Multi-modal encryption jewelry appraisal system and interaction method based on federal learning framework

Abstract

The invention provides a multi-modal encryption jewelry appraisal system based on a federal learning framework and an interaction method. According to the system, a user-collaborative private training data set is constructed, and in combination with ciphertext feature processing and block chain non-tampering evidence storage, the indexes that the AI identification response is less than or equal to 5 seconds and the manual recheck period is less than or equal to 3 workdays are realized. And an SM4-TLS dual-protocol end-to-end encryption and differential privacy mechanism is adopted, so that the data leakage risk is reduced by 78%. The federal learning framework enables the model training efficiency to be improved by 57% compared with traditional centralized training, and serves the fields of jewelry retail and judicial evidence collection. The system efficiently extracts features by using an AI server cluster, and completes jewelry multi-modal analysis at a millisecond level. And when high-confidence-coefficient features are detected, a structured report is automatically generated, and meanwhile, an expert rechecking mechanism is embedded to guarantee the accuracy. And the block chain evidence storage module carries out timestamp evidence storage on each round of federal learning parameter updating and identification results to ensure that an evidence chain is complete and traceable. Not only is the accuracy of manual identification kept, but also the overall time consumption is shortened through the technical means.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of jewelry identification, and specifically relates to a multi-modal encrypted jewelry identification system and an interaction method based on a federated learning framework. BACKGROUND

[0002] Jewelry identification refers to a process of systematically analyzing and evaluating the types, authenticity, quality, technology, origin and historical value of jewelry and jade through scientific means, aiming to provide authoritative and accurate identification results and market reference for consumers, collectors, auction houses and the jewelry industry. Jewelry identification usually relies on professional equipment such as refractometers, microscopes, spectrometers and density meters, combining traditional visual inspection methods with modern scientific methods to comprehensively detect the physical properties (including refractive index, hardness, density), chemical composition, internal structure and processing technology of gemstones. Identification agencies issue legally binding identification certificates based on international or national standards to ensure transaction safety and consumer rights. In addition, jewelry identification also involves the identification of imitation products, optimized processing products and natural gemstones, as well as the evaluation of historical jewelry cultural value, which is an indispensable part of the jewelry industry.

[0003] However, the existing technology in the field of jewelry identification has several shortcomings: first, the data security protection is weak, traditional methods often use plaintext transmission or simple encryption mechanisms, and user privacy data is at risk of leakage during transmission and storage; second, the identification efficiency is low, relying on manual operation and professional equipment, which takes a long time from sample shooting to issuing results; third, there is a contradiction between model training and data sharing, centralized training requires the collection of user original data, which significantly increases the risk of privacy leakage and limits the continuous optimization ability of the model; fourth, the identification process lacks traceable full-process records, making it difficult for identification results to be fully recognized in judicial evidence, insurance damage assessment and other scenarios. SUMMARY

[0004] The purpose of the present application is to solve the above problems and provide a multi-modal encrypted jewelry identification system and an interaction method based on a federated learning framework.

[0005] The technical solution adopted by the present application is as follows: a multi-modal encrypted jewelry identification system and an interaction method based on a federated learning framework, the system comprising:

[0006] AI server cluster: used for implementing encryption processing, feature extraction and model training of jewelry data, containing the following core units:

[0007] Federated learning training unit: adopts a hybrid architecture of jewelry feature-aware Transformer-CNN, supports dynamic incremental training based on fluctuations in jewelry market value (e.g. triggers emergency training when gold price fluctuation >5%) and generates global model parameter update files; when users participate in data co-construction, jewelry pictures are processed with Laplacian noise differential privacy, training data pool is protected with differential privacy parameters ε=2, δ=1e-5, and model parameter update files are synchronized daily to at least 100 edge nodes via HTTPS protocol.

[0008] Feature vector generator: deploys a multi-modal fusion model EfficientNet-B4+ViT-S, inputs encrypted jewelry picture data, outputs polynomial mask, CKKS+(PolyMask) double encryption; if multi-modal fusion confidence ≥98%, triggers automatic report generation process.

[0009] Homomorphic encryption chip: custom jewelry identification dedicated encryption chip JadeCrypto-1, supports CKKS homomorphic encryption scheme, ensures that the server only operates on encrypted feature vectors.

[0010] Expert review terminal: used to realize manual review and identification record evidence, contains the following modules:

[0011] Micro-video acquisition device: integrates a 40x optical zoom lens, a 3840x2160 resolution 30fps frame rate sensor, and a facet intersection recognition accelerator to verify jewelry features.

[0012] Blockchain evidence module: based on Tencent Cloud TBaaS platform, solidifies identification process screen recording files, operation logs, and feature labeling information, and uses SHA-256 hash algorithm.

[0013] Encrypted communication module: used to realize data transmission and storage security, contains the following modules:

[0014] Transmission layer encryption: uses SM4 national encryption algorithm and TLS1.3 protocol to build a double-layer encryption channel.

[0015] Storage layer key management: keys are dynamically generated by Tencent Cloud KMS management system.

[0016] Abnormal traffic fuse mechanism: triggers traffic restriction when detecting single IP address request frequency >100 times / sec; switches to TLS1.3+SM3 backup encryption channel when packet loss rate >0.1%.

[0017] One-way data gate: set between AI server cluster and expert review terminal, controlled by STM32F407 chip, only allows one-way transmission of ciphertext feature vectors from AI server cluster to expert review terminal.

[0018] In a preferred embodiment, the multi-modal encrypted jewelry identification system based on federated learning framework and its interaction method, its multi-modal data secure interaction method covers the following steps:

[0019] S1: The user takes pictures of the jewelry facets through the mobile application, preparing for identification.

[0020] S2: The user uploads the jewelry pictures through the SM4 national encryption algorithm and the double-layer encryption channel constructed by the TLS1.3 protocol to the AI server cluster for encrypted transmission, and the transmission layer key is dynamically generated by the Tencent cloud KMS management system.

[0021] S3: The feature vector generator of the AI server cluster deploys the multi-modal fusion model EfficientNet-B4+ViT-S to process the encrypted jewelry picture data, and the facet intersection feature enhancement layer (FacetEnhancer) input data expansion: encrypted image + facet intersection space coordinates (from the microscopic video acquisition device); generate a polynomial mask, CKKS+(PolyMask) double encryption; if the multi-modal fusion confidence is ≥98%, trigger the automatic report generation process.

[0022] S4: When the user chooses to participate in data co-construction, the jewelry pictures are processed by Laplace noise differential privacy, and the training data pool is protected according to the differential privacy parameters ε=2, δ=1e-5; Incremental training is started at 02:00 every day, and the federated learning training unit generates global model parameter update files and synchronizes them to at least 100 edge nodes through the HTTPS protocol.

