A lightweight federated learning palm vein pad anti-counterfeiting method and system

Abstract

The application discloses a kind of lightweight federated learning's palm vein PAD anti-fake method and system, wherein the method is: collecting palm vein image and pre-processing, obtain three kinds of feature fragments of blood vessel topology, blood flow velocity and texture details and encryption;Obtain environmental parameters, align the ciphertext domain to encrypted fragments;Asymmetric federation architecture is constructed, and the model is iteratively updated to convergence by federated collaborative training;Collaborative decryption fragments after model convergence, generate complete palm vein characteristics by matching weighted fusion;Operation log and algorithm parameters are recorded in each node, traceability compliance verification is performed, and tamper-proof traceability link is generated.The application combines layered weighted homomorphic encryption, environment adaptive alignment and asymmetric federated collaborative updating technology, realizes the secure encryption transmission and lightweight training of palm vein data, balances anti-fake accuracy and terminal adaptability, simultaneously improves the compliance and traceability of the method through whole-process traceability verification, and is suitable for various PAD terminal palm vein anti-fake scenarios.

Description

Technical Field

[0001] This invention relates to the field of biometric anti-counterfeiting technology, specifically to a lightweight federated learning method and system for palm vein PAD anti-counterfeiting. Background Technology

[0002] Biometric technology, with its uniqueness and convenience, has been widely used in the field of identity authentication. Among them, palm vein recognition, with its inherent advantages of liveness detection, difficulty in forgery, and high user acceptance, has become an important development direction for anti-counterfeiting authentication technology. With the popularization of mobile terminals and IoT technology, PAD devices, as portable identity verification carriers, place higher demands on the lightweight deployment, data security protection, and cross-scenario adaptability of palm vein anti-counterfeiting systems. Federated learning technology, through a distributed training mode, achieves data usability without visibility, providing an effective path to resolve the contradiction between cross-device data sharing and privacy protection. Combining it with palm vein recognition technology can improve the generalization ability of anti-counterfeiting models while ensuring the security of users' biometric information, meeting the actual needs of multi-terminal collaborative anti-counterfeiting, and has become an important research hotspot in the current field of biometric anti-counterfeiting.

[0003] Traditional palm vein anti-counterfeiting technologies have many limitations and are difficult to adapt to the needs of complex application scenarios. Some technologies adopt a centralized data processing mode, which requires transmitting the user's original palm vein data to a central server for training and recognition. This not only increases the risk of leakage during data transmission, but also poses privacy and security risks due to centralized data storage. Other technologies do not effectively decompose and encrypt palm vein features, making feature information easy to be stolen or tampered with, resulting in insufficient anti-counterfeiting reliability. At the same time, traditional systems mostly adopt a unified model architecture, which does not take into account the differences in computing resources of terminal devices. It is difficult to achieve a balance between lightweight deployment and recognition accuracy, and they lack the ability to adapt to environmental factors. Feature alignment deviations are prone to occur under different environmental conditions, affecting recognition accuracy. In addition, most technologies have not established a sound compliance traceability mechanism, and the entire data processing process lacks effective supervision, making it difficult to meet the relevant data security regulations. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides a lightweight federated learning method for palm vein PAD anti-counterfeiting. This method can improve the accuracy and stability of anti-counterfeiting in different scenarios, while meeting compliance and traceability requirements, and providing secure and efficient technical support for various identity authentication scenarios.

[0005] Another objective of this invention is to provide a lightweight, federated learning-based palm vein PAD anti-counterfeiting system.

[0006] The technical solution of this invention to solve the above-mentioned technical problems is:

[0007] A lightweight federated learning-based anti-counterfeiting method for palm vein PADs includes the following steps:

[0008] Step 1: Acquire palm vein images and preprocess them to obtain three types of feature fragments: vascular topology, blood flow velocity, and texture details; assign unique identifiers to each type of feature fragment and establish associations; use a hierarchical weighted fully homomorphic encryption algorithm to encrypt the feature fragments to obtain encrypted feature fragment combinations;

[0009] Step 2: Obtain the environmental parameters during palm vein image acquisition, and call the environment adaptive ciphertext field alignment algorithm to perform ciphertext field alignment calculation on the encrypted fragments in the encrypted feature fragment combination to obtain the aligned encrypted features;

[0010] Step 3: Construct a federated collaborative training model, employing an asymmetric federated splitting architecture comprising a central server and multiple PAD terminals for training. This involves deploying an anti-counterfeiting backbone network on the central server and lightweight adapters on each PAD terminal. Each PAD terminal receives the aligned encrypted features output from Step 2, uses these features to train its own lightweight adapter, and uploads the encrypted parameter update data to the central server. The central server aggregates the encrypted parameter update data uploaded by multiple PAD terminals, fuses it with the parameter update data of the anti-counterfeiting backbone network using an asymmetric federated collaborative update algorithm, generates a global parameter update, and synchronously distributes it to update the anti-counterfeiting backbone network of the central server and the lightweight adapters of each PAD terminal. This training process is repeated until the federated collaborative training model converges.

