each other via a referenced API.

[0006] The beneficial effects of the technical solutions provided in some embodiments of this specification include at least the following:

[0007] This specification's embodiments obtain a content digest of the inference image and record an anchor digest. Specifically, the content digest of the inference image is used as an immutable traceability anchor for the model supply chain, and a signature proof corresponding to the anchor digest is generated. This signature proof is then attached to the image list identified by the anchor digest via an application programming interface (API). In other words, during anchor registration, key parameters of the model supply chain are bound to the anchor digest as a single proof object. Unlike existing technologies that rely on variable labels to reference images, leading to the silent replacement of traceability anchors, this specification's embodiments cryptographically bind the traceability anchor to the image content, fundamentally blocking supply chain replacement attacks from different images with the same label, and achieving strong consistency between deployed and registered objects.

[0008] Furthermore, the embodiments of this specification can parse the preceding summaries carried in the proofs at each stage, constructing a directed acyclic model supply chain traceability graph with stage summaries as nodes and summary citation relationships as edges. This connects the training baseline, weight derivation, and inference mirror into a reconstructable complete supply chain path. Unlike existing technologies where proofs at each stage are stored in isolation and the evidence chain is easily broken when tracing across systems, the embodiments of this specification enable the complete reconstruction of the entire chain of AI model data from training to deployment through a graph structure, effectively supporting the root cause localization and impact scope assessment of security incidents.

[0009] Furthermore, this embodiment intercepts container group creation requests at the admission control point of the container orchestration engine. Based on the source map, it performs closure verification on the digest of the image to be deployed, including packaging consistency, weight consistency, signature trust, and training baseline policies. If the verification fails, container group creation is rejected. Unlike existing technologies that only perform vulnerability scanning at runtime or post-event auditing, this embodiment moves source verification to the deployment scheduling stage, blocking non-compliant models from entering the production environment at the source. It also employs a failure-closing strategy when the referenced application programming interface is unreachable, preventing security verification from being bypassed.

[0010] Furthermore, the embodiments in this specification read the anchor digest within the inference container group and generate a runtime proof payload containing a request digest, a response digest, and an anchor digest. After signing, this payload is written back to the image list identified by the anchor digest. Unlike existing technologies that lack cryptographic binding between inference output and model version, the embodiments in this specification ensure that each inference output is traceable to the specific model version that generated it, providing verifiable cryptographic evidence for model behavior auditing and output compliance verification.

[0011] Furthermore, the embodiments in this specification form a verifiable traceability closed loop from model registration to inference output through a complete step-by-step chain of anchor point registration, traceability graph construction, deployment verification, runtime proof, and audit verification. During audits or compliance reviews, the proof data of each step can be traced back, such as anchor point binding proof, proofs at each stage, root proof, runtime proof payload, etc., to achieve traceability, verifiability, and auditability throughout the entire supply chain lifecycle.

[0012] Furthermore, the embodiments in this specification employ a dual guarantee mechanism of closure verification during the deployment phase and proof write-back during the runtime phase, ensuring that traceability verification is carried out throughout the entire lifecycle of the model from deployment to inference. Compared to existing technologies that only generate traceability information during the development and building phase and lack continuous verification during the deployment and runtime phases, the embodiments in this specification prevent non-compliant models from going online during the admission phase and continuously record inference evidence during the runtime phase, thereby upgrading the security of the AI ​​model supply chain from single-point protection to full-link closed-loop protection. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this specification, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a schematic diagram of the architecture of an AI model supply chain verifiable traceability system provided in this manual.

[0015] Figure 2 This is a flowchart illustrating a model supply chain verifiable traceability method provided in this specification.

[0016] Figure 3 This is a schematic diagram of the node and edge types of the model supply chain traceability graph provided in this manual.

[0017] Figure 4 This is a flowchart illustrating the anchor point registration and referencing process provided in this manual.

[0018] Figure 5 This is a logical diagram illustrating the summary closure verification logic performed on the traceability graph by the deployment verification module provided in this manual.

[0019] Figure 6 This is a timing diagram of the deployment access and runtime verification provided in this manual. Detailed Implementation

[0020] The technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings.

[0021] The terms "first," "second," etc., in the description, claims, and accompanying drawings are used to distinguish different objects and not to describe a particular order. Furthermore, the term "comprising" and any variations thereof are intended to cover a non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such processes, methods, products, or apparatus.

[0022] This specification provides a model supply chain verifiable traceability method through several embodiments. The executing entity of this model supply chain verifiable traceability method can be the model supply chain verifiable traceability system provided in the embodiments of this invention.

[0023] Before this manual elaborates on the verifiable traceability method for model supply chains in conjunction with one or more embodiments, it first introduces the application scenarios of this verifiable traceability method for model supply chains.

[0024] Please see Figure 1 , Figure 1 This is a schematic diagram of the architecture of an AI model supply chain verifiable traceability system provided in an embodiment of the present invention. In this embodiment, the AI ​​model supply chain verifiable traceability system can be integrated into an electronic device, such as a terminal or a server. The terminal can be a mobile phone, tablet computer, smart Bluetooth device, laptop computer, or personal computer (PC); the server can be a single server or a server cluster composed of multiple servers.

