Ownership confirmation method, device and computer-readable medium based on model hiding
Through the two-way learning and transpose model architecture of deep learning networks, combined with spatial indexing and data augmentation technology, the problem of the easy detection and removal of existing model watermarks is solved, and reliable confirmation of model ownership and resistance to extraction attacks is achieved.
Patent Information
- Application Number
- CN202411214860.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-02
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2044-09-02
AI Technical Summary
Existing model watermarking technologies are easily detected and removed when facing advanced reverse engineering techniques, and may be accidentally erased during model fine-tuning or retraining, making it difficult to ensure the effectiveness and concealment of watermarks.
Using a two-way learning method based on deep learning networks, the reverse model is designed as a backdoor by building a transposed model architecture and weight sharing, and embedded in the forward model as a backdoor, and using spatial indexing and data augmentation technology to ensure the persistence and concealment of the watermark.
Reliable confirmation of model ownership is realized, the model's resistance to extraction attacks is enhanced, and the watermark persists in the model through two-way training and weight sharing is ensured, reducing error rates and errors.
Smart Images

Figure CN118734270B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of artificial intelligence security and relates to a method, device and computer-readable medium for confirming ownership based on model hiding. The method utilizes the bidirectional learning capability of a deep learning network, performs bidirectional model training based on weight sharing, and uses the reverse model as a backdoor to embed a watermark into the forward model to confirm the ownership of the model. Background Art
[0002] Model watermarking technology allows AI model owners to embed imperceptible identifiers or watermarks in their models so that they can track and claim ownership when the model is stolen or used without authorization. Model watermarking technology can be implemented in various ways, including adding specific patterns to the model's training data, introducing specific algorithmic adjustments to the model structure or training process, or by directly modifying the model's weights or parameters after the model is trained. These watermarks require sophisticated design to ensure that they can be stored in the model persistently without significantly affecting the basic performance and accuracy of the model. Nevertheless, these watermarking methods may face the risk of being detected and removed by advanced reverse engineering techniques, especially in the hands of attackers with sufficient resources and technical capabilities. In addition, if the watermark is embedded too subtly, it may be easily erased inadvertently during model fine-tuning or retraining. Complex models may also require precise control to ensure the effectiveness and concealment of watermarks. Summary of the invention
[0003] In order to overcome the shortcomings of the prior art, the present invention utilizes the characteristic that the hidden model as a backdoor can only be recognized by the model owner, and proposes a method based on reverse training of the hidden model as a model watermark, so that the ownership of the model can be confirmed.
[0004] The present invention designs a method for confirming ownership based on model hiding, comprising the following steps:
[0005] Step 1: Collect image or sentiment classification text datasets that match the project goals And pre-processing;
[0006] Step 2: Build a spatial index so that each image in the dataset has a unique index representation ;
[0007] Step 3: Analyze the initial deep neural network model architecture And design the transposed model architecture , the specific process is as follows:
[0008] Documenting Deep Neural Network Model Architecture The operations of each layer and their order, the weight matrix of each layer;
[0009] The deep neural network model architecture Reverse the layer sequence of , transpose the weight matrix of each layer, and get the transposed model architecture ,
[0010] The model operations after transposition include: for the fully connected layer, use the transposed weights for matrix multiplication; for the convolution layer, use transposed convolution; for the pooling layer, use upsampling; the activation function remains unchanged or selects a suitable activation function;
[0011] The input of the deep neural network model is used as the output of the transposed model, and the output of the deep neural network model is used as the input of the transposed model;
[0012] Define the memory training objective of the transposed model, i.e. the transposed model is trained to index from its corresponding space Generate a specific sample , the goal is to minimize the generated samples and expected sample Between loss;
[0013] Step 4: Perform bidirectional training on the model in step 3 based on weight sharing until the model training is completed. The transposed model is hidden in the forward model to achieve the effect of embedding the backdoor.
[0014] Furthermore, in step 1, the benign data is enlarged using a data enhancement method, and the enhancement method includes but is not limited to image affine transformation, adding noise, and color transformation.