[0023] S5: The homomorphic encryption chip of the AI server cluster uses the CKKS homomorphic encryption scheme to encrypt the generated polynomial mask, CKKS+(PolyMask) double encryption, to ensure that the server only operates on the encrypted feature vector.

[0024] The specific process of the homomorphic encryption chip of the AI server cluster using the CKKS scheme to encrypt the polynomial mask, CKKS+(PolyMask) double encryption, is as follows:

[0025] S5-1: Key generation: dynamically generate the public key (pk), private key (sk), and evaluation key (evk) required by the CKKS algorithm through the Tencent cloud KMS management system, where the public key is used for encryption, the private key is used for decryption, and the evaluation key supports subsequent homomorphic operations (such as feature vector similarity comparison).

[0026] S5-2: Feature vector encoding: map the 512-dimensional real-valued feature vector v=(v1, v2,..., v 512 ) to the complex domain polynomial ring Z[X] / (X^ N+1) In the middle, the vector elements are embedded into the polynomial coefficients by "slot filling" technique, forming the plaintext polynomial m(X).

[0027] S5-3: Encryption processing: Perform encryption operation on the plaintext polynomial m(X) using the public key, generate ciphertext: C=(a(X), b(X));

[0028] Where the ciphertext components satisfy b(X)=a(X)·s(X)+e(X)+m(X)·Δ;

[0029] Where s(X) is the private key polynomial, e(X) is the Gaussian noise, and Δ is the scaling factor, which ensures the real value precision.

[0030] S5-4: Homomorphic operation support: The encrypted feature vector ciphertext can directly participate in the model parameter update in federated learning or the feature comparison of the expert review terminal, without decryption, the inner product, addition and other operations can be completed, avoiding plaintext leakage.

[0031] The polynomial encoding formula of the 512-dimensional feature vector is as follows:

[0032]

[0033] In the formula, each component vk of the 512-dimensional feature vector v is encoded as the coefficient of the polynomial m(X), where X ^{k-1} is the polynomial base. This encoding method realizes the compact representation of high-dimensional vectors, adapts to the encryption requirements of CKKS on polynomial ring elements, while preserving the dimension structure of the vector, supporting subsequent homomorphic vector operations;

[0034] S5-5: Encrypt the 512-dimensional feature vector v by CKKS first, and then add nonlinear noise through PolyMask:

[0035] venc=PolyMask(CKKS-Enc(v),k,M)

[0036] PolyMask calculation formula:

[0037] PolyMask(C,k,M)=(Ck)modM

[0038] Where:

[0039] C: original ciphertext (such as CKKS encrypted feature vector);

[0040] k: dynamically generated polynomial index (positively related to data sensitivity);

[0041] M: modulus of encryption scheme (determined by key parameters).

[0042] Even if the attacker cracks the outer CKKS, he still needs to solve the polynomial discrete logarithm problem to obtain the original data; The integration mode of CKKS: PolyMask as the second layer of encryption, superimposed with CKKS linear homomorphism; Facet feature mapping accelerator (processing speed increased to 18GB / s), quantum attack resistant module (based on NTRU algorithm, responding to the traffic fuse requirement of document S9).

[0043] S6: The AI server cluster is connected to the expert review terminal through a one-way data gate controlled by an STM32F407 chip, allowing only the encrypted polynomial mask, CKKS+(PolyMask) double encryption to be transmitted to the expert review terminal at a rate of not more than 1000 data packets / second, and automatically triggering a fuse mechanism for abnormal traffic.

[0044] S7: The expert review terminal monitors the appraiser's load state in real time through a Redis ZSET dynamic allocator, determines the task assignment order according to a priority calculation formula, and automatically reassigns tasks to other appraiser terminals if they have not responded for more than 30 minutes.

[0045] S8: If the user applies for review, the system assigns the task to an idle appraiser through dynamic task allocation; the appraiser uses a microscopic video capture device to verify the jewelry features; the blockchain storage module is based on the Tencent Cloud TBaaS platform, which solidifies the appraisal process screen recording file, operation log and feature labeling information. The judicial evidence storage chain: (1) AI real-time verification storage (automatic comparison of microscopic video and feature vector), (2) judicial agency node access (courts / quality inspection bureaus can verify in real time, see task allocation process Figure 3 ).

[0046] S9: The encrypted communication module monitors transmission traffic in real time and automatically triggers traffic restriction when detecting that the request frequency of a single IP address is >100 times / second; when the data packet loss rate is >0.1%, switch to a TLS1.3+SM3 backup encryption channel to ensure communication security.

[0047] In a preferred embodiment, step S1 requires the user to take a facet image of the jewelry to be identified through a mobile application, and the image resolution should meet or exceed the 4K standard to ensure that the details of the jewelry surface are clear and identifiable, thereby providing high-quality raw data for subsequent feature extraction.

[0048] In a preferred embodiment, step S2 uses a double encryption transmission channel to ensure data security: the transmission layer integrates the SM4 national encryption algorithm and the TLS1.3 protocol to build an encryption mechanism; the storage layer key is dynamically generated and hosted by the Tencent Cloud KMS management system, with a key rotation period strictly limited to 24 hours to effectively reduce the risk of key leakage and ensure the security of the data transmission process.

[0049] In a preferred embodiment, step S3 is performed by the AI server cluster deploying a multi-modal fusion model EfficientNet-B4+ViT-S to process the encrypted jewelry image data and generate a polynomial mask, double-encrypted by CKKS+(PolyMask). According to laboratory tests on Kirin 9000S chips, the time taken by the mobile terminal to generate the feature vector is less than 0.8 seconds; when the confidence level of the feature vector reaches or exceeds 95%, the system automatically triggers the report generation process, efficiently completing the preliminary identification.

[0050] In a preferred embodiment, step S4 provides that when the user participates in data co-construction, the jewelry image provided by the user needs to be processed by differential privacy, that is, data desensitization is achieved by adding Laplace noise. The training data pool uses differential privacy parameters of ε=2 and δ=1e-5 to implement protection, avoiding reverse restoration of the user's original data; the federal learning training unit starts incremental training at 2 a.m. every day, and the generated global model parameter update file has a file size of no more than 500MB, and is synchronized and distributed to at least 100 edge nodes through the HTTPS protocol, ensuring the continuous optimization and collaborative update of the model.