[0025] In some embodiments, the AI ​​model supply chain verifiable traceability system can also be integrated into multiple electronic devices. For example, the AI ​​model supply chain verifiable traceability system can be integrated into multiple servers, and the AI ​​model supply chain verifiable traceability method of this application can be implemented by multiple servers.

[0026] In some embodiments, the server may also be implemented as a terminal. The terminal may be a device such as a mobile phone, tablet computer, laptop computer, or personal computer. The terminal includes a central processing unit, graphics processor, memory, storage device, network communication module, display screen, battery and power management module, etc.

[0027] refer to Figure 1 The electronic device may include an artifact warehouse 102, a continuous integration and continuous deployment (CI / CD) pipeline 103, an anchor registration service (ARS) 104, a traceability graph service (MGS) 105, a Kubernetes cluster 106, and an audit client (AC) 107. The various components communicate with each other via an application programming interface (API) and a digest, using the digest as a content addressing identifier for storing, querying, and retrieving artifacts and supporting documents. The referrer API is an API interface based on the HTTPS protocol.

[0028] Specifically, the anchor registration service 104, the traceability graph service 105, the Kubernetes cluster 106, and the audit client 107 communicate with the artifact repository 102 via reference application programming interfaces (APIs) to store, query, and retrieve artifacts and supporting documents using digests as content addressing identifiers. The Kubernetes cluster 106 and the audit client 107 are connected via trust and control flow to synchronize verification policies and the list of trusted root certificates. The Kubernetes cluster 106 and the audit client 107 retrieve artifacts and related attachments from the artifact repository 102 via reference APIs and digest communication. The admission control point of the Kubernetes cluster 106 deploys a deployment verification module 1061, which, during the admission verification process, retrieves all references associated with the anchor digest from the artifact repository 102 via reference APIs and digest communication. The runtime proof agent 1063 deployed within the inference container group (Pod) 1062 writes the signed runtime proof payload back to the image list identified by the anchor digest in the artifact repository 102 via reference APIs and digest communication. Audit client 107 retrieves a reference set from the workpiece warehouse 102 for traceability verification via application programming interface and summary communication. The modules synchronize their verification strategies and trust root information through trust and control flow synchronization strategies, ensuring consistency across modules. The traceability reports are exported from audit client 107 via report export flow for use by external audit systems.

[0029] The continuous integration and continuous deployment pipeline 103 is used to trigger the anchor registration and traceability graph construction process during the model packaging stage. Specifically, the continuous integration and continuous deployment pipeline 103 triggers the execution of anchor registration and traceability graph construction by calling the anchor registration service 104 and the traceability graph service 105 through application programming interfaces.

[0030] The artifact repository 102 stores AI model artifacts and related accessories, including software bills of materials, signatures, and certification documents. The artifact repository 102 supports the Open Container Initiative (OCI) application programming interface (API), allowing querying and retrieving all references associated with a summary. The artifact repository 102 includes an artifact parsing and summary calculation module for parsing the stored AI model artifacts and calculating their content summaries; an anchor registration and signature generation module for performing anchor registration and certification document signing operations; a metadata on-chain and registration module for registering and storing artifact metadata information; and a query and verification API for external components to query artifact information and verify artifact integrity. Finally, the artifact repository 102 includes a policy and trust root management module for storing and managing verification policies and trust root certificate lists.

[0031] Anchor registration service 104 is used to parse the list and extract content summaries from the inference service image, record anchor summary, collect key parameters of the model supply chain to generate anchor binding proof, sign the anchor binding proof, and attach it to the image list identified by the anchor summary via the application programming interface. Anchor registration service 104 includes a product parsing and summary calculation module for parsing the list and extracting content summaries from the inference service image; an anchor registration and signature generation module for registering anchors and generating signatures for anchor binding proofs; and a metadata on-chain and registration module for storing anchor registration-related metadata on the blockchain.

[0032] The traceability graph service 105 is used to parse the pre-concept summaries carried in the proofs of each stage. It constructs a directed acyclic model supply chain traceability graph using stage summaries as nodes and summary reference relationships as edges. After calculating the graph summary of the traceability graph, it writes the root proof and attaches the root proof to the mirror list identified by the anchor summary. The traceability graph service 105 has a built-in traceability graph construction module for performing traceability graph construction and graph summary calculation.

[0033] The Kubernetes cluster 106's admission control point deploys a deployment verification module 1061. This module, implemented as a dynamic admission controller network hook, intercepts container group creation requests and extracts image digests. Based on the source graph, it performs closure verification on the image digests to be deployed. Within the deployment verification module 1061 are: a request parsing module to parse container group creation requests and extract image digests; a policy verification module to perform policy matching in closure verification; a source verification module to reconstruct the source graph and perform digest closure verification; a signature and certificate verification module to verify the validity of signatures and certificates at each stage; and an admission decision module to make admission decisions to allow or deny container group creation based on the verification results.

[0034] The Kubernetes cluster 106 also includes an inference container group 1062, which deploys an inference service container and a runtime proof agent 1063. The inference container group 1062 runs in a trusted execution environment or relies on operating system security mechanisms. The runtime proof agent 1063 reads the anchor digest, generates a runtime proof payload for the inference session containing a request digest, a response digest, and an anchor digest, signs the runtime proof payload, and writes it back to the image manifest identified by the anchor digest via a referenced application programming interface. The runtime proof agent 1063 deploys an environment measurement module to measure the security status of the runtime environment of the inference container group 1062; an evidence generation and signing module to generate and sign runtime proof payloads; a periodic reporting and event triggering module to periodically or by event triggering of the generated runtime proof payloads; and a remote proof module to execute remote proof processes related to the trusted execution environment.