[0011] Step 4: After the federated collaborative training model converges, based on the threshold decryption mechanism, the central server and each PAD terminal collaboratively decrypt the encrypted feature fragment combination generated in Step 1; the decrypted three types of feature fragments—vascular topology, blood flow velocity, and texture details—are matched according to their unique identifiers and weighted and fused according to preset weights to generate complete palm vein features.

[0012] Step 5: Record the operation logs and algorithm parameters of Steps 1 to 4 in real time on the central server and each PAD terminal, perform traceability compliance verification, and generate an unalterable responsibility traceability link.

[0013] Preferably, in step 1, the preprocessing includes denoising, grayscale correction, and feature fragmentation, specifically as follows:

[0014] A 3×3 computational template is used to denoise the acquired palm vein image. The pixel weights in the computational template are distributed according to the rule of decreasing from the center to the edge, and the weight decay range is set between 0.8 and 1.2.

[0015] The denoised palm vein image is then subjected to grayscale correction by adjusting pixels with grayscale values ​​less than 10 to 10, pixels with grayscale values ​​greater than 245 to 245, and linearly mapping the grayscale values ​​of the remaining pixels to the range of 0 to 255.

[0016] The grayscale-corrected palm vein image was divided into three types of feature fragments. The pixel regions of the vascular skeleton and branch nodes were extracted as vascular topological fragments, the regions of dynamic changes in pixel grayscale values ​​within the blood vessel were extracted as blood flow velocity fragments, and the regions of texture and grayscale gradient changes around the blood vessel were extracted as texture detail fragments. The pixel proportions of vascular topological fragments, blood flow velocity fragments, and texture detail fragments were 65%, 20%, and 15%, respectively. Finally, a 256-bit unique identifier was generated for each of the vascular topological fragments, blood flow velocity fragments, and texture detail fragments to establish the association between them.

[0017] Preferably, in step 3, the anti-counterfeiting backbone network of the central server includes a convolutional layer, a pooling layer, and a fully connected layer; the lightweight scene adapter of the PAD terminal includes two fully connected layers.

[0018] Preferably, in step 3, the specific process of generating the global parameter update is as follows:

[0019] After receiving the encrypted parameter update data uploaded by each PAD terminal, the central server performs an integrity check on the encrypted parameter update data. After the check passes, the encrypted parameter update data is decrypted to obtain the plaintext parameter update data corresponding to each PAD terminal.

[0020] The asymmetric federated collaborative update algorithm is invoked to perform a weighted fusion calculation of the plaintext parameter update amounts of each decrypted PAD terminal and the plaintext parameter update amounts of the central server's anti-counterfeiting backbone network. During the calculation process, the federated collaborative weight coefficients in the asymmetric federated collaborative update algorithm are dynamically adjusted so that the weight ratio of the parameter update amounts of the central server's anti-counterfeiting backbone network and the parameter update amounts of the PAD terminals is optimized in real time according to the global training error to generate the global parameter update amount.

[0021] The central server synchronously distributes the global parameter updates in encrypted form to its own anti-counterfeiting backbone network and the lightweight adapters of all PAD terminals. The central server's anti-counterfeiting backbone network completes global parameter iteration based on the global parameter updates, while the lightweight adapters of each PAD terminal complete local parameter iteration based on the global parameter updates, thereby completing this round of collaborative training. If convergence is not achieved, a new round of collaborative training is triggered.

[0022] Preferably, in step 4, the specific steps of the threshold decryption mechanism are as follows:

[0023] Decryption thresholds of 0.8, 0.6, and 0.5 are set for vascular topology fragments, blood flow velocity fragments, and texture detail fragments, respectively. During decryption, the similarity between the decryption result and the preset template is calculated. If the similarity is higher than the corresponding threshold, the decryption is considered successful. If the similarity is lower than the corresponding threshold, a decryption retry is triggered. The number of decryption retryes shall not exceed 3.

[0024] The encrypted fragments adopt a fragmented distributed storage. The central server stores 40% of the total number of vascular topology fragments, 30% of the total number of blood flow velocity fragments, and 30% of the total number of texture detail fragments. The remaining feature fragments are distributed and stored in each PAD terminal, and each PAD terminal must not store complete data of any type of feature fragment.

[0025] Preferably, in step 4, the specific process of matching and weighted fusion of the decrypted vascular topology fragments, blood flow velocity fragments, and texture detail fragments is as follows:

[0026] The character sequences of the decrypted vascular topology fragments, blood flow velocity fragments, and texture detail fragments are extracted. The consistency of each character sequence is compared, and interrelated vascular topology fragments, blood flow velocity fragments, and texture detail fragments are filtered out. After matching, a weighted average algorithm is used to fuse the vascular topology fragments, blood flow velocity fragments, and texture detail fragments. The fusion weight is consistent with the fragment alignment weight in the environment adaptive ciphertext field alignment algorithm, that is, the weight of vascular topology fragments is 0.5, the weight of blood flow velocity fragments is 0.3, and the weight of texture detail fragments is 0.2.

[0027] During the fusion calculation, a weighted summation operation is performed on the corresponding pixel positions of vascular topology fragments, blood flow velocity fragments, and texture detail fragments to obtain the fused pixel grayscale values. After fusion, the generated palm vein feature information is normalized to map all pixel grayscale values ​​to the [0,1] interval, and abnormal pixels with grayscale values ​​exceeding the [0,1] interval or with a difference greater than 0.2 from the grayscale values ​​of surrounding pixels are removed.