[0051] In a preferred embodiment, in step S5, the homomorphic encryption chip in the AI server cluster uses the CKKS homomorphic encryption scheme to implement encryption processing on the generated polynomial mask, double-encrypted by CKKS+(PolyMask). The ciphertext processing speed of the homomorphic encryption chip is not less than 12GB / s, ensuring that the server can only operate on the encrypted feature vector, thereby guaranteeing the privacy of the original data at the hardware level.

[0052] In a preferred embodiment, in step S6, the AI server cluster and the expert review terminal are connected through a one-way data gate controlled by an STM32F407 chip. This connection only allows the double-encrypted polynomial mask, CKKS+(PolyMask), to be transmitted from the AI server cluster to the expert review terminal in one direction, with a transmission rate limit of no more than 1000 data packets per second; if abnormal traffic is detected, the system will automatically trigger a fuse mechanism to block abnormal data transmission.

[0053] In a preferred embodiment, in step S7, the expert review terminal monitors the load state of the appraiser in real time through a Redis ZSET queue, and the monitoring indicators include task backlog, response delay, and historical accuracy. The system dynamically calculates the task priority and adjusts the dispatch order based on the above indicators; if the appraiser does not respond to the assigned task within 30 minutes, the system will automatically reassign the task to other idle appraiser terminals to ensure the efficiency of the review.

[0054] The calculation formula used for task priority calculation is:

[0055]

[0056] Task timeout mechanism: tasks that have not been responded to for more than 30 minutes are automatically transferred to other appraiser terminals.

[0057] In a preferred embodiment, in step S8, the system assigns the task to an idle appraiser by a dynamic task assignment mechanism; the appraiser performs manual verification using a microscopic video acquisition device, which is equipped with a 40x optical zoom lens, a 30fps frame rate sensor with a resolution of 3840x2160, and a facet intersection recognition accelerator, for observing microscopic details such as jewelry package distribution; during the appraisal process, the blockchain storage module relies on the Tencent Cloud TBaaS platform to solidify the storage of AI and manual dual-channel storage files (including confidence conflict markers), operation logs, and feature labeling information, ensuring that the appraisal records cannot be tampered with;

[0058] In step S9, the encrypted communication module monitors the transmission traffic in real time, and when it detects that the request frequency of a single IP address exceeds 100 times per second, the system automatically triggers traffic limiting measures; if the packet loss rate exceeds 0.1%, the system switches to a TLS1.3+SM3 backup encryption channel to continue transmission, ensuring the stability and security of data transmission through a dual-channel protection mechanism.

[0059] Based on the above technical solutions, the present application has the following advantages:

[0060] 1. The present application constructs an efficient appraisal process collaboration mechanism. Compared with traditional jewelry appraisal, which relies on manual operation and requires a large amount of time and professional instruments, the present application uses the rapid feature extraction capability of the AI server cluster to complete preliminary analysis in a short time and automatically generate a report when high-confidence features are detected. At the same time, in combination with an expert review mechanism, the system dynamically assigns idle appraisers to perform manual verification, thereby significantly shortening the overall process time while retaining the accuracy of manual appraisal, forming an efficient collaboration mode of "AI rapid processing + manual accurate review", and meeting the needs of timeliness in jewelry retail, insurance damage assessment, and other scenarios.

[0061] 2. The present application constructs a full-link data privacy protection system. In view of the deficiencies of traditional automated appraisal systems, such as clear transmission of user data and centralized training leading to leakage, the system uses multiple technical measures to ensure data privacy and security: a double-encryption channel is constructed by SM4 national encryption algorithm and TLS1.3 protocol in the transmission process, and the storage layer key is periodically rotated by a dynamic management system; only encrypted feature vectors are operated in the feature processing link, and differential privacy technology is used to desensitize user data involved in model co-construction, ensuring that the original information cannot be reversed. The above measures block the risk of privacy leakage from data collection, transmission to processing, and meet the requirements of data security related regulations.

[0062] 3. The present application enhances the credibility and authority of the identification results. Traditional AI identification is difficult to serve as judicial evidence due to the lack of traceable process records; and the present system, through microscopic video verification of the expert review terminal and the block chain storage module, solidifies the storage of key information of the identification process (such as operation log, feature labeling result), and the millisecond level timestamp ensures that the record cannot be tampered with. At the same time, the collaborative mechanism of artificial review and AI analysis, while taking advantage of the fast processing of AI, makes up for the limitations of AI through the experience of professional appraisers, making the identification conclusion more reliable, and meeting the needs of higher authority scenes such as judicial evidence and quality certification. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1 The system architecture topology of the present application is shown in the figure;

[0064] Figure 2 The working principle diagram of the one-way data gate in the present application is shown in the figure;

[0065] Figure 3 The flow chart of dynamic task allocation in the present application is shown in the figure. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical scheme and advantages of the present application clearer and more definite, the present application will be described in detail in combination with the attached figures and specific embodiments. It should be noted that the specific embodiments described herein are only used to illustrate the present application and do not constitute any limitation on the present application.

[0067] Reference Figure 1 ,

[0068] The multi-modal encrypted jewelry identification system and interaction method based on the federated learning framework, the system comprises:

[0069] AI server cluster: used for completing the encryption processing, feature extraction and model training of jewelry data, including the following core units:

[0070] Federated learning training unit: adopts a mixed architecture of jewelry feature perception Transformer-CNN (model parameter quantity 1.2 billion), supports dynamic incremental training based on jewelry market value fluctuation (for example, triggering emergency training when gold price fluctuation > 5%) and generates global model parameter update file (volume ≤ 500MB); when users participate in data co-construction, the jewelry pictures are processed by Laplace noise (μ = 0.1, σ = 0.05) differential privacy, the training data pool is protected according to the differential privacy parameters of ε = 2, δ = 1e-5, and the model parameter update file is synchronized to at least 100 edge nodes daily through HTTPS protocol.

[0071] Feature vector generator: Deploy the multi-modal fusion model EfficientNet-B4+ViT-S, input encrypted jewelry picture data, output polynomial mask, CKKS+(PolyMask) double encryption (mobile terminal generation time ≤0.8 seconds, based on Kirin 9000S chip actual measurement); if the multi-modal fusion confidence is ≥98%, trigger automatic report generation.

[0072] Homomorphic encryption chip: Customized jewelry identification special encryption chip JadeCrypto-1, supports CKKS homomorphic encryption scheme (ciphertext processing speed ≥12GB / s), ensures that the server only operates encrypted feature vectors.

[0073] Expert review terminal: Used for manual review and identification record storage, including the following modules:

[0074] Micro-video acquisition device: Integrates a 40x optical zoom lens, a 3840x2160 resolution 30fps frame rate sensor, and a facet intersection recognition accelerator (recognition accuracy ±0.01mm) for verifying jewelry features such as inclusion distribution.