[0035] Audit client 107 is used to retrieve a reference set from the product warehouse 102 based on anchor summary or inference session credentials, reconstruct the traceability graph, verify the signature chain and summary reference chain of the reconstructed traceability graph, and output a traceability report based on the verification results. Audit client 107 includes a query and visualization module for inputting query conditions and visually displaying the traceability results; a traceability verification module for reconstructing the traceability graph and verifying the signature chain and summary reference chain; a report export module for exporting verified traceability reports as structured format files; and a policy check module for performing compliance checks on the traceability results according to preset policies.

[0036] The anchor registration service 104 may include processors and memory. The anchor registration service 104 can perform manifest parsing on the inference service image, extract content summaries to record anchor summaries; collect key parameters of the model supply chain bound to the image, including at least a model weight file summary, training data Merkle root, training baseline environment image summary, hyperparameter summary, and stage identifier; generate anchor binding proofs based on the key parameters and anchor summaries; sign the anchor binding proofs using the registered private key; and attach the signed anchor binding proofs to the image manifest identified by the anchor summaries via an application programming interface.

[0037] The traceability graph service 105 may include processors and memory, etc. The traceability graph service 105 can obtain proofs for each stage, including proofs for the training stage, quantization stage, distillation stage, and packaging stage; it parses the preceding summaries carried in each stage proof, constructing a directed acyclic model supply chain traceability graph with stage summaries as nodes and preceding summary reference relationships in each stage proof as edges; each stage proof forms a signature chain through signatures and a summary reference chain through preceding summary reference relationships; after calculating the graph summary of the traceability graph, the graph summary is written into the root proof, and the root proof is attached to the mirror list identified by the anchor summary.

[0038] The deployment verification module 1061 may include processors and memory, etc. The deployment verification module 1061 can intercept container group creation requests and extract image digests at the admission control point of the Kubernetes cluster 106; call the referenced application programming interface to pull all proofs associated with the anchor digest; perform closure verification on the image digest to be deployed according to the traceability graph, including packaging consistency verification, weight consistency verification, signature trust verification, and training baseline policy verification; if all verifications pass, the inference image is deployed as an inference container group and subsequent steps are continued; otherwise, the process is terminated and a verifiable failure credential is output.

[0039] The runtime proof agent 1063 may include a processor and memory. The runtime proof agent 1063 can read anchor digests within the inference container group 1062, which are injected into the inference container group 1062 via an application programming interface or environment variable from the container orchestration engine. It generates a runtime proof payload for the inference session, containing a request digest, a response digest, and an anchor digest. The runtime proof payload also includes an inference session identifier, model name, weight digest, and timestamp. After signing the runtime proof payload, it writes the signed runtime proof payload back to the image manifest identified by the anchor digest via the referenced application programming interface.

[0040] The audit client 107 may include a processor and memory, etc. The audit client 107 can retrieve a reference set from the artifact repository 102 based on the anchor digest or reasoning session credentials; reconstruct the traceability graph; verify the signature chain and digest reference chain of the reconstructed traceability graph, including: verifying that the signature verification of each stage proof is successful and that the corresponding public keys all belong to the preset trust anchor list; starting from the anchor digest, tracing back step by step along the preceding digests in each stage proof until the training baseline node. If each node can be traced back to the training baseline node along the preceding digest, the digest reference chain verification is successful; if all verifications are successful, a traceability report is output.

[0041] In this embodiment, the Anchor Registration Service (ARS) 104 corresponds to the Anchor Registration Module, the Source Graph Service (MGS) 105 corresponds to the Source Graph Construction Module, the Runtime Proof Proxy (RAA) 1063 corresponds to the Runtime Proof Module, and the Audit Client (AC) 107 corresponds to the Audit Verification Module. The Anchor Registration Module is used to obtain the content digest of the inference mirror and record the anchor summary; generate an anchor binding proof based on the anchor summary, sign the anchor binding proof, and attach the signed anchor binding proof to the mirror list identified by the anchor summary via the application programming interface. The Source Graph Construction Module is used to parse the preceding summaries carried in each stage proof, construct a directed acyclic model supply chain source graph using stage summaries as nodes and preceding summary reference relationships in each stage proof as edges; each stage proof forms a signature chain through signatures and a summary reference chain through preceding summary reference relationships; calculates the graph summary of the source graph, writes the graph summary into the root proof, and attaches the root proof to the mirror list identified by the anchor summary. The Deployment Verification Module runs on the deployment control node and is used to verify the source graph. The image digest to be deployed undergoes closure verification, which includes package consistency verification, weight consistency verification, signature trust verification, and training baseline policy verification. If all verifications pass, the inference image is deployed as an inference container group and subsequent steps continue; otherwise, the process terminates and a verifiable failure credential is output. The runtime proof module, deployed within the inference container group, reads the anchor digest within the inference container group and generates a runtime proof payload containing a request digest, response digest, and anchor digest for the inference session. After signing the runtime proof payload, it writes the signed runtime proof payload back to the image list identified by the anchor digest via the referenced application programming interface. The audit verification module pulls a reference set from the image repository based on the anchor digest or inference session credentials, reconstructs the traceability graph, and verifies the signature chain and digest reference chain of the reconstructed traceability graph. If both verifications pass, a traceability report is output. The anchor registration module, traceability graph construction module, deployment verification module, runtime proof module, and audit verification module all use the anchor digest as an associated identifier and communicate with each other via the referenced application programming interface.