[0028] Preferably, in step 5, the operation log records the operation time, operation node number, parameter settings, and data flow of steps 1 to 4, wherein the operation time recording accuracy needs to be accurate to milliseconds; the operation node number uses a unique numerical code to identify all nodes in the system; the recorded algorithm parameters include the α value, key parameters, and encryption depth of the hierarchical weighted fully homomorphic encryption algorithm, and the weights of the environment-adaptive ciphertext field alignment algorithm. The transformation coefficients, as well as the β value, learning rate, and parameter update amount of the asymmetric federated collaborative update algorithm; the specific steps of the traceability compliance verification are as follows: in steps 1 to 4, verification nodes are set up respectively, and the verification dimensions cover data format compliance, parameter value validity, and operation permission legality; if any step fails the verification, the current operation is immediately terminated, the abnormal information is recorded synchronously and reported to the central server; the abnormal information includes the abnormal step, the abnormal reason and the corresponding operation node number.

[0029] A lightweight federated learning-based palm vein PAD anti-counterfeiting system includes:

[0030] The palm vein feature encryption module is used to acquire palm vein images and perform preprocessing to obtain three types of feature fragments: blood vessel topology, blood flow velocity, and texture details. It assigns unique identifiers to each type of feature fragment and establishes associations. It then calls a hierarchical weighted fully homomorphic encryption algorithm to encrypt the feature fragments, resulting in a combination of encrypted feature fragments, which is then output.

[0031] The encrypted domain feature alignment module is used to receive the combination of encrypted feature fragments output by the palm vein feature encryption module, obtain the environmental parameters when the palm vein image is acquired, call the environment adaptive encrypted domain alignment algorithm to perform encrypted domain alignment calculation on the encrypted fragments in the combination of encrypted feature fragments, and output the aligned encrypted features.

[0032] The asymmetric federated training module includes a central server unit and multiple PAD terminal units. The central server unit is deployed with an anti-counterfeiting backbone network. Each PAD terminal unit is deployed with a lightweight adapter head. Each PAD terminal unit receives aligned encrypted features, uses these features to train its own lightweight adapter head, and uploads the encrypted parameter update data to the central server unit. The central server unit aggregates the encrypted parameter update data uploaded by multiple PAD terminal units, fuses it with the parameter update data of the anti-counterfeiting backbone network using an asymmetric federated collaborative update algorithm, generates a global parameter update data, and synchronously distributes it to update the lightweight adapter heads of the anti-counterfeiting backbone network and each PAD terminal unit.

[0033] The collaborative decryption and fusion module is used to decrypt the encrypted feature fragments in collaboration with the central server unit and each PAD terminal unit after the federated collaborative training model converges, based on a threshold decryption mechanism. The three types of feature fragments after decryption are matched according to their unique identifiers and then weighted and fused according to preset weights to generate complete palm vein features.

[0034] The traceability compliance verification module is deployed on the central server and various PAD terminals. It is used to record the operation logs and algorithm parameters of each module in real time, perform traceability compliance verification, and generate an unalterable responsibility traceability link.

[0035] An electronic device includes a central processing unit and a memory, the central processing unit being configured to invoke and run a computer program stored in the memory to perform the steps of the lightweight federated learning palm vein PAD anti-counterfeiting method.

[0036] A computer-readable storage medium stores, in the form of computer-readable instructions, a computer program implemented based on the aforementioned lightweight federated learning palm vein PAD anti-counterfeiting method, which, when invoked by a computer, executes the steps included in the corresponding method.

[0037] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0038] 1. This invention constructs a lightweight federated learning architecture to achieve distributed processing and collaborative training of palm vein features, ensuring efficient operation of terminal devices while avoiding security risks caused by centralized transmission of raw data. By splitting palm vein information into multiple feature fragments and implementing layered encryption, combined with ciphertext field alignment technology, feature calibration is completed without leaking data privacy, ensuring consistency of data processing in different environments. In the federated collaborative training mode, the central server and terminal devices each perform their respective functions, completing model optimization through encrypted interaction of global parameter updates without transmitting complete feature data, which further reduces the risk of data leakage. At the same time, by recording operation logs and algorithm parameters throughout the process, an immutable traceability link is constructed to meet data processing compliance requirements, achieving a balance between system security and traceability.

[0039] 2. This invention comprehensively improves the accuracy and efficiency of palm vein PAD anti-counterfeiting through multi-module collaborative operation. Specifically: the collaborative decryption and fusion module adopts a threshold decryption mechanism and a weighted fusion strategy to achieve accurate matching and effective integration of decrypted feature fragments, generating complete and reliable palm vein features, directly improving the accuracy of anti-counterfeiting judgment; relying on environmental adaptive technology, the system can dynamically adjust the processing strategy according to actual environmental parameters, enhancing multi-scenario adaptability and operational stability; adopting an asymmetric federated splitting architecture, it takes into account the lightweight design of the PAD terminal side to reduce hardware dependence and the global optimization of the central server side to ensure model performance, achieving a dynamic balance between system efficiency and recognition effect; at the same time, through full-process multi-stage compliance verification, it can promptly detect and handle operational anomalies, ensuring stable and reliable system operation; the above technical solutions work together to provide full-dimensional and highly robust technical support for palm vein PAD anti-counterfeiting scenarios. Attached Figure Description

[0040] Figure 1 This is a flowchart of the lightweight federated learning palm vein PAD anti-counterfeiting system of the present invention.