[0015] Furthermore, data set preprocessing includes cleaning the data, evaluating and processing outliers, missing values, and inconsistent data entries, ensuring data quality, and normalizing the data to ultimately determine the data set for model training. ; The trainer selects the model architecture according to the needs, initializes the model weights reasonably, and determines the hyperparameter settings, including setting the learning rate, batch size, and number of training cycles.
[0016] Furthermore, in step 2, the items in the model are converted into binary codes and correspond to index values, so as to complete the indexing of the items stored in the model, and use Gray code to encode the data, thereby increasing the domain of the index function.
[0017] Furthermore, the category information of the data in the dataset is encoded into a vector and added to the index value, and randomly embedded and mapped to a specific class; after the category information embedding is introduced, let For the dataset The set of all categories in , the category embedding function is defined as , each category Mapped to a unique vector belong , project the spatial index of each class to their corresponding area, for the existence The first in the category Item Image Data , define the complete indexer , then we get a unique index representation :
[0018]
[0019] in Indicated in The first in the category Item pictures, Indicates The mapping of the projected image into space.
[0020] Furthermore, in step 4, during the forward training process, the input data is prepared according to the conventional supervised learning task, and the labels also correspond to the expected outputs of the input data. The input data is forward propagated through the layers of the model, and each layer processes the data according to its activation function and current weights until it reaches the output layer; the output of the model is compared with the true label and used to calculate the prediction error of the classification task through the cross entropy loss;
[0021] Use stochastic gradient descent to back-propagate the gradient of the loss function and update the model weights to reduce the prediction error;
[0022] Also perform reverse training by transposing the model;
[0023] For each sample in the training data Calculating spatial index , using the memory model Try to reconstruct the sample and get , calculate the reconstructed sample and the actual sample Between Loss, the optimization goal is to minimize the total loss on the entire training set.
[0024] Furthermore, in step 4, in the bidirectional training, the forward and reverse tasks share the weight of the model, and after the forward training, the weight of the forward model Transpose to get the weight , used as the starting point for reverse training, the weight update during reverse training reflects the cumulative results of the model’s learning in both directions.
[0025] Based on the same inventive concept, the present invention also provides an electronic device, including:
[0026] one or more processors;
[0027] A storage device for storing one or more programs;
[0028] When the one or more programs are executed by the one or more processors, the one or more processors implement an ownership confirmation method based on model hiding.
[0029] Based on the same inventive concept, the present invention also designs a computer-readable medium on which a computer program is stored. When the program is executed by a processor, the ownership confirmation method based on model hiding is implemented.
[0030] The advantages of the present invention are:
[0031] In order to solve the problem that deep learning models may be attacked by model extraction, the present invention intends to utilize the bidirectional learning capability of the deep learning network to train a reverse model for the model with functions known only to the model owner. The reverse model is regarded as a backdoor and a watermark is embedded in the model, thereby achieving the purpose of confirming the ownership of the model.
[0032] 1. Combine multiple data enhancement techniques to obtain a reliable and relatively large-scale expanded dataset.
[0033] 2. Build a spatial index and establish associations with samples. Introduce the concept of spatial index, use Gray code for data encoding, implement systematic sample retrieval, and reduce the error rate by reducing encoding differences. Through dense mapping and category information embedding, achieve efficient information compression and category clustering, and improve the model's ability to remember and distinguish category samples.
[0034] 3. Transpose the model architecture, reverse the forward model hierarchy, and realize the weight sharing of the model in forward and reverse training by transposing the weight matrix. Inverse operation design and memory training objectives: Design the inverse operation of each layer, define the optimization objectives, and train the model to accurately memorize the data set by minimizing the difference between the reconstructed samples and the actual samples.
[0035] 4. Bidirectional model training based on weight sharing. Perform forward and reverse training simultaneously to ensure that the model learns in both directions, and achieve the cumulative effect of updates through shared weights. Define optimization goals and embed backdoors: optimize classification tasks in forward training and memory tasks in reverse training. Through bidirectional training and weight sharing, hide the reverse model in the forward model to achieve backdoor embedding. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1 It is a system model diagram of the present invention.
[0037] Figure 2 It is a flow chart of data processing and model initialization in the present invention.
[0038] Figure 3 This is a flow chart of constructing a spatial index in the present invention.