[0042] In some embodiments, the deployment control node is a Coopernets cluster, and the deployment verification module is a network hook of the Coopernets cluster's dynamic admission controller, used to intercept container group creation requests and extract image digests; when the closure verification fails, the network hook returns a Hypertext Transfer Protocol Forbidden status code, and the response body carries a failure credential in JavaScript object abbreviation format, which includes a failure reason code, anchor digest, and graph digest.

[0043] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, the embodiment of the model supply chain verifiable traceability system is relatively simple in description because it is fundamentally similar to the embodiment of the model supply chain verifiable traceability method; relevant parts can be referred to the description of the method embodiment.

[0044] It should be noted that, Figure 1 The schematic diagram of the model supply chain verifiable traceability system shown is merely an example. The model supply chain verifiable traceability system and scenarios described in the embodiments of the present invention are for the purpose of more clearly illustrating the technical solutions of the embodiments of the present invention, and do not constitute a limitation on the technical solutions provided by the embodiments of the present invention. As those skilled in the art will know, with the evolution of the model supply chain verifiable traceability system and the emergence of new scenarios, the technical solutions provided by the embodiments of the present invention are also applicable to similar technical problems.

[0045] Please see Figure 2 , Figure 2 This is a flowchart illustrating a model supply chain verifiable traceability method provided in an embodiment of the present invention. This model supply chain verifiable traceability method can be derived from... Figure 1 The model supply chain verifiable traceability system shown performs as described. This model supply chain verifiable traceability method may include at least the following steps:

[0046] 200. Obtain the content summary of the inference mirror and record the anchor summary; generate an anchor binding certificate based on the anchor summary, sign the anchor binding certificate, and attach the signed anchor binding certificate to the mirror list identified by the anchor summary by referencing the application programming interface;

[0047] 210. Analyze the preceding digests carried in each stage proof, and construct a directed acyclic model supply chain traceability graph with stage digests as nodes and preceding digest reference relationships in each stage proof as edges; each stage proof forms a signature chain through signatures and a digest reference chain through preceding digest reference relationships; calculate the graph digest for the traceability graph, write the graph digest into the root proof, and attach the root proof to the mirror list identified by the anchor digest;

[0048] 220. Based on the deployment control node, perform closure verification according to the traceability diagram of the image to be deployed. The closure verification includes packaging consistency verification, weight consistency verification, signature trust verification and training baseline policy verification. If all verifications pass, the inference image is deployed as an inference container group and the subsequent steps are continued. Otherwise, the method flow is terminated and a verifiable failure credential is output.

[0049] 230. Read the anchor digest within the inference container group, generate a runtime proof payload for the inference session that includes a request digest, a response digest, and an anchor digest, sign the runtime proof payload, and write the signed runtime proof payload back to the image list identified by the anchor digest via the application programming interface.

[0050] 240. Pull the reference set from the mirror repository based on the anchor digest or inference session credentials, reconstruct the source graph, verify the signature chain and digest reference chain of the reconstructed source graph, and output the source report if both pass the verification.

[0051] Please see Figure 3 , Figure 3 This is a schematic diagram of the node and edge types of the model supply chain traceability graph provided in this embodiment of the invention. For example... Figure 3 As shown, the model supply chain traceability graph is a directed acyclic graph. The complete link from the training baseline to the inference instance includes: training baseline nodes, weight derivation nodes, packaging nodes, and inference instance nodes. Node types include: Open Container Initiative (OCI) Digest nodes, used to identify mirror-type nodes such as baseline models, inference images (MSAD), and runtime instances; and BlobDigest nodes, used to identify non-mirror-type nodes such as model weight files. Edge types include: TRAIN_BASE edge, representing the baseline training relationship from training to generating the baseline model; WEIGHT_DERIVE edge, representing the weight derivation relationship from deriving root weights from the baseline model; PACK edge, representing the packaging relationship of packaging weights and configurations into a model supply chain anchor summary; and DEPLOY edge, representing the deployment relationship of deploying the model supply chain anchor summary to the inference instance. Through the above organization of nodes and edges, the traceability graph completely records all derivation paths and dependencies of the AI ​​model from the training baseline to final deployment, and any anomalies in any node or edge can be detected and located.