[0075] Blockchain storage module: Based on Tencent Cloud TBaaS platform, solidifies identification process screen recording files, operation logs (records the identification expert's zoom, annotation, and comparison operations on the micro-video), and feature annotation information (timestamp accurate to milliseconds, stores facet intersection coordinates, inclusion distribution, and risk level annotation results in JSON format), and the hash algorithm is SHA-256.

[0076] Encrypted communication module: Used to ensure data transmission and storage security, including the following modules:

[0077] Transmission layer encryption: Uses SM4 national encryption algorithm (block length 128 bits) and TLS1.3 protocol to build a double-layer encrypted channel;

[0078] Storage layer key management: The key is dynamically generated by the Tencent Cloud KMS management system (key rotation period ≤24 hours);

[0079] Abnormal traffic fuse: Trigger traffic restriction when detecting that the request frequency of a single IP address is >100 times / second; switch to a TLS1.3+SM3 backup encryption channel when the packet loss rate is >0.1%.

[0080] One-way data gate: Set between the AI server cluster and the expert review terminal, the AI server cluster and the expert review terminal are connected through a one-way data gate controlled by an STM32F407 chip, allowing only encrypted feature vectors (encrypted length 1024 bytes) to be transmitted from the AI server cluster to the expert review terminal in one direction (transmission rate limited to ≤1000 packets per second).

[0081] The multi-modal encrypted jewelry identification system and interaction method based on a federal learning framework includes the following steps:

[0082] S1: The user takes a jewelry facet picture through a mobile application, preparing for identification.

[0083] S2: The jewelry picture uploaded by the user is encrypted and transmitted to the AI server cluster through a double-layer encryption channel constructed by the SM4 national encryption algorithm and the TLS1.3 protocol. The transmission layer key is dynamically generated by the Tencent cloud KMS management system.

[0084] S3: The feature vector generator of the AI server cluster deploys the multi-modal fusion model EfficientNet-B4+ViT-S to process the encrypted jewelry picture data. The FacetEnhancer input data expansion: encrypted image + facet intersection space coordinates (from the microscopic video acquisition device); generate a polynomial mask, CKKS+(PolyMask) double encryption; if the multi-modal fusion confidence is ≥98%, trigger automatic report generation.

[0085] S4: When the user chooses to participate in data co-construction, the jewelry picture is processed by Laplace noise differential privacy, and the training data pool is protected according to the differential privacy parameters of ε=2 and δ=1e-5; incremental training is started at 02:00 every day, and the federal learning training unit generates global model parameter update files and synchronizes them to at least 100 edge nodes through the HTTPS protocol.

[0086] S5: The homomorphic encryption chip of the AI server cluster uses the CKKS homomorphic encryption scheme to encrypt the generated polynomial mask, CKKS+(PolyMask) double encryption, ensuring that the server only operates on encrypted feature vectors.

[0087] The specific process of the homomorphic encryption chip of the AI server cluster using the CKKS scheme to encrypt the polynomial mask, CKKS+(PolyMask) double encryption, is as follows:

[0088] S5-1: Key generation: dynamically generate the public key (pk), private key (sk), and evaluation key (evk) required by the CKKS algorithm through the Tencent cloud KMS management system. The public key is used for encryption, the private key is used for decryption, and the evaluation key supports subsequent homomorphic operations (such as feature vector similarity comparison).

[0089] S5-2: Feature vector encoding: map the 512-dimensional real-valued feature vector v=(v1, v2,..., v 512 ) to the complex domain polynomial ring Z[X] / (X N+1) In the middle, the vector elements are embedded into the polynomial coefficients by "slot filling" technique, forming the plaintext polynomial m(X).

[0090] S5-3: Encryption processing: Perform encryption operation on the plaintext polynomial m(X) using the public key, generate ciphertext: C=(a(X), b(X);

[0091] Where the ciphertext components satisfy b(X)=a(X)·s(X)+e(X)+m(X)·Δ;

[0092] Where s(X) is the private key polynomial, e(X) is the Gaussian noise, and Δ is the scaling factor, which ensures the real value precision.

[0093] S5-4: Homomorphic operation support: The encrypted feature vector ciphertext can directly participate in the model parameter update in federated learning or the feature comparison of the expert review terminal, without decryption, the inner product, addition and other operations can be completed, avoiding plaintext leakage.

[0094] The polynomial encoding formula of the 512-dimensional feature vector is:

[0095]

[0096] In the formula: Each component vk of the 512-dimensional feature vector v is encoded as the coefficient of the polynomial m(X), where X k-1 is the polynomial base. This encoding method realizes the compact representation of high-dimensional vectors, adapts to the encryption requirements of CKKS for polynomial ring elements, while preserving the dimension structure of the vector, supporting subsequent homomorphic vector operations;

[0097] S5-5: Encrypt the 512-dimensional feature vector v by CKKS first, and then add nonlinear noise through PolyMask:

[0098] venc=PolyMask(CKKS-Enc(v),k,M)

[0099] PolyMask calculation formula:

[0100] PolyMask(C,k,M)=(Ck)modM

[0101] Where:

[0102] C: Original ciphertext (such as CKKS encrypted feature vector);

[0103] k: Dynamically generated polynomial index (positively related to data sensitivity);

[0104] M: Modulus of encryption scheme (determined by key parameters).

[0105] Even if the attacker cracks the outer CKKS, he still needs to solve the polynomial discrete logarithm problem to obtain the original data; The integration mode of CKKS: PolyMask as the second layer of encryption, superimposed with CKKS linear homomorphism; Facet feature mapping accelerator (processing speed increased to 18GB / s), quantum attack resistant module (based on NTRU algorithm, responding to the traffic fuse requirement of document S9).

[0106] S6: The AI server cluster is connected with the expert review terminal through the STM32F407 chip controlled one-way data gate, only allowing the one-way transmission of the encrypted polynomial mask, CKKS+(PolyMask) double encryption to the expert review terminal at a rate of ≤1000 data packets per second, and the abnormal traffic automatically triggers the fuse mechanism.

[0107] S7: The expert review terminal monitors the appraiser load state in real time through the RedisZSET dynamic allocator, determines the task dispatching order according to the priority calculation formula, and automatically reallocates the task to other appraiser terminals if it is not responded for more than 30 minutes.