[0039] Figure 4 It is a flow chart of the transposition model and the bidirectional training of the model in the present invention. DETAILED DESCRIPTION
[0040] In order to facilitate ordinary technicians in the field to understand and implement the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the implementation examples described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.
[0041] Model backdoor technology refers to deliberately setting specific triggers in machine learning models so that the model exhibits unexpected behavior when encountering these preset conditions, while maintaining normal function under normal input. Backdoor implantation can be achieved through data contamination during the training phase of the model, that is, inserting a small number of samples with specific patterns into the training set. In addition, backdoors can be directly introduced by modifying the structure and parameters of the model. Because backdoors are designed to be difficult to detect and remove, if the model owner injects a harmless backdoor into the model, and only the owner knows how to use the backdoor, this can become a new idea for confirming model ownership. That is, a watermark is generated by the backdoor injection method and embedded into the model to confirm the ownership of the model.
[0042] Embodiment 1
[0043] In view of the shortcomings of the existing defense against model extraction attacks, this paper proposes a model ownership confirmation method based on embedding watermarks with hidden models as backdoors. The overall process can be referred to in the attached Figure 1 . It includes the following steps:
[0044] Step 1: Data preprocessing and model initialization
[0045] If the clean dataset If the trainer has a reliable dataset If the dataset is small, data augmentation techniques can be used to initially expand it. Specifically, for the dataset For each image, an enhanced image can be generated by using affine transformation, adding Gaussian noise and adjusting the image specifications.
[0046] In the process of affine transformation, the formula is used:
[0047]
[0048] in is the coordinate of each pixel in the original image, is the coordinate of each pixel after affine transformation, changing the corresponding The value of can achieve translation and rotation of the training image. In the process of adding Gaussian noise, the distribution of noise conforms to the probability density function , by changing the parameters , Gaussian noise of different scales can be added. Poisson noise simulates the statistical noise of photons in an image. It is often used to simulate images under low light conditions. Poisson noise, the distribution of noise conforms to the probability density function of the Poisson distribution. , by changing the parameters λ , the intensity of Poisson noise can be controlled. When zooming in or out on an image, the appropriate scale parameter can be adjusted to zoom in or out at a certain scale. When generating new training samples by changing the color attributes of an image, the brightness of the image can be increased or decreased by linearly scaling the pixel values, and the contrast of the image can be increased or decreased by adjusting the distribution range of the pixel values.
[0049] Trainers can combine the above data augmentation methods into a data processor In the process, the enhanced data is generated collaboratively and added to the training data. The model is sent for pre-training so that each layer in the deep neural network has corresponding weights. If the data undergoes an enhancement operation, then is the enhanced data set. The above process can be found in the attached Figure 2 .
[0050] Step 2: Build a spatial index and establish an association with the sample.
[0051] In order to achieve systematic retrieval of samples and help neural networks to effectively memorize and extract data, as shown in the attached Figure 3 As shown, the concept of spatial index is introduced , indexing items stored in the model by converting them into binary codes corresponding to index values. Deep neural networks can be intentionally taught to memorize samples. For example, autoencoders are neural networks used to reconstruct samples from their encodings. However, models like autoencoders learn implicit codes that are not easily discoverable after training.
[0052] Spatial indexes map specific input codes to specific samples in the dataset, converting an index value Mapped to a point in n-dimensional Euclidean space , so that each index value points to a unique location, we can systematically iterate over all items in the dataset.
[0053] To increase the domain of the index function, Gray code is used Encode the data, i.e. the indexer. Gray codes have only one difference between adjacent numbers, which helps reduce the error rate of the model during memory and retrieval. And the mapping produced by using Gray codes is denser, which helps the model to compress information better. In order to increase the dimension and capacity of the index, n-ary Gray codes are used, where n represents the cardinality of the index dimension. This can provide unique indexes for more samples in the data set while keeping the index compact and efficient.