[0052] The logic for constructing the source graph uses chain head proof (Attestation) n The entry point is the chain head proof, which is the latest generated proof file and represents the end (latest) of the chain. The construction logic specifically includes: obtaining the chain head proof (Attestation). n Chainhead proof Attestation n Includes the current digest n and prevDigest n According to the pre-digest (prevDigest) n Locate and obtain the preceding proof Attestation n-1 Precedence proof of Attestationn-1 Includes the current digest n-1 and prevDigest n-1 ; Continue analyzing the preorder proof Attestation n-1 The current digest included n-1 and prevDigest n-1 According to the pre-digest (prevDigest) n-1 Locate and obtain the preceding proof Attestation n-2 Precedence proof of Attestation n-2 Includes the current digest n-2 and prevDigest n-2 Repeat the above parsing steps recursively to parse the preceding proof, its current digest, and the preceding digest, until the root node (genesis node) with an empty preceding digest (prevDigest=∅) is reached. This root node is the training baseline node and the earliest node in the chain. Here, digest is the digest value of the current proof, prevDigest is the value pointing to the preceding proof digest, weight is the dataset weight, and sig is the signature.

[0053] In this embodiment, the rules for constructing the source graph include: proving Attestation using the chain head. n The process begins with the `prevDigest` field as the entry point; preceding proofs are recursively parsed; each proof is treated as a node; if multiple proofs' `prevDigest` points to the same node, a merged structure is formed, creating a directed acyclic graph (DAG); this continues until the root node (genesis node) with `prevDigest=∅` is reached. Through this recursive parsing, each proof is treated as a single node, and the preceding digest references in each proof are used as directed edges to construct a DAG-form supply chain traceability graph. Proofs at each stage form a signature chain through signatures and a digest reference chain through preceding digest references. After computing the graph digest of the traceability graph, the root proof is written and attached to the mirror list identified by the anchor digest.

[0054] In some embodiments, determining the anchor digest corresponding to the inference image, generating a signature certificate corresponding to the anchor digest, and attaching the signature certificate to the image list identified by the anchor digest via a referenced application programming interface (API) includes: performing list parsing on the inference image to extract the content digest of the inference image and recording it as the anchor digest; collecting key parameters of the model supply chain bound to the inference image, generating an anchor binding certificate based on the key parameters of the model supply chain and the anchor digest; signing the anchor binding certificate using the registered private key, and attaching the signed anchor binding certificate to the image list identified by the anchor digest via a referenced API; the key parameters of the model supply chain include at least a model weight file digest, training data Merkle root, training baseline environment image digest, hyperparameter digest, and stage identifier.

[0055] Please see Figure 4 , Figure 4 This is a schematic diagram of the anchor point registration and reference mounting process provided in an embodiment of the present invention. In this embodiment, the anchor point registration process includes the following steps:

[0056] 400. Package the model and related files into an Open Container Initiative Artifact (OCI Artifact) to obtain the inference service image.

[0057] 410. Perform manifest parsing on the inference service image, calculate and extract the Open Container Initiative Inventory Digest (OCIManifest Digest) as the Model Supply-chain Anchor Digest (MSAD).

[0058] 420. Collect and mirror key parameters of the model supply chain. Key parameters include model weight file summary, training data Merkle root, training baseline environment mirror summary, hyperparameter summary, and stage identifier.

[0059] 430. Generate an anchor binding certificate based on the key parameters and anchor digest. The anchor binding certificate is in the software supply chain proof (in-toto Attestation) format. Sign the anchor binding certificate using the registered private key.

[0060] 440. Invoke the Open Container Initiative Referrers API to mount the signed anchor binding proof as a referrer to the image manifest identified by the anchor digest.

[0061] 450. Determine if registration was successful. If registration is successful, the anchor point registration is complete, and the anchor point binding certificate has been successfully mounted to the anchor point summary, making it verifiable. If registration fails, perform error handling and remedial operations, including: recording error logs and context information, verifying signatures, certificates, and permission configurations, regenerating or repairing the anchor point binding certificate, retrying the mounting operation, or contacting the repository administrator.

[0062] In some embodiments, each stage of the stage proof includes at least a training stage, a quantization stage, a distillation stage, and a packaging stage; the anchor binding proof is a proof generated by the packaging stage; each stage proof records the summary of the current stage and the summary of the previous stage it depends on.

[0063] In some embodiments, the closure verification includes: package consistency verification: verifying that the digest of the inference image to be deployed is consistent with the anchor digest, and that the anchor digest corresponds to the package node in the traceability graph; weight consistency verification: verifying that the weight digest pointed to by the weight derived edge in the traceability graph is consistent with the weight digest declared in the anchor binding proof; signature trust verification: verifying that the signature verification of each stage proof is successful, and that the public key corresponding to the signature of each stage proof belongs to the preset trust anchor list; training baseline policy verification: verifying that the set of training baseline image digests allowed by the policy engine contains the training baseline node digest of the traceability graph.

[0064] In this embodiment, please refer to Figure 5 , Figure 5 This is a logical diagram illustrating the digest closure verification logic performed on the source graph by the deployment verification module provided in this embodiment of the invention. The deployment verification module receives the input source graph, which is a directed acyclic graph from the head to the tail of the chain. The digest closure verification includes the following four assertions:

[0065] Packaging Consistency Verification: This verifies that the digest of the inference image to be deployed is consistent with the anchor digest, and that the anchor digest corresponds to a packaging node in the source graph. Specifically, the digest of the image to be deployed is extracted and compared with the anchor digest; if they match, the verification is successful. Simultaneously, it verifies whether the node type corresponding to the anchor digest in the source graph is a packaging node. This assertion ensures that the actually deployed inference image is consistent with the inference image bound during anchor registration.