[0108] S8: If the user applies for review, the system assigns the task to the idle appraiser through dynamic task allocation; the appraiser uses the microscopic video acquisition device to verify the jewelry features; the blockchain storage module is based on the Tencent cloud TBaaS platform, and the appraisal process screen recording file, operation log and feature labeling information are solidified. The judicial evidence special storage chain: (1) AI real-time verification storage (automatic comparison of microscopic video and feature vector), (2) judicial agency node access (courts / quality inspection bureaus can verify in real time, see task allocation process Figure 3 ).

[0109] S9: The encrypted communication module monitors the transmission traffic in real time, and automatically triggers traffic restriction when detecting that the request frequency of a single IP address is >100 times / second; when the data packet loss rate is >0.1%, switch to the TLS1.3+SM3 backup encryption channel to ensure communication security.

[0110] In step S1, the user takes a picture of the facet of the jewelry to be identified through the mobile application, and the picture resolution needs to reach or exceed the 4K standard to ensure that the details of the jewelry surface are clear and identifiable, providing high-quality original data for subsequent feature extraction.

[0111] In step S2, the jewelry picture taken by the user is transmitted securely through a double-layer encryption channel at the transmission layer, where the SM4 national encryption algorithm (group length 128 bits) and the TLS1.3 protocol are used together to build an encryption mechanism, and the storage layer key is dynamically generated and managed by the Tencent cloud KMS management system. The rotation period of the key is strictly controlled within 24 hours to reduce the risk of key leakage and ensure the security of data transmission.

[0112] In step S3, the feature vector generator in the AI server cluster deploys the multi-modal fusion model EfficientNet-B4+ViT-S to process the encrypted jewelry picture data, and expands the input data of the facet intersection feature enhancement layer (FacetEnhancer) : encrypted image + facet intersection spatial coordinates (from the microscopic video acquisition device); a feature vector with a dimension of 512 is generated; according to the actual measurement of the Kirin 9000S chip, the time consumed by the mobile terminal to generate the feature vector does not exceed 0.8 seconds; if the confidence of the generated feature vector reaches or exceeds 95%, the system will automatically trigger the report generation process to quickly complete the preliminary identification.

[0113] In step S4, when the user chooses to participate in data co-construction, the jewelry picture provided by the user needs to be processed through differential privacy, specifically by adding Laplace noise (noise mean μ = 0.1, standard deviation σ = 0.05) to achieve data desensitization; the training data pool is protected by differential privacy parameters ε = 2 and δ = 1e-5 to prevent the original data of the user from being reversed; the federal learning training unit starts incremental training at 02:00 every day, and the generated global model parameter update file has a size of not more than 500MB, and is synchronized to at least 100 edge nodes through the HTTPS protocol, ensuring the continuous optimization and sharing of the model.

[0114] In step S5, the homomorphic encryption chip (model HES-2200) in the AI server cluster uses the CKKS homomorphic encryption scheme to encrypt the generated polynomial mask, CKKS+(PolyMask) double encryption, the ciphertext processing speed of this chip is not less than 12GB / s, ensuring that the server can only operate on encrypted feature vectors, and protecting the privacy of the original data from the hardware level.

[0115] In step S6, the AI server cluster and the expert review terminal are connected through a one-way data gate controlled by an STM32F407 chip, which only allows the encrypted polynomial mask, CKKS+(PolyMask) double encryption (after encryption, the length is 1024 bytes) to be transmitted from the AI server cluster to the expert review terminal in one direction, and the transmission rate is limited to not more than 1000 data packets per second; if abnormal traffic is detected, the system will automatically trigger the fuse mechanism to block abnormal data transmission.

[0116] In step S7, the expert review terminal monitors the load state of the appraiser in real time through the Redis ZSET queue, and the monitoring indicators include task backlog, response delay, and historical accuracy rate, based on these indicators, the task priority is dynamically calculated and the dispatching order is adjusted; if the appraiser does not respond to the assigned task within 30 minutes, the system will automatically reassign the task to other idle appraiser terminals to ensure the efficiency of the review.

[0117] The calculation formula used for task priority calculation is:

[0118]

[0119] Task timeout mechanism: tasks that are not responded to for more than 30 minutes are automatically reassigned to other appraiser terminals.

[0120] In step S8, the system assigns the task to an idle appraiser through a dynamic task allocation mechanism; the appraiser performs manual verification using a microscopic video acquisition device, which integrates a 40x optical zoom lens, a 30fps frame rate sensor with a resolution of 3840x2160, and a facet intersection recognition accelerator (recognition accuracy ±0.01mm), for observing details such as jewelry inclusions distribution; during the appraisal process, the blockchain storage module based on Tencent Cloud TBaaS platform, solidifies the appraisal process screen recording file, operation log (records magnification, annotation and comparison operations) and feature annotation information (timestamp accurate to millisecond level, stores facet intersection coordinates, inclusions distribution and risk level annotation results in JSON format), ensures that the appraisal record cannot be tampered with;

[0121] In step S9, the encrypted communication module monitors the transmission traffic in real time, and when the request frequency of a single IP address exceeds 100 times per second, the system automatically triggers traffic limiting measures; if the data packet loss rate exceeds 0.1%, switch to TLS1.3+SM3 backup encryption channel for continuous transmission, maintain the stability and security of data transmission through dual-channel protection mechanism.

[0122] Example 1: AI jewelry identification process:

[0123] In the user upload link, when the user operates through the mobile terminal application, the interface will provide a shooting guide function - a circular focus frame is displayed in the center of the screen, and prompt texts such as "keep the light uniform" and "avoid glare" are marked around it, to ensure that the user's diamond facet picture resolution reaches the 4K standard (3840x2160 pixels), clearly presenting details such as facet edges and surface texture. After the picture is taken, the system automatically calls the transmission layer encryption module: first, use the SM4 national encryption algorithm to group encrypt the picture data (128 bits per group), generate ciphertext data packets; then establish a secure transmission channel through the TLS1.3 protocol, encapsulate the ciphertext data packets into TLS encrypted frames, and transmit them to the AI server cluster via the mobile network. This process does not require the user to manually operate the encryption steps, and only needs to click the "upload" button to complete.

[0124] The feature extraction link is performed by a feature vector generator in the AI server cluster. The EfficientNet-B4 model deployed by the generator is optimized for lightweight, and before input, the encrypted picture is preprocessed: first restored to plaintext by the decryption module (only temporarily stored in memory), then adjusted to the model-adapted 224x224 pixel size, and normalized (mean [0.485, 0.456, 0.406], standard deviation [0.229, 0.224, 0.225]). The model extracts features layer by layer through convolutional layers, pooling layers, and fully connected layers, and finally outputs a polynomial mask, double encrypted by CKKS+ (PolyMask). The system has a built-in confidence evaluation mechanism, which calculates the cosine similarity between the feature vector and the standard jewelry feature library (threshold set to 0.95). If it reaches or exceeds the threshold, the system automatically calls the report generation module to generate an electronic report containing the jewelry type (such as diamond, sapphire), main features (facet number, edge regularity), and preliminary identification conclusion. Users can view the report in real-time within the application.