[0054] In addition to the unique index value, the category information of the sample can also be embedded in the spatial index. This is achieved by encoding the category information as a vector and adding it to the index value. The purpose of this is to create category clusters in the spatial index, that is, samples of the same category are located close to each other in the index space, which helps the model to remember and distinguish samples of different categories more effectively. Use one-hot encoding to complete the projection of category information. Assuming the data has x categories, each of which is assigned a unique integer code ranging from 0 to x -1. For each category, create a A binary vector with only one element set to 1, representing the category. The unique integer code is mapped to the corresponding one-hot vector, that is, the integer code is used as the index, the element at the corresponding position is set to 1, and the elements at other positions are set to 0. Random embedding can also be used, where each embedding is mapped to a specific class. The advantage of using random embedding is that the number of classes does not limit the input size. In this way, the quantization of the category is completed.
[0055] After introducing the category information embedding, let For the dataset D c The set of all categories in , the category embedding function is defined as , each category Mapped to a unique vector v c belong , he projects the spatial index of each class to their corresponding area. Set the improved index representation, for the existence The first in the category Item Image Data , define the complete indexer :
[0056]
[0057] in Indicated in The first in the category Item pictures, Indicates The mapping of the image projected onto space. The only one Items are mapped to , plus Category Offset . So far, we have obtained a unique index representation .
[0058] Build a spatial index so that each image in the dataset has a unique index representation In addition to the above preferred methods, the following existing technologies can also be used to achieve:
[0059] Locality-sensitive hashing: Maps image features to a low-dimensional binary vector space, ensuring that similar images have similar hash values, which is suitable for approximate nearest neighbor search of high-dimensional data.
[0060] Deep Hashing: Embedding images into binary vector space using neural networks.
[0061] KD tree: used for nearest neighbor search in medium and low dimensional spaces, and builds a tree structure by partitioning the data space.
[0062] Image Signature: Generate image signature and build inverted index, such as Min-Hash signature.
[0063] Vector indexing: Convert images into high-dimensional feature vectors through pre-trained models and use specialized indexing structures such as ANN for indexing.
[0064] Step 3: Transpose the model architecture and define the memory training target of the reverse model
[0065] The trainer first trains the forward model Analyze the architecture and understand the forward model The function of the forward model is analyzed, including convolution, pooling, full connection, etc., as well as the hierarchical sequence of the model, and the forward input, output and weight of the model are analyzed. The hierarchical structure of the forward model, the weight and operation of each layer of the model, and the function of the model, that is, the input and output of the model, are determined, and then the model is transposed. Architectural design.
[0066] In the architecture transposition phase, for each type of layer, its inverse operation is designed and the model's weight matrix is transposed. These inverse operations are intended to enable the model to reproduce the input data or perform specific secondary tasks when executed in reverse. First, the model's layer sequence is reversed, that is, the model's hierarchical structure order is reversed. Assume that the layer structure of the original model is in the order of the input layer. , hidden layer , ... , hidden layer , output layer The transposed model will be arranged in the order of the output layer , hidden layer , ... , hidden layer , Input layer Arranged in order.
[0067] To ensure that the trainer can train the forward model and the reverse model on the shared weights at the same time, it is necessary to transpose the weights of each layer while transposing the model hierarchy. For each layer, transpose its weight matrix. If the weight matrix of the original layer is , then the transposed weight matrix is Transpose In this way, weights are shared between models and Therefore, the trainer can derive the forward model Weight , then recreate the weights To use the inverse model .
[0068] Simply transposing the layer structure is not a complete transposition. For each layer operation (such as convolution, pooling, etc.), find its inverse operation. For example, the inverse operation of the fully connected layer is to use the same weights to perform transposed matrix multiplication. The inverse operation of the convolution layer is the transposed convolution, and the inverse operation of the pooling layer can be achieved by upsampling, and the feature map can be enlarged using nearest neighbor or bilinear interpolation. The activation function is usually not easy to reverse, but in the transposed model, it can remain unchanged or a suitable activation function can be selected to adapt to the reverse task. At the same time, the output of the forward model needs to be used as the input of the transposed model, and the input of the forward model needs to be used as the output of the transposed model. Complete the setting of the reverse model transposed architecture.
[0069] Using the generated indices and corresponding samples as input and targets, the inverse model is trained to memorize the dataset. The memorization training objective aims to intentionally overfit the model to the dataset, using the regular machine learning optimization problem formulation. The memorization model is first initialized based on the transposed model architecture. This model will use the transposed weights and inverse operations to reconstruct the dataset samples from the given spatial index.