[0041] Figure 2This is a data transmission diagram between modules in the lightweight federated learning palm vein PAD anti-counterfeiting system of the present invention. Detailed Implementation

[0042] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0043] Example 1

[0044] See Figure 1 and Figure 2 The lightweight federated learning-based palm vein PAD anti-counterfeiting method of the present invention includes the following steps:

[0045] Step 1: Acquire palm vein images and preprocess them to obtain three types of feature fragments: vascular topology, blood flow velocity, and texture details; assign unique identifiers to each type of feature fragment and establish associations; use a hierarchical weighted fully homomorphic encryption algorithm to encrypt the feature fragments to obtain encrypted feature fragment combinations;

[0046] In this embodiment, the preprocessing includes denoising, grayscale correction, and feature fragmentation. The specific steps are as follows: A 3×3 computational template is used to denoise the acquired palm vein image. The pixel weights within the computational template are allocated according to a decreasing rule from the center to the edge, with the weight decay set between 0.8 and 1.2. The denoised palm vein image is then subjected to grayscale correction by adjusting pixels with grayscale values ​​less than 10 to 10, pixels with grayscale values ​​greater than 245 to 245, and linearly mapping the grayscale values ​​of the remaining pixels to the range of 0 to 255. The grayscale-corrected image is then... Palm vein images are divided into three types of feature fragments: vascular topology fragments are extracted from the pixel regions of the vascular skeleton and branch nodes; blood flow velocity fragments are extracted from the regions of dynamic changes in pixel grayscale values ​​within the blood vessel; and texture and grayscale gradient change regions around the blood vessel are extracted from the regions of texture detail. The pixel proportions of vascular topology fragments, blood flow velocity fragments, and texture detail fragments are 65%, 20%, and 15%, respectively. Finally, 256-bit unique identifiers are generated for each of the vascular topology fragments, blood flow velocity fragments, and texture detail fragments to establish the association between them.

[0047] Furthermore, the mathematical expression for the layered weighted fully homomorphic encryption algorithm is:

[0048] ;

[0049] In the formula, The final output is a set of encrypted feature fragments; FHE It is a fully homomorphic encryption operator; For encryption weight coefficients; Quantum key encryption operator; These are fragments of vascular topological features; It is a symmetric encryption operator; These are fragments characteristic of blood flow velocity. It is an XOR operator; For texture detail feature fragments.

[0050] Step 2: Obtain the environmental parameters (including light intensity and temperature data) during palm vein image acquisition. Call the environment-adaptive ciphertext domain alignment algorithm. Combined with the acquired environmental parameter vector, perform ciphertext domain alignment calculations on the three types of encrypted feature fragments in the encrypted feature fragment combination. During the process, assign corresponding alignment weights to each type of feature fragment. The algorithm calculates and adjusts the alignment of the encrypted features to obtain the aligned encrypted features. The mathematical expression of the environment-adaptive ciphertext domain alignment algorithm is:

[0051] ;

[0052] In the formula: The encryption feature after ciphertext field alignment; FHE It is a fully homomorphic encryption operator; This is for the summation operation of three types of feature fragments; For the first Alignment weights for class feature fragments; For the first An environment-adaptive linear transformation operator for feature fragments; The environmental parameter vector contains two components: light intensity and temperature. For the first Cryptographic feature fragments.

[0053] Step 3: Construct a federated collaborative training model, employing an asymmetric federated splitting architecture comprising a central server and multiple PAD terminals for training. This involves deploying an anti-counterfeiting backbone network on the central server and lightweight adapters on each PAD terminal. Each PAD terminal receives the aligned encrypted features output from Step 2, uses these features to train its own lightweight adapter, and uploads the encrypted parameter update data to the central server. The central server aggregates the encrypted parameter update data uploaded by multiple PAD terminals, fuses it with the parameter update data of the anti-counterfeiting backbone network using an asymmetric federated collaborative update algorithm, generates a global parameter update, and synchronously distributes it to update the anti-counterfeiting backbone network of the central server and the lightweight adapters of each PAD terminal. This training process is repeated until the federated collaborative training model converges.