[0066] Weight Consistency Verification: This verifies that the weight digests pointed to by the weight-derived edges in the source graph are consistent with the weight digests declared in the anchor binding proof. Specifically, the weight digests pointed to by the weight-derived edges are extracted from the source graph, and the declared weight digests are extracted from the anchor binding proof. The two are compared, and if they match, the verification passes. This assertion ensures that the deployed model weight file is consistent with the weight file bound when the anchor is registered.

[0067] Signature Trust Verification: This verifies that the signature verification of each stage of the proof is successful, and that the public keys corresponding to the signatures of each stage of the proof belong to the preset trust anchor list, with the trust chain traceable to the trusted root chain. Specifically, a signature verification operation is performed on each stage of the proof to confirm the validity of the signature; the public key corresponding to each signature is extracted and compared with the public keys in the preset trust anchor list; if all belong to the trust anchor list, the verification is successful; the signature chain is traced back to the trusted root chain to confirm that the chain is complete and unbroken. This assertion ensures that the source of each stage of the proof is trustworthy and has not been tampered with.

[0068] Training baseline policy validation: This verifies that the set of training baseline mirror summaries allowed by the policy engine contains summaries of the training baseline nodes in the source graph. Specifically, it obtains summaries of the training baseline nodes in the source graph, queries the set of allowed training baseline mirror summaries in the policy engine, and confirms that the training baseline node summaries belong to the set; if they do, the validation passes. This assertion ensures that the baseline environment upon which the model training is based meets the preset policy requirements.

[0069] If all four assertions pass the verification, the output result is "Pass", the inference image is deployed as an inference container group and the subsequent steps are continued; if any assertion fails the verification, the output result is "Failure" and the reason for failure, the method flow is terminated and a verifiable failure credential is output.

[0070] In some embodiments, the runtime proof payload is generated by a runtime proof agent, which is deployed in the form of a sidecar container within the inference container group. The anchor summary is injected into the inference container group via the container orchestration engine's application programming interface or environment variables. The runtime proof payload also includes an inference session identifier, model name, weight summary, and timestamp.

[0071] In some embodiments, the inference session credential contains a runtime proof payload; reconstructing the source graph by pulling a reference set from the mirror repository based on the anchor digest or the inference session credential includes: pulling the corresponding runtime proof payload from the mirror repository based on the inference session credential; extracting the anchor digest from the runtime proof payload; and pulling the reference set based on the anchor digest to reconstruct the source graph.

[0072] In some embodiments, verifying the signature chain and digest reference chain of the reconstructed traceability graph includes: when the signature verification of each stage proof passes and the corresponding public key belongs to the preset trust anchor list, it indicates that the signature chain verification is passed; starting from the anchor digest, trace back step by step along the preceding digest in each stage proof until the training baseline node; if each node can be traced back to the training baseline node along the preceding digest, the digest reference chain verification is passed.

[0073] In some embodiments, when performing closure verification in the deployment control node, if the referenced application programming interface is unreachable, the method flow is terminated and an error credential indicating that the verification service is unavailable is output.

[0074] In some embodiments, please refer to Figure 6 , Figure 6 This is a timing diagram of deployment access and runtime verification provided in an embodiment of the present invention. Figure 6 The sequence diagram clearly illustrates the complete interaction process from user Pod submission to runtime attestation write-back, involving several steps among five stakeholders. Specifically, the deployment admission and runtime attestation process involves interactions between the user or operations personnel, the Kubernetes API Server, the Deployment Verification Module Network Hook (DVMWebhook), the Artifact Registry, and the Runtime Attestation Agent (RAA) in the inference container group. This process includes the following steps:

[0075] Users or operations personnel submit a container group creation request (submit a Pod) to the Kubernetes API server. The container group creation request includes image references.

[0076] The Kubernetes API server calls the deployment verification module's network hook (by calling the DVM Webhook) to forward the container group creation request to the deployment verification module;

[0077] The deployment verification module calls the reference application programming interface to query the product warehouse for all references associated with the anchor summary (query Referrers) and obtains the model supply chain traceability diagram and related attachments (obtain MSAG and other attachments).

[0078] Deploy the verification module to verify the supply chain traceability diagram of the model (verify MSAG). The verification includes signature verification, policy verification and integrity verification (signature, policy and integrity).

[0079] If all verifications pass, the deployment verification module returns an admission review: Allowed to the Kubernetes API server.

[0080] The Kubernetes API server creates a container group, schedules and starts the inference container group, and the runtime proof agent deployed within the inference container group starts running (creating a Pod, scheduling and starting the RAA).

[0081] The runtime proof agent generates and signs the runtime proof payload (generates and signs the runtime AttestationMSA-RT).

[0082] The runtime proof agent writes the signed runtime proof payload back to the mirror list identified by the anchor digest in the artifact repository by referencing the application programming interface (write-back to Attestation, ORAS push), and admission is successful.