[0125] The triggering condition for the artificial review link is that the user actively selects "Apply for professional review" or the system detects that the feature vector confidence is less than 95% (requires human assistance to confirm). The system monitors the online status and task load of the appraiser terminal through the Redis queue, and preferentially assigns it to the appraiser with the least current task load and the highest historical review accuracy. After the appraiser logs in to the review terminal, the user uploads the original picture and the AI-generated feature vector, and verifies it through a microscopic video capture device: adjust the 40x optical zoom lens to the appropriate focal length, start the high-definition sensor at 30 frames per second, and real-time capture microscopic images of the diamond surface (resolution 3840x2160), observe the distribution pattern (such as cloud-like, needle-like), location (internal / near surface), and size (measured by facet intersection identification accelerator, accuracy ±0.01 mm). During the verification process, the system synchronously records the appraiser's operation screen (including lens adjustment, use of marking tools, etc.), and records the timestamp (accurate to milliseconds) of each step of operation. After verification is complete, the appraiser needs to input the review conclusion (such as "consistent with natural diamond features" or "suspected synthetic") on the terminal, and the system combines the AI preliminary conclusion and the review conclusion to generate the final identification report, which is pushed to the user through the application and stored in the blockchain storage module to ensure that the results are traceable.

[0126] Example 2: Federal learning data co-construction:

[0127] The data desensitization link is triggered after the user selects "participate in data co-construction" and checks the "data sharing agreement". The system first preprocesses the jewelry image uploaded by the user: identifies the effective area in the image (excludes background interference), and extracts the jewelry main part; then applies differential privacy technology for desensitization processing - according to the distribution range of image pixel values (0-255), the amount of Laplace noise added is calculated (mean μ = 0.1, standard deviation σ = 0.05), and a random number generator is used to add noise values conforming to the distribution to each pixel point, so that the original pixel value is slightly disturbed (the disturbance amplitude is not more than ±0.5). This processing retains the overall features of the image (such as facet shape, texture direction), while effectively avoiding the reverse restoration of user privacy information (such as shooting location, device model) through pixel values. The desensitized image is stored on the user's local device and uploaded to the AI server cluster's training data pool through a secure channel only during model training; the data pool uses differential privacy parameters ε = 2, δ = 1e-5 for global protection, ensuring that the leakage of any single data does not substantially affect the overall data privacy.

[0128] The model training link uses a federated learning framework, and the incremental training process is automatically started at 02:00 every day. First, the AI server cluster randomly selects 10% of the desensitized images (about 50,000) from the training data pool and distributes them to more than 100 edge nodes distributed throughout the country (including jewelry retail store terminals, third-party identification agency servers, etc.); each edge node performs local training based on the locally stored historical model parameters (the global model of the previous day) and the newly received desensitized data (5 iterations, learning rate 0.001), generating updated local model parameters. Subsequently, the server cluster collects all edge node local parameters through an HTTPS secure channel (encrypted using the TLS1.3 protocol, and the certificate is issued by a trusted CA authority), and uses the federated averaging algorithm (FedAvg) for aggregation: calculates the weighted average of each node's parameters (the weight is determined according to the proportion of the node's training data), and generates new global model parameters (volume about 480MB). The new model parameters are synchronized to all edge nodes through HTTPS, and each node automatically downloads and replaces the local model to complete the model update. Throughout the process, the user only needs to ensure that the device is connected to the network, and no manual operation is required; the data always participates in the training in a desensitized form, and the server cannot obtain the original image content, thereby balancing the user data privacy and model training efficiency.

[0129] Example 3: Forensic Scene

[0130] 1. The user uploads the case-related jewelry image, and the system automatically marks it as "high sensitivity" ($\varepsilon = 1.0$)

[0131] 2. The microscopic video acquisition device starts the judicial mode:

[0132] 1) 40 times lens automatic scanning facet code (such as GIA certificate number),

[0133] 2) Chain of custody synchronous upload to court node (response time <2 seconds).

[0134] 3. Generate a judicial special report:

[0135] 1) Contains AI and 3 independent signatures of appraisers,

[0136] 2) Blockchain storage hash value is automatically embedded in judicial documents.

[0137] Test data: compare traditional judicial appraisal cycle of 7 days vs. 12 hours of the system.

[0138] From the above analysis, it can be seen that:

[0139] In the present application, efficient collaboration of the appraisal process is first realized. Traditional jewelry appraisal relies on manual operation, which requires a lot of time and the help of professional instruments. However, the method can complete preliminary analysis in a short time through the rapid feature extraction capability of the AI server cluster, and automatically generate a report if high confidence features are detected. At the same time, in combination with the expert review mechanism, the system dynamically allocates idle appraisers for manual verification, while retaining the accuracy of manual appraisal, the overall process time is significantly shortened through technical means, forming an efficient collaboration mode of "AI rapid processing + manual accurate review", which meets the timeliness requirements of jewelry retail, insurance loss assessment and other scenarios.

[0140] In the present application, a full-link data privacy protection system is constructed. In view of the potential risk of revealing the user data in clear text transmission and centralized training in traditional automatic appraisal system, the system uses multi-dimensional technology to ensure privacy security: in the transmission process, SM4 national encryption algorithm and TLS1.3 protocol are used to construct double-encryption channels, and the storage layer key is periodically replaced by a dynamic management system; only encrypted feature vectors are operated in the feature processing link, and differential privacy technology is used to desensitize the user data involved in model co-construction, so that the original information cannot be reversed. The above measures block the privacy leakage path from data collection, transmission to processing, which meets the requirements of data security related regulations.

[0141] In the present application, the credibility and authority of the identification result are strengthened. The traditional AI identification is difficult to be used as judicial evidence due to the lack of traceability of process records; and the system solidifies the storage of key information (such as operation logs, feature labeling results) of the identification process through the microscopic video verification of the expert review terminal and the block chain storage module, and the time stamp is accurate to millisecond level and the record is tamper-proof. At the same time, the cooperative mechanism of artificial review and AI analysis not only plays the advantage of fast processing of AI, but also supplements the limitations of AI through the experience of professional appraisers, so that the identification conclusion is more reliable and can meet the needs of scenes with high requirements for authority such as judicial evidence and quality certification.