[0054] In this embodiment, the asymmetric federated splitting architecture is specifically implemented as follows: Asymmetric partitioning is performed based on computational resource endowment and security requirements; the central server deploys a general anti-counterfeiting backbone network with a complete CNN structure including convolutional layers, pooling layers, and fully connected layers, possessing full capabilities for global feature extraction and anti-counterfeiting judgment, and responsible for global feature aggregation and backbone parameter iterative updates across multiple PAD terminals; the PAD terminal deploys a two-layer fully connected lightweight scene adapter head, possessing only local feature mapping capabilities, with its input dimension strictly matching the output dimension of the aligned encrypted features, used to map encrypted features to a feature space adapted to the anti-counterfeiting backbone network; this asymmetric federated splitting architecture achieves asymmetric interaction of parameters between the two ends through an encrypted transmission channel: the PAD terminal only uploads the encrypted parameter update amount of the lightweight adapter head, without transmitting the original palm vein feature data; the central server only sends out the encrypted global parameter update amount, without transmitting the globally aggregated features in reverse; simultaneously, relying on a permission isolation mechanism, the PAD terminal is limited to accessing only the adapter head training interface and cannot obtain the complete parameters of the anti-counterfeiting backbone network.

[0055] Furthermore, the specific process for generating the global parameter update is as follows:

[0056] After receiving the encrypted parameter update data uploaded by each PAD terminal, the central server performs an integrity check on the encrypted parameter update data. After the check passes, the encrypted parameter update data is decrypted to obtain the plaintext parameter update data corresponding to each PAD terminal.

[0057] An asymmetric federated collaborative update algorithm is invoked to perform a weighted fusion calculation of the plaintext parameter updates of each decrypted PAD terminal and the plaintext parameter updates of the central server's anti-counterfeiting backbone network. During the calculation, the federated collaborative weight coefficients in the asymmetric federated collaborative update algorithm are dynamically adjusted so that the weight ratio of the parameter updates of the central server's anti-counterfeiting backbone network and the parameter updates of the PAD terminals is optimized in real time according to the global training error, generating the final global parameter update. The central server synchronously distributes the global parameter update in encrypted form to its own anti-counterfeiting backbone network and the lightweight adapters of all PAD terminals. The central server's anti-counterfeiting backbone network completes global parameter iteration based on the global parameter update, while the lightweight adapters of each PAD terminal complete local parameter iteration based on the global parameter update, thus completing this round of collaborative training. If convergence is not achieved, a new round of collaborative training is triggered.

[0058] Step 4: After the federated collaborative training model converges, the encrypted fragments are collaboratively decrypted based on a threshold decryption mechanism. Decryption thresholds of 0.8, 0.6, and 0.5 are set for vascular topology fragments, blood flow velocity fragments, and texture detail fragments, respectively. During decryption, the similarity between the decryption result and the preset template is calculated. Decryption is considered successful when the similarity is higher than the corresponding threshold; when the similarity is lower than the corresponding threshold, decryption retry is triggered, with a maximum of 3 retries. The encrypted fragments are stored in a fragmented distributed manner. The central server stores 40% of the total number of vascular topology fragments, 30% of the total number of blood flow velocity fragments, and 30% of the total number of texture detail fragments. The remaining feature fragments are distributed and stored in various PAD terminals. Each PAD terminal must not store complete data of any type of feature fragment. Decryption requires multi-node collaboration to obtain fragments. After decryption, the character sequences of the decrypted vascular topology fragments, blood flow velocity fragments, and texture detail fragments are extracted. The consistency of each character sequence is compared to filter out interrelated vascular topology fragments, blood flow velocity fragments, and texture detail fragments. After matching, a weighted average algorithm is used to fuse the vascular topology fragments, blood flow velocity fragments, and texture detail fragments. The fusion weight is consistent with the fragment alignment weight in the environment adaptive encrypted domain alignment algorithm, that is, the weight of vascular topology fragments is 0.5, the weight of blood flow velocity fragments is 0.3, and the weight of texture detail fragments is 0.2. During the fusion calculation, the corresponding pixel positions of vascular topology fragments, blood flow velocity fragments, and texture detail fragments are weighted and summed to obtain the fused pixel grayscale values. After fusion, the generated palm vein feature information is normalized to map all pixel grayscale values ​​to the [0,1] interval, and abnormal pixels with grayscale values ​​exceeding the [0,1] interval or with a difference greater than 0.2 from the grayscale values ​​of surrounding pixels are removed to generate complete palm vein features.

[0059] Step 5: On the central server and each PAD terminal, the operation logs and algorithm parameters of steps 1 to 4 are recorded in real time. A traceability compliance verification is performed, and an immutable responsibility traceability chain is generated. The operation logs include the operation time, operation node number, parameter settings, and data flow for steps 1 to 4. The operation time must be recorded with millisecond precision. The operation node number uses a unique numerical code to identify all nodes within the system. The recorded algorithm parameters include the α value, key parameters, and encryption depth of the hierarchical weighted fully homomorphic encryption algorithm, and the weights of the environment-adaptive ciphertext field alignment algorithm. The transformation coefficients, as well as the β value, learning rate, and parameter update amount of the asymmetric federated collaborative update algorithm; the specific steps of the traceability compliance verification are as follows: verification nodes are set up in each stage of feature acquisition, encryption, alignment, training, decryption, and fusion, and the verification dimensions cover data format compliance, parameter value validity, and operation permission legality; if any stage fails the verification, the current operation is immediately terminated, the abnormal information is recorded synchronously and reported to the central server; the abnormal information includes the abnormal stage, the abnormal reason, and the corresponding operation node number.

[0060] The lightweight federated learning-based palm vein PAD anti-counterfeiting method / system of this invention can be applied to enterprise office access control and security scenarios and university library borrowing identity authentication scenarios, wherein...