[0083] This specification's embodiments obtain a content digest of the inference image and record an anchor digest. Specifically, the content digest of the inference image is used as an immutable traceability anchor for the model supply chain, and a signature proof corresponding to the anchor digest is generated. This signature proof is then attached to the image list identified by the anchor digest via an application programming interface (API). In other words, during anchor registration, key parameters of the model supply chain are bound to the anchor digest as a single proof object. Unlike existing technologies that rely on variable labels to reference images, leading to the silent replacement of traceability anchors, this specification's embodiments cryptographically bind the traceability anchor to the image content, fundamentally blocking supply chain replacement attacks from different images with the same label, and achieving strong consistency between deployed and registered objects.

[0084] Furthermore, the embodiments of this specification can parse the preceding summaries carried in the proofs at each stage, constructing a directed acyclic model supply chain traceability graph with stage summaries as nodes and summary citation relationships as edges. This connects the training baseline, weight derivation, and inference mirror into a reconstructable complete supply chain path. Unlike existing technologies where proofs at each stage are stored in isolation and the evidence chain is easily broken when tracing across systems, the embodiments of this specification enable the complete reconstruction of the entire chain of AI model data from training to deployment through a graph structure, effectively supporting the root cause localization and impact scope assessment of security incidents.

[0085] Furthermore, this embodiment intercepts container group creation requests at the admission control point of the container orchestration engine. Based on the source map, it performs closure verification on the digest of the image to be deployed, including packaging consistency, weight consistency, signature trust, and training baseline policies. If the verification fails, container group creation is rejected. Unlike existing technologies that only perform vulnerability scanning at runtime or post-event auditing, this embodiment moves source verification to the deployment scheduling stage, blocking non-compliant models from entering the production environment at the source. It also employs a failure-closing strategy when the referenced application programming interface is unreachable, preventing security verification from being bypassed.

[0086] Furthermore, the embodiments in this specification read the anchor digest within the inference container group and generate a runtime proof payload containing a request digest, a response digest, and an anchor digest. After signing, this payload is written back to the image list identified by the anchor digest. Unlike existing technologies that lack cryptographic binding between inference output and model version, the embodiments in this specification ensure that each inference output is traceable to the specific model version that generated it, providing verifiable cryptographic evidence for model behavior auditing and output compliance verification.

[0087] Furthermore, the embodiments in this specification form a verifiable traceability closed loop from model registration to inference output through a complete step-by-step chain of anchor point registration, traceability graph construction, deployment verification, runtime proof, and audit verification. During audits or compliance reviews, the proof data of each step can be traced back, such as anchor point binding proof, proofs at each stage, root proof, runtime proof payload, etc., to achieve traceability, verifiability, and auditability throughout the entire supply chain lifecycle.

[0088] Furthermore, the embodiments in this specification employ a dual guarantee mechanism of closure verification during the deployment phase and proof write-back during the runtime phase, ensuring that traceability verification is carried out throughout the entire lifecycle of the model from deployment to inference. Compared to existing technologies that only generate traceability information during the development and building phase and lack continuous verification during the deployment and runtime phases, the embodiments in this specification prevent non-compliant models from going online during the admission phase and continuously record inference evidence during the runtime phase, thereby upgrading the security of the AI ​​model supply chain from single-point protection to full-link closed-loop protection.

[0089] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0090] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, the embodiments of the model supply chain verifiable traceability method are relatively simple in description because they are fundamentally similar to the model supply chain verifiable traceability system embodiments; relevant parts can be referred to the descriptions in the system embodiments.

[0091] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks. Unless otherwise specified, the technical features of this embodiment and its implementation can be combined arbitrarily.

[0092] The above embodiments are merely preferred embodiments described in this specification and are not intended to limit the scope of this specification. Any modifications and improvements made by those skilled in the art to the technical solutions of this specification without departing from the spirit of this specification should fall within the protection scope defined by the claims of this specification.

Claims

1. A verifiable traceability method for a model supply chain, characterized in that, include: Obtain the content summary of the inference mirror and record the anchor point summary; An anchor binding proof is generated based on the anchor digest, the anchor binding proof is signed, and the signed anchor binding proof is mounted to the mirror list identified by the anchor digest via the application programming interface. The preceding digests carried in each stage proof are parsed, and a directed acyclic model supply chain traceability graph is constructed with stage digests as nodes and preceding digest reference relationships in each stage proof as edges. Each stage proof forms a signature chain through signatures and a digest reference chain through preceding digest reference relationships. After calculating the graph digest of the traceability graph, the graph digest is written into the root proof and the root proof is attached to the mirror list identified by the anchor digest. Based on the deployment control node, a closure verification is performed on the digest of the image to be deployed according to the traceability graph. The closure verification includes packaging consistency verification, weight consistency verification, signature trust verification and training baseline policy verification. If all verifications pass, the inference image is deployed as an inference container group and the subsequent steps are continued. Otherwise, the method flow is terminated and a verifiable failure credential is output. The anchor digest is read within the inference container group, and a runtime proof payload containing a request digest, a response digest, and the anchor digest is generated for the inference session. After signing the runtime proof payload, the signed runtime proof payload is written back to the image list identified by the anchor digest via the referenced application programming interface. Based on the anchor digest or inference session credentials, pull the reference set from the mirror repository, reconstruct the source graph, verify the signature chain and digest reference chain of the reconstructed source graph, and output a source report if both pass the verification.