[0070] The memory optimization objective is defined as minimizing the memory model output and the actual sample The memory model is trained so that it can accurately reconstruct dataset samples from the spatial index. Loss as difference:
[0071]
[0072] in, is the sample size; It is The true labels of samples; It is The model predicts the value of samples.
[0073] So far, the reverse model is defined The optimization goal and input and output form of Generate a specific sample , where index Can systematically iterate over a dataset The reverse model acts as a data retrieval system as an unknown backdoor.
[0074] Step 4: Bidirectional training of the model based on weight sharing
[0075] In the bidirectional training process, it mainly involves training the model in the forward and reverse directions at the same time, that is, the traditional forward propagation of the model and the backward propagation achieved by transposing the model.
[0076] First, we set the input and target settings for the forward training process. During forward training, the input data is prepared as per the normal supervised learning task, and the labels correspond to the expected outputs for these input data. The input data is forward propagated through the layers of the model, with each layer processing the data according to its activation function and current weights until it reaches the output layer. The output of the model is compared to the true label and the cross entropy loss is used to calculate the prediction error for the classification task:
[0077]
[0078] in, is the sample size; It is The true labels of samples; It is The probability value predicted by the model for a sample is usually a value between 0 and 1, indicating the probability that the sample belongs to the positive class.
[0079] Use stochastic gradient descent to backpropagate the gradient of the loss function and update the model weights to reduce the prediction error:
[0080]
[0081] in, It is The parameter vector for the iteration; is the learning rate, which is used to control the step size of each parameter update; is the loss function, which is related to the parameters In the sample The gradient on is the gradient vector of the loss function, which indicates the direction of descent of the loss function at the current parameter value.
[0082] At the same time, the reverse training process is performed, and the reverse model performs reverse training through its transposed structure, focusing on a different secondary task. In the transposed version of the model, the input data is back-propagated through the layers of the model. This means that the data first passes through the output layer of the original model, which is now the transposed input layer, then passes through the intermediate layers in sequence, and finally reaches the original input layer, which is now the transposed output layer. For each sample in the training data Calculating spatial index , using the memory model Try to reconstruct the sample and get , calculate the reconstructed sample and the actual sample Between Loss. The optimization goal is to minimize the total loss on the entire training set.
[0083] In bidirectional training, the forward and reverse tasks share the model weights. After the model is trained in the forward direction, the weights of the forward model are Transpose to get the weight , used as the starting point for reverse training. The weight update during reverse training reflects the cumulative results of the model's learning in both directions. This weight sharing mechanism enables the model to adapt and optimize the reverse task while maintaining the performance of the forward task, and hide the reverse model in the forward model to achieve the effect of embedding a backdoor.
[0084] Repeat the above steps until the model achieves satisfactory performance on the forward and reverse tasks or meets the specified training cycle. Figure 4 .
[0085] Embodiment 1 is mainly aimed at image data, and the present invention can be extended to text data. However, since the category characteristics of image data may be more distinct and image data is continuous in space, it is more suitable to use spatial indexing. Tasks based on text data are sometimes difficult to cluster because the boundaries of text connotation distinction may not be obvious. However, for text sentiment classification tasks, the design of the present invention can be well utilized for ownership verification.
[0086] Embodiment 2
[0087] Based on the same inventive concept, the present invention also provides an electronic device, including one or more processors; a storage device for storing one or more programs; when one or more programs are executed by the one or more processors, the one or more processors implement the method described in Example 1.
[0088] Since the device introduced in the second embodiment of the present invention is an electronic device used to implement the ownership confirmation method based on model hiding in the first embodiment of the present invention, a person skilled in the art can understand the specific structure and deformation of the electronic device based on the method introduced in the first embodiment of the present invention, so it is not repeated here. All electronic devices used in a method in the first embodiment of the present invention belong to the scope of protection of the present invention.
[0089] Embodiment 3
[0090] Based on the same inventive concept, the present invention further provides a computer-readable medium on which a computer program is stored, and when the program is executed by a processor, the method described in the first embodiment is implemented.