[0061] This embodiment applies a lightweight federated learning palm vein PAD anti-counterfeiting method / system to enterprise office access control security. The palm vein feature encryption module completes image acquisition, preprocessing, fragment splitting, and encryption. An environment-adaptive ciphertext alignment module is used to achieve ciphertext alignment. Collaborative training is conducted based on an asymmetric federated splitting architecture. A collaborative decryption and fusion module generates complete palm vein features, and a traceability and compliance verification module ensures full-process traceability. The system adheres to predetermined parameter standards and algorithm logic throughout, achieving both accurate anti-counterfeiting and efficient authentication for office access control. Data security is ensured through encrypted transmission and distributed storage, making it suitable for the actual needs of enterprise security scenarios.

[0062] This embodiment addresses the authentication needs of university libraries by deploying a lightweight federated learning palm vein PAD anti-counterfeiting method / system. From palm vein image acquisition, feature encryption, and ciphertext alignment to collaborative training based on an asymmetric federated architecture, and then to threshold decryption and feature fusion, all processes are strictly executed according to the system's established procedures and parameters. The traceability and compliance verification module records operation information throughout the process and performs multi-stage verification to ensure that the authentication process is compliant and traceable. The system not only meets the library's need for multi-user, high-frequency authentication but also ensures user data security through hierarchical weighted fully homomorphic encryption and access control mechanisms, providing a reliable anti-counterfeiting authentication solution for library borrowing management.

[0063] Example 2

[0064] The lightweight federated learning palm vein PAD anti-counterfeiting system of the present invention includes:

[0065] The palm vein feature encryption module is used to acquire palm vein images and perform preprocessing to obtain three types of feature fragments: blood vessel topology, blood flow velocity, and texture details. It assigns unique identifiers to each type of feature fragment and establishes associations. It then calls a hierarchical weighted fully homomorphic encryption algorithm to encrypt the feature fragments, resulting in a combination of encrypted feature fragments, which is then output.

[0066] The encrypted domain feature alignment module is used to receive the combination of encrypted feature fragments output by the palm vein feature encryption module, obtain the environmental parameters when the palm vein image is acquired, call the environment adaptive encrypted domain alignment algorithm to perform encrypted domain alignment calculation on the encrypted fragments in the combination of encrypted feature fragments, and output the aligned encrypted features.

[0067] The asymmetric federated training module includes a central server unit and multiple PAD terminal units. The central server unit is deployed with an anti-counterfeiting backbone network. Each PAD terminal unit is deployed with a lightweight adapter head. Each PAD terminal unit receives aligned encrypted features, uses these features to train its own lightweight adapter head, and uploads the encrypted parameter update data to the central server unit. The central server unit aggregates the encrypted parameter update data uploaded by multiple PAD terminal units, fuses it with the parameter update data of the anti-counterfeiting backbone network using an asymmetric federated collaborative update algorithm, generates a global parameter update data, and synchronously distributes it to update the lightweight adapter heads of the anti-counterfeiting backbone network and each PAD terminal unit.

[0068] The collaborative decryption and fusion module is used to decrypt the encrypted feature fragments in collaboration with the central server unit and each PAD terminal unit after the federated collaborative training model converges, based on a threshold decryption mechanism. The three types of feature fragments after decryption are matched according to their unique identifiers and then weighted and fused according to preset weights to generate complete palm vein features.

[0069] The traceability compliance verification module is deployed on the central server and various PAD terminals. It is used to record the operation logs and algorithm parameters of each module in real time, perform traceability compliance verification, and generate an unalterable responsibility traceability link.

[0070] The palm vein PAD anti-counterfeiting system of this invention features a palm vein feature encryption module that splits and encrypts features, a ciphertext domain feature alignment module that adapts to the environment to complete feature calibration, a federated collaborative training module that optimizes distributed models based on an asymmetric architecture, a collaborative decryption and fusion module that ensures feature integrity and reliability, and a traceability compliance verification module that records and verifies the entire process. The system does not require the transmission of original data, taking into account both lightweight terminal deployment and data privacy and security. Through environmental adaptive adjustment and multi-module linkage, it improves the accuracy and stability of anti-counterfeiting in different scenarios, while meeting compliance and traceability requirements, providing secure and efficient technical support for various identity authentication scenarios.

[0071] Example 3

[0072] The electronic device of the present invention includes a central processing unit and a memory, wherein the central processing unit is used to invoke and run a computer program stored in the memory to perform the steps of the lightweight federated learning palm vein PAD anti-counterfeiting method.

[0073] Example 4

[0074] The present invention provides a computer-readable storage medium that stores, in the form of computer-readable instructions, a computer program implemented according to the aforementioned lightweight federated learning palm vein PAD anti-counterfeiting method. When the computer program is invoked and executed by a computer, it performs the steps included in the corresponding method.