2. The method according to claim 1, characterized in that, The process of determining the anchor digest corresponding to the inference mirror, generating a signature proof corresponding to the anchor digest, and attaching the signature proof to the mirror list identified by the anchor digest via an application programming interface includes: Perform manifest parsing on the inference mirror to extract the content summary of the inference mirror, and record it as the anchor summary; Collect key parameters of the model supply chain that are bound to the inference mirror, and generate an anchor binding proof based on the key parameters of the model supply chain and the anchor summary; The anchor binding proof is signed using the registered private key, and the signed anchor binding proof is mounted to the mirror list identified by the anchor digest through the referenced application programming interface; The key parameters of the model supply chain include at least the model weight file summary, training data Merkle root, training baseline environment mirror summary, hyperparameter summary, and stage identifier.

3. The method according to claim 1, characterized in that, The stages in the proof at each stage include at least a training stage, a quantization stage, a distillation stage, and a packaging stage; the anchor binding proof is the proof generated by the packaging stage; each stage proof records the summary of the current stage and the summary of the previous stage it depends on.

4. The method according to claim 1, characterized in that, In the closure verification: The packaging consistency verification is as follows: verifying that the digest of the inference image to be deployed is consistent with the anchor digest, and that the anchor digest corresponds to the packaging node in the source graph; The weight consistency verification is as follows: verifying that the weight digest pointed to by the weight-derived edge in the source graph is consistent with the weight digest declared in the anchor binding proof; The signature trust verification is as follows: verifying that the signature verification of each stage of the proof passes and that the public key corresponding to the signature of each stage of the proof belongs to the preset trust anchor list. The training baseline policy verification is as follows: the set of training baseline mirror summaries allowed by the verification policy engine contains the training baseline node summaries of the source graph.

5. The method according to claim 1, characterized in that, The runtime proof payload is generated by the runtime proof agent, which is deployed in the inference container group as a sidecar container. The anchor summary is injected into the inference container group via the container orchestration engine's application programming interface or environment variables. The runtime proof payload also includes an inference session identifier, model name, weight summary, and timestamp.

6. The method according to claim 1, characterized in that, The inference session credential contains the runtime proof payload; the step of reconstructing the source graph by pulling the reference set from the mirror repository based on the anchor digest or the inference session credential includes: Based on the inference session credentials, the corresponding runtime proof payload is retrieved from the mirror repository; Extract the anchor point summary from the runtime proof payload; The reference set is pulled from the anchor summary to reconstruct the source graph.

7. The method according to claim 1, characterized in that, The signature chain and digest citation chain of the source map obtained from the verification and reconstruction include: When the signature verification of each stage of the proof passes and the corresponding public key belongs to the preset trust anchor list, it indicates that the signature chain verification is successful. Starting from the anchor point summary, trace back step by step along the preceding summaries in each stage of the proof until the training baseline node. If each node can be traced back to the training baseline node along the preceding summaries, then the summary reference chain verification is successful.

8. The method according to claim 1, characterized in that, When performing closure verification in the deployment control node, if the referenced application programming interface is unreachable, the method flow is terminated and an error credential indicating that the verification service is unavailable is output.

9. A verifiable traceability system for a model supply chain, characterized in that, include: The anchor point registration module is used to obtain the content summary of the inference mirror and record the anchor point summary; An anchor binding proof is generated based on the anchor digest, the anchor binding proof is signed, and the signed anchor binding proof is mounted to the mirror list identified by the anchor digest via the application programming interface. The traceability graph construction module is used to parse the pre-concept summaries carried in each stage proof, and construct a directed acyclic model supply chain traceability graph with stage summaries as nodes and pre-concept summary reference relationships in each stage proof as edges; each stage proof forms a signature chain through signatures and a summary reference chain through pre-concept summary reference relationships; after calculating the graph summary of the traceability graph, the graph summary is written into the root proof and the root proof is attached to the mirror list identified by the anchor point summary; The deployment verification module runs on the deployment control node and is used to perform closure verification on the digest of the image to be deployed according to the traceability graph. The closure verification includes packaging consistency verification, weight consistency verification, signature trust verification and training baseline policy verification. If all verifications pass, the inference image is deployed as an inference container group and the subsequent steps are continued. Otherwise, the method flow is terminated and a verifiable failure credential is output. The runtime proof module is deployed within the inference container group. It is used to read the anchor digest within the inference container group, generate a runtime proof payload for the inference session that includes a request digest, a response digest, and the anchor digest, sign the runtime proof payload, and write the signed runtime proof payload back to the image list identified by the anchor digest via the referenced application programming interface. The audit verification module is used to pull the reference set from the mirror repository based on the anchor digest or reasoning session credential, reconstruct the source graph, verify the signature chain and digest reference chain of the reconstructed source graph, and output a source report if both pass the verification. The anchor point registration module, the source map construction module, the deployment verification module, the runtime verification module, and the audit verification module all use the anchor point summary as an associated identifier and communicate with each other through the referenced application programming interface.

10. The system according to claim 9, characterized in that, The deployment control node is a Coopernets cluster, and the deployment verification module is a network hook of the dynamic admission controller of the Coopernets cluster, used to intercept container group creation requests and extract image digests; when the closure verification fails, the network hook returns a Hypertext Transfer Protocol Forbidden status code, and the response body carries a failure credential in JavaScript object abbreviation format, which includes a failure reason code, the anchor digest, and a graph digest.

Images