[0142] It should be particularly noted that the relational terms such as "first", "second" and the like used herein are only intended to distinguish different entities or operations from each other, and do not necessarily require or imply any actual relationship or chronological order between these entities or operations. In addition, the term "comprising" or any variant thereof means non-exclusive inclusion, i.e. a process, method, article or device including a series of elements, which includes not only the explicitly listed elements, but also other elements not explicitly listed or inherent to the process, method, article or device. Without further limitation, the elements defined by "including a" and the like do not exclude the presence of other identical elements in the process, method, article or device containing the elements.

[0143] The above examples are only used to illustrate the technical solutions of the present application and do not constitute a limitation thereof; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that it can modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part of the technical features; such modification or replacement does not cause the essence of the corresponding technical solution to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A multimodal encrypted jewelry authentication system and interaction method based on a federated learning framework, characterized by: The system includes: an AI server cluster, an expert review terminal, an encrypted communication module, and a one-way data gate; The AI ​​server cluster is responsible for encrypting, extracting features from, and training models for jewelry data, and includes the following core units: Federated learning training unit: It adopts a jewelry feature-aware Transformer-CNN hybrid architecture, supports dynamic incremental training based on jewelry market value fluctuations (e.g., triggering emergency training when gold price fluctuations > 5%) and generates global model parameter update files; Feature vector generator: Deploys the multimodal fusion model EfficientNet-B4+ViT-S, with encrypted jewelry image data as input and output polynomial mask, double encrypted with CKKS+ (PolyMask); when the multimodal fusion confidence is ≥98%, a report is triggered and a 3D topology map of micro-features is attached. Homomorphic encryption chip: JadeCrypto-1, a custom-designed encryption chip for jewelry authentication; The expert review terminal is used for manual review and evidence storage of appraisal records, and includes the following modules: Microscopic video acquisition device: integrates a 40x optical zoom lens, a 3840×2160 resolution 30fps frame rate sensor and a faceted intersection recognition accelerator, used to verify jewelry features; Blockchain Evidence Storage Module: Based on Tencent Cloud TBaaS platform, it permanently stores evidence files (including confidence conflict markers), operation logs and feature annotation information through both AI and human channels, and adopts SHA-256 hash algorithm; The encrypted communication module is used to ensure the security of data transmission and storage, and includes the following modules: Transport layer encryption: A two-layer encryption channel is constructed using the SM4 national cryptographic algorithm and the TLS 1.3 protocol; Storage layer key management: Keys are dynamically generated by the Tencent Cloud KMS management system; Abnormal traffic circuit breaker mechanism: When a single IP address is detected to request more than 100 times per second, traffic limiting is triggered; when the packet loss rate is greater than 0.1%, the system switches to a TLS 1.3+SM3 backup encrypted channel. The one-way data gate is deployed between the AI ​​server cluster and the expert review terminal. The AI ​​server cluster and the expert review terminal are connected through the one-way data gate controlled by the STM32F407 chip, which only allows the encrypted feature vector to be transmitted unidirectionally from the AI ​​server cluster to the expert review terminal. Jewelry category weighting module (GemWeighting): The formula is: $$\omega_c=\frac{\log(\text{value}_c+1)}{\sum_{c\in C}\log(\text{value}_c+1)}$$ Where $c$ represents the jewelry category (diamonds / sapphires, etc.), and $\text{value}_c$ represents the historical average value.

2. The multimodal encrypted jewelry authentication system and interaction method based on a federated learning framework as described in claim 1, characterized in that: The method includes the following steps: S1: The user takes an image of the faceted jewelry using a mobile application, ready for authentication; S2: The jewelry images uploaded by users are transmitted to the AI ​​server cluster through a two-layer encrypted channel built with the SM4 national cryptographic algorithm and the TLS1.3 protocol. The transmission layer key is dynamically generated by the Tencent Cloud KMS management system. S3: The AI ​​server cluster's feature vector generator deploys the multimodal fusion model EfficientNet-B4+ViT-S to process the encrypted jewelry image data, generating a polynomial mask with CKKS+ (PolyMask) double encryption; if the multimodal fusion confidence level is ≥98%, automatic report generation is initiated. S4: When a user chooses to participate in data co-construction, the jewelry image is processed for differential privacy using the Laplacian noise mechanism, and the training data pool is protected according to the differential privacy parameters ε=2 and δ=1e-5; incremental training starts at 02:00 every day, the federated learning training unit generates a global model parameter update file, and synchronizes it to no less than 100 edge nodes via HTTPS protocol; The S5:AI server cluster's homomorphic encryption chip uses the CKKS homomorphic encryption scheme to encrypt the generated polynomial mask with CKKS+ (PolyMask) double encryption, ensuring that the server only operates on the encrypted feature vector. The homomorphic encryption chip in the AI ​​server cluster uses the CKKS scheme for multinomial masking and CKKS+ (PolyMask) double encryption. The specific process is as follows: S5-1: Key Generation: The public key (pk), private key (sk), and evaluation key (evk) required by the CKKS algorithm are dynamically generated through the Tencent Cloud KMS management system. The public key is used for encryption, the private key is used for decryption, and the evaluation key supports subsequent homomorphic operations (such as feature vector similarity comparison). S5-2: Feature Vector Encoding: Encode the 512-dimensional real-valued feature vector v = (v1, v2, ..., v...). 512 Mapping to the complex field polynomial ring Z[X] / (X^ N In +1), the vector elements are embedded into the polynomial coefficients through the "slot filling" technique to form the plaintext polynomial m(X); S5-3: Encryption process: Use the public key to perform an encryption operation on the plaintext polynomial m(X) to generate ciphertext C=(a(X),b(X)); The ciphertext components satisfy b(X)=a(X)·s(X)+e(X)+m(X)·Δ; Where s(X) is the private key polynomial, e(X) is Gaussian noise, and Δ is a scaling factor to ensure the accuracy of the real value; S5-4: Homomorphic operation support: The encrypted feature vector ciphertext can be directly used in the model parameter update in federated learning or the feature comparison of the expert review terminal. It can complete operations such as inner product and addition without decryption, preventing the leakage of original data. The polynomial encoding formula for a 512-dimensional feature vector is as follows: In the formula: each component vk of the 512-dimensional feature vector v is encoded as the coefficient of the polynomial m(X), where X^ {k-1} Using a polynomial basis, this encoding method achieves a compact representation of high-dimensional vectors, adapts to CKKS's encryption requirements for polynomial ring elements, and preserves the dimensional structure of the vector, supporting subsequent homomorphic vector operations. S5-5: First, CKKS encryption is applied to the 512-dimensional feature vector v, and then nonlinear noise is added using PolyMask: venc=PolyMask(CKKS-Enc(v),k,M) PolyMask calculation formula: PolyMask(C,k,M) = (Ck)modM in: C: Original ciphertext (such as the feature vector after CKKS encryption); k: Dynamically generated polynomial exponent (positively correlated with data sensitivity); M: The modulus of the encryption scheme (determined by the key parameters); Even if attackers crack the outer CKKS layer, they still need to solve the polynomial discrete logarithm problem to obtain the original data; the integration method with CKKS is: PolyMask as the second layer of encryption, which is linearly homomorphically superimposed with CKKS; faceted feature mapping accelerator (processing speed increased to 18GB / s), and quantum attack resistance module (based on the NTRU algorithm, responding to the traffic circuit breaking requirements of document S9); S6: The AI ​​server cluster and the expert review terminal are connected via a one-way data gate controlled by the STM32F407 chip. Only encrypted polynomial mask and CKKS+ (PolyMask) double encryption are allowed to be transmitted unidirectionally to the expert review terminal at a rate of no more than 1,000 data packets per second. Abnormal traffic will automatically trigger the circuit breaker mechanism. S7: The expert review terminal monitors the appraiser's load status in real time through the Redis ZSET dynamic allocator and determines the task dispatch order according to the priority calculation formula; tasks that have not responded for more than 30 minutes are automatically reassigned to other appraiser terminals. S8: If a user applies for review, the system will assign the task to an available appraiser through dynamic task allocation; the appraiser uses a microscopic video acquisition device to verify the jewelry features; the blockchain evidence storage module is based on the Tencent Cloud TBaaS platform and permanently stores AI and manual dual-channel evidence storage files (including confidence conflict markers), operation logs and feature annotation information; the judicial evidence collection dedicated evidence storage chain: (1) AI real-time verification of evidence storage (automatic comparison of microscopic video and feature vector), (2) access of judicial institution nodes; S9: The encrypted communication module monitors the transmission traffic in real time. When it detects that the frequency of a single IP address request exceeds 100 times / second, it automatically triggers traffic limiting. When the packet loss rate exceeds 0.1%, it switches to the TLS1.3+SM3 backup encrypted channel to ensure communication security.