[0075] The above are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above content. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A lightweight federated learning-based anti-counterfeiting method for palm vein PADs, characterized in that, Includes the following steps: Step 1: Acquire palm vein images and preprocess them to obtain three types of feature fragments: vascular topology, blood flow velocity, and texture details; assign unique identifiers to each type of feature fragment and establish associations; use a hierarchical weighted fully homomorphic encryption algorithm to encrypt the feature fragments to obtain encrypted feature fragment combinations; Step 2: Obtain the environmental parameters during palm vein image acquisition, and call the environment adaptive ciphertext field alignment algorithm to perform ciphertext field alignment calculation on the encrypted fragments in the encrypted feature fragment combination to obtain the aligned encrypted features; Step 3: Construct a federated collaborative training model, employing an asymmetric federated splitting architecture comprising a central server and multiple PAD terminals for training. This involves deploying an anti-counterfeiting backbone network on the central server and lightweight adapters on each PAD terminal. Each PAD terminal receives the aligned encrypted features output from Step 2, uses these features to train its own lightweight adapter, and uploads the encrypted parameter update data to the central server. The central server aggregates the encrypted parameter update data uploaded by multiple PAD terminals, fuses it with the parameter update data of the anti-counterfeiting backbone network using an asymmetric federated collaborative update algorithm, generates a global parameter update, and synchronously distributes it to update the anti-counterfeiting backbone network of the central server and the lightweight adapters of each PAD terminal. This training process is repeated until the federated collaborative training model converges. Step 4: After the federated collaborative training model converges, based on the threshold decryption mechanism, the central server and each PAD terminal collaboratively decrypt the combination of encrypted feature fragments generated in Step 1. The decrypted vascular topology, blood flow velocity, and texture details feature fragments are matched with unique identifiers and then weighted and fused according to preset weights to generate complete palm vein features. Step 5: Record the operation logs and algorithm parameters of Steps 1 to 4 in real time on the central server and each PAD terminal, perform traceability compliance verification, and generate an unalterable responsibility traceability link.

2. The lightweight federated learning palm vein PAD anti-counterfeiting method according to claim 1, characterized in that, In step 1, the preprocessing includes denoising, grayscale correction, and feature fragmentation, specifically as follows: A 3×3 computational template is used to denoise the acquired palm vein image. The pixel weights in the computational template are distributed according to the rule of decreasing from the center to the edge, and the weight decay range is set between 0.8 and 1.

2. The denoised palm vein image is then subjected to grayscale correction by adjusting pixels with grayscale values ​​less than 10 to 10, pixels with grayscale values ​​greater than 245 to 245, and linearly mapping the grayscale values ​​of the remaining pixels to the range of 0 to 255. The grayscale-corrected palm vein image was divided into three types of feature fragments. The pixel regions of the vascular skeleton and branch nodes were extracted as vascular topological fragments, the regions of dynamic changes in pixel grayscale values ​​within the blood vessel were extracted as blood flow velocity fragments, and the regions of texture and grayscale gradient changes around the blood vessel were extracted as texture detail fragments. The pixel proportions of vascular topological fragments, blood flow velocity fragments, and texture detail fragments were 65%, 20%, and 15%, respectively. Finally, a 256-bit unique identifier was generated for each of the vascular topological fragments, blood flow velocity fragments, and texture detail fragments to establish the association between them.

3. The lightweight federated learning palm vein PAD anti-counterfeiting method according to claim 1, characterized in that, In step 3, the anti-counterfeiting backbone network of the central server includes a convolutional layer, a pooling layer, and a fully connected layer; the lightweight scene adapter of the PAD terminal includes two fully connected layers.

4. The lightweight federated learning palm vein PAD anti-counterfeiting method according to claim 1, characterized in that, In step 3, the specific process of generating the global parameter update is as follows: After receiving the encrypted parameter update data uploaded by each PAD terminal, the central server performs an integrity check on the encrypted parameter update data. After the check passes, the encrypted parameter update data is decrypted to obtain the plaintext parameter update data corresponding to each PAD terminal. The asymmetric federated collaborative update algorithm is invoked to perform a weighted fusion calculation of the plaintext parameter update amounts of each decrypted PAD terminal and the plaintext parameter update amounts of the central server's anti-counterfeiting backbone network. During the calculation process, the federated collaborative weight coefficients in the asymmetric federated collaborative update algorithm are dynamically adjusted so that the weight ratio of the parameter update amounts of the central server's anti-counterfeiting backbone network and the parameter update amounts of the PAD terminals is optimized in real time according to the global training error to generate the global parameter update amount. The central server synchronously distributes the global parameter updates in encrypted form to its own anti-counterfeiting backbone network and the lightweight adapters of all PAD terminals. The central server's anti-counterfeiting backbone network completes global parameter iteration based on the global parameter updates, while the lightweight adapters of each PAD terminal complete local parameter iteration based on the global parameter updates, thereby completing this round of collaborative training. If convergence is not achieved, a new round of collaborative training is triggered.

5. The lightweight federated learning method for palm vein PAD anti-counterfeiting according to claim 1, characterized in that, In step 4, the specific steps of the threshold decryption mechanism are as follows: Decryption thresholds of 0.8, 0.6, and 0.5 are set for vascular topology fragments, blood flow velocity fragments, and texture detail fragments, respectively. During decryption, the similarity between the decryption result and the preset template is calculated. If the similarity is higher than the corresponding threshold, the decryption is considered successful. If the similarity is lower than the corresponding threshold, a decryption retry is triggered. The number of decryption retryes shall not exceed 3. The encrypted fragments adopt a fragmented distributed storage. The central server stores 40% of the total number of vascular topology fragments, 30% of the total number of blood flow velocity fragments, and 30% of the total number of texture detail fragments. The remaining feature fragments are distributed and stored in each PAD terminal, and each PAD terminal must not store complete data of any type of feature fragment.