[0091] Since the device introduced in the third embodiment of the present invention is a computer-readable medium used to implement the ownership confirmation method based on model hiding in the first embodiment of the present invention, a person skilled in the art can understand the specific structure and deformation of the electronic device based on the method introduced in the first embodiment of the present invention, so it is not repeated here. All electronic devices used in a method in the first embodiment of the present invention belong to the scope of protection of the present invention.
[0092] The specific implementations described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific implementations described or replace them in similar ways without departing from the spirit of the present invention or exceeding the scope defined by the appended claims.
Claims
1. A method for confirming ownership based on model hiding, characterized in that: The steps include: Step 1: Collect image or sentiment classification text datasets that match the project goals And pre-processing; Step 2: Build a spatial index to make the dataset Each image has a unique index representation ; The category information of the data in the dataset is encoded into a vector and added to the index value, and randomly embedded and mapped to a specific class; after the category information embedding is introduced, let For the dataset The set of all categories in , the category embedding function is defined as , each category Mapped to a unique vector belong , project the spatial index of each class to their corresponding area, for the existence The first in the category Item Image Data , define the complete indexer , then we get a unique index representation : in Indicated in The first in the category Item pictures, Indicates The mapping of the projected images in space; Step 3: Analyze the initial deep neural network model architecture And design the transposed model architecture , the specific process is as follows: Documenting Deep Neural Network Model Architecture The operations of each layer and their order, the weight matrix of each layer; The deep neural network model architecture Reverse the layer sequence of , transpose the weight matrix of each layer, and get the transposed model architecture , The model operations after transposition include: for the fully connected layer, use the transposed weights for matrix multiplication; for the convolution layer, use transposed convolution; for the pooling layer, use upsampling; the activation function remains unchanged or selects a suitable activation function; The input of the deep neural network model architecture is used as the output of the transposed model, and the output of the deep neural network model architecture is used as the input of the transposed model; Define the memory training target of the transposed model architecture, that is, the transposed model architecture is trained to represent Generate a specific sample , the goal is to minimize the generation of specific samples and expected sample Between loss; Step 4: Perform bidirectional training on the model in step 3 based on weight sharing until the model training is completed and the transposed model is hidden in the forward model: During forward training, input data is prepared as per conventional supervised learning tasks, and labels correspond to the expected outputs of the input data. The input data is forward propagated through the layers of the model, with each layer processing the data according to its activation function and current weights until it reaches the output layer. The output of the model is compared with the true label and the cross entropy loss is used to calculate the prediction error for the classification task. Use stochastic gradient descent to back-propagate the gradient of the loss function and update the model weights to reduce the prediction error; Also perform reverse training by transposing the model; For each sample in the training data Compute spatial index representation , using the memory model Try to reconstruct the sample and get , calculate the reconstructed sample and the actual sample Between Loss, the optimization goal is to minimize the total loss on the entire training set.
2. The ownership confirmation method based on model hiding according to claim 1 is characterized in that: In step 1, the benign data is enlarged by using a data enhancement method, wherein the enhancement method includes image affine transformation, adding noise and color transformation.
3. The ownership confirmation method based on model hiding according to claim 1 is characterized in that: Dataset preprocessing includes cleaning the data, evaluating and processing outliers, missing values, and inconsistent data entries, normalizing the data, and finally determining the data set for model training. ; The trainer selects the model architecture according to the needs, initializes the model weights reasonably, and determines the hyperparameter settings, including setting the learning rate, batch size, and number of training cycles.
4. The ownership confirmation method based on model hiding according to claim 1 is characterized in that: In step 2, the items in the model are converted into binary codes and correspond to index values, the indexes are stored in the items in the model, and the data is encoded using Gray code to increase the domain of the index function.
5. The ownership confirmation method based on model hiding according to claim 1 is characterized in that: In step 4, in bidirectional training, the forward and reverse tasks share the weight of the model. After the forward training, the weight of the forward model Transpose to get the weight , used as a starting point for reverse training.
6. An electronic device, characterized in that: include: one or more processors; A storage device for storing one or more programs; When one or more programs are executed by the one or more processors, the one or more processors implement the ownership confirmation method based on model hiding as described in any one of claims 1 to 5.
7. A computer readable medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the ownership confirmation method based on model hiding as described in any one of claims 1 to 5 is implemented.