3. The multimodal encrypted jewelry authentication system and interaction method based on a federated learning framework as described in claim 2, characterized in that: In step S1, the user takes a facet image of the jewelry to be appraised using a mobile application. The image resolution must be no less than 4K to ensure that the details on the surface of the jewelry are clearly distinguishable, providing high-quality raw data for subsequent feature extraction.

4. The multimodal encrypted jewelry authentication system and interaction method based on a federated learning framework as described in claim 2, characterized in that: In step S2, the jewelry images taken by the user are securely transmitted through a dual-layer encryption channel in the transmission layer. The transmission layer uses the SM4 national cryptographic algorithm and the TLS1.3 protocol to jointly construct an encryption mechanism. The storage layer key is dynamically generated and managed by the Tencent Cloud KMS management system, and the key rotation cycle is strictly limited to no more than 24 hours to reduce the risk of leakage and ensure the security of data transmission.

5. The multimodal encrypted jewelry authentication system and interaction method based on a federated learning framework as described in claim 2, characterized in that: In step S3, the feature vector generator of the AI ​​server cluster deploys the multimodal fusion model EfficientNet-B4+ViT-S to process the encrypted jewelry image data and expand the input data of the FacetEnhancer layer: encrypted image + facet intersection spatial coordinates (from the microscopic video acquisition device); generating a feature vector with a dimension of 512; according to the actual test of the Kirin 9000S chip, the time taken for the mobile device to generate this feature vector is no more than 0.8 seconds.

6. The multimodal encrypted jewelry authentication system and interaction method based on a federated learning framework as described in claim 2, characterized in that: In step S4, when a user chooses to participate in data co-construction, the jewelry images they provide need to undergo differential privacy processing, specifically by adding Laplace noise to achieve data desensitization; the training data pool uses differential privacy parameters of ε=2 and δ=1e-5 for protection.

7. The multimodal encrypted jewelry authentication system and interaction method based on a federated learning framework as described in claim 2, characterized in that: In step S5, the homomorphic encryption chip of the AI ​​server cluster implements the CKKS homomorphic encryption scheme to encrypt the generated polynomial mask with CKKS+ (PolyMask) double encryption. The chip's ciphertext processing speed is no less than 12GB / s, ensuring that the server can only process the encrypted feature vector, thus protecting the privacy of the original data from the hardware level.

8. The multimodal encrypted jewelry authentication system and interaction method based on a federated learning framework as described in claim 2, characterized in that: In step S6, the AI ​​server cluster and the expert review terminal are connected through a one-way data gate controlled by the STM32F407 chip, allowing only the encrypted polynomial mask and CKKS+ (PolyMask) double encryption to be transmitted unidirectionally from the AI ​​server cluster to the expert review terminal.

9. The multimodal encrypted jewelry authentication system and interaction method based on a federated learning framework as described in claim 2, characterized in that: In step S7, the expert review terminal monitors the appraiser's load status in real time through the Redis ZSET queue. The monitoring indicators include task backlog, response latency and historical accuracy. Based on the above indicators, the task priority is dynamically calculated to adjust the dispatch order. If an appraiser fails to respond to the assigned task within 30 minutes, the system will automatically reassign the task to another available appraiser terminal to ensure review efficiency. The formula used to calculate task priority is as follows: Task timeout mechanism: If a task does not receive a response for more than 30 minutes, it will be automatically reassigned to another appraiser terminal.

10. The multimodal encrypted jewelry authentication system and interaction method based on a federated learning framework as described in claim 2, characterized in that: In step S8, the system assigns tasks to available appraisers through a dynamic task allocation mechanism. The appraisers use a microscopic video acquisition device to perform manual verification. This device integrates a 40x optical zoom lens, a 3840×2160 resolution 30fps frame rate sensor, and a faceted intersection recognition accelerator to observe the details of the jewelry inclusions. During the appraisal process, the blockchain evidence storage module, based on the Tencent Cloud TBaaS platform, solidifies the screen recording files, operation logs, and feature annotation information of the appraisal process to ensure that the appraisal records are tamper-proof. In step S9, the encrypted communication module monitors the transmission traffic in real time. When it detects that the frequency of a single IP address request exceeds 100 times / second, the system automatically triggers traffic restriction measures. If the packet loss rate exceeds 0.1%, the system will switch to a TLS1.3+SM3 backup encrypted channel to continue transmission, ensuring the stability and security of data transmission through a dual-channel protection mechanism.

Images