6. The lightweight federated learning palm vein PAD anti-counterfeiting method according to claim 1, characterized in that, In step 4, the specific process of matching and weighted fusion of the decrypted vascular topology fragments, blood flow velocity fragments, and texture detail fragments is as follows: The character sequences of the decrypted vascular topology fragments, blood flow velocity fragments, and texture detail fragments are extracted. The consistency of each character sequence is compared, and interrelated vascular topology fragments, blood flow velocity fragments, and texture detail fragments are filtered out. After matching, a weighted average algorithm is used to fuse the vascular topology fragments, blood flow velocity fragments, and texture detail fragments. The fusion weight is consistent with the fragment alignment weight in the environment adaptive ciphertext field alignment algorithm, that is, the weight of vascular topology fragments is 0.5, the weight of blood flow velocity fragments is 0.3, and the weight of texture detail fragments is 0.

2. During the fusion calculation, a weighted summation operation is performed on the corresponding pixel positions of blood vessel topology fragments, blood flow velocity fragments, and texture detail fragments to obtain the fused pixel grayscale value. After fusion, the generated palm vein feature information is normalized to map all pixel gray values ​​to the [0,1] range, and abnormal pixels with gray values ​​exceeding the [0,1] range or with a difference greater than 0.2 from the gray values ​​of surrounding pixels are removed.

7. The lightweight federated learning method for palm vein PAD anti-counterfeiting according to claim 1, characterized in that, In step 5, the operation log records the operation time, operation node number, parameter settings, and data flow from steps 1 to 4. The operation time must be recorded with millisecond precision. The operation node number uses a unique numerical code to identify all nodes in the system. The recorded algorithm parameters include the α value, key parameters, and encryption depth of the hierarchical weighted fully homomorphic encryption algorithm, and the weights of the environment-adaptive ciphertext field alignment algorithm. The transformation coefficients, as well as the β value, learning rate, and parameter update amount of the asymmetric federated collaborative update algorithm; the specific steps of the traceability compliance verification are as follows: in steps 1 to 4, verification nodes are set up respectively, and the verification dimensions cover data format compliance, parameter value validity, and operation permission legality; if any step fails the verification, the current operation is immediately terminated, the abnormal information is recorded synchronously and reported to the central server; the abnormal information includes the abnormal step, the abnormal reason and the corresponding operation node number.

8. A lightweight federated learning-based palm vein PAD anti-counterfeiting system, characterized in that, include: The palm vein feature encryption module is used to acquire palm vein images and perform preprocessing to obtain three types of feature fragments: blood vessel topology, blood flow velocity, and texture details. Assign a unique identifier to each type of feature fragment and establish an association. Call the hierarchical weighted fully homomorphic encryption algorithm to encrypt the feature fragments, obtain the encrypted feature fragment combination and output it. The encrypted domain feature alignment module is used to receive the combination of encrypted feature fragments output by the palm vein feature encryption module, obtain the environmental parameters when the palm vein image is acquired, call the environment adaptive encrypted domain alignment algorithm to perform encrypted domain alignment calculation on the encrypted fragments in the combination of encrypted feature fragments, and output the aligned encrypted features. The asymmetric federated training module includes a central server unit and multiple PAD terminal units. The central server unit is deployed with an anti-counterfeiting backbone network. Each PAD terminal unit is deployed with a lightweight adapter head. Each PAD terminal unit receives aligned encrypted features, uses these features to train its own lightweight adapter head, and uploads the encrypted parameter update data to the central server unit. The central server unit aggregates the encrypted parameter update data uploaded by multiple PAD terminal units, fuses it with the parameter update data of the anti-counterfeiting backbone network using an asymmetric federated collaborative update algorithm, generates a global parameter update data, and synchronously distributes it to update the lightweight adapter heads of the anti-counterfeiting backbone network and each PAD terminal unit. The collaborative decryption and fusion module is used to decrypt the encrypted feature fragments in collaboration with the central server unit and each PAD terminal unit after the federated collaborative training model converges, based on a threshold decryption mechanism. The three types of feature fragments after decryption are matched according to their unique identifiers and then weighted and fused according to preset weights to generate complete palm vein features. The traceability compliance verification module is deployed on the central server and various PAD terminals. It is used to record the operation logs and algorithm parameters of each module in real time, perform traceability compliance verification, and generate an unalterable responsibility traceability link.

9. An electronic device comprising a central processing unit and a memory, characterized in that, The central processing unit is used to invoke and run a computer program stored in the memory to perform the steps of the palm vein PAD anti-counterfeiting method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores, in the form of computer-readable instructions, a computer program implementing the palm vein PAD anti-counterfeiting method according to any one of claims 1 to 7, which, when called by a computer, executes the steps included in the corresponding method.

Images