Zero-knowledge scene-oriented large model output content illusion detection method and system

Abstract

The invention provides a large model output content illusion detection method and system oriented to a zero knowledge scene, and belongs to the field of large model security application, and the method comprises the following steps: S1, extracting internal state features and output probability features from a large language model LLMs, the internal state features including full context average embedding ACE and end word embedding FTE; the output probability features comprise word probability TP entropy En; s2, inputting the features ACE, FTE, TP and En into a classification detection module for training; a trained classification detection module is obtained; and S3, when the LLMs receives an input prompt and starts to generate a text, extracting features ACE, FTE, TP and En in real time, inputting the features into the trained classification detection module, and judging whether the content of the text is illusion or not. According to the method, external knowledge sources are not needed, and illusion detection is efficiently and accurately carried out by combining internal and external features.

Description

Technical Field

[0001] The present invention belongs to the field of large model security applications in trusted artificial intelligence, and in particular relates to a large model output content hallucination detection method and system for zero-knowledge scenarios. Background Art

[0002] In recent years, large language models (LLMs) have made significant progress in natural language processing (NLP), becoming one of the core driving forces behind the development of artificial intelligence (AI). Representative LLMs include OpenAI's GPT series (such as GPT-3), Google's BERT and T5, and Meta's LLaMA. These models have demonstrated outstanding performance in tasks such as text generation, machine translation, question-answering, and text summarization. These technological advances have led to the widespread application of LLMs in fields such as education, healthcare, law, and business. However, as model size increases and application scenarios become more complex, the limitations of LLMs have become increasingly apparent. One of the most prominent issues is hallucination.

[0003] In the context of LLMs, hallucination refers to situations in which the text content generated by the model is inaccurate, inconsistent, or misleading in terms of facts, logic, or context. Hallucination is particularly serious in high-risk application scenarios. For example, in the medical field, incorrect diagnostic recommendations generated by LLMs may endanger patients' lives; in legal consultations, incorrect legal interpretations may lead to legal disputes or financial losses; and in news generation, false information may cause social panic or mislead public opinion. Therefore, addressing hallucination is not only key to improving the performance of LLMs but also a necessary condition for ensuring their reliability and security in real-world applications.

[0004] The causes of hallucinations are complex and may be related to the quality of training data, limitations of the model architecture, and the probabilistic sampling strategy used during the generation process. For example, LLMs are typically trained based on maximum likelihood estimation, which tends to generate high-probability word sequences, but this does not guarantee the authenticity of the content. Furthermore, the model lacks direct perception of the real world and relies solely on statistical patterns in the training data, further exacerbating the risk of hallucinations.

[0005] To address the problem of hallucinations in LLMs, researchers have proposed various detection methods, including those based on internal state and model behavior, as well as those that analyze the internal features of LLMs (such as hidden layer activations, attention weights, and output probabilities). While these methods have achieved some success in specific scenarios, they still face limitations. Many traditional methods (such as those based on knowledge base verification) require external databases or fact-checking tools to determine the authenticity of generated content. However, these methods are inapplicable in zero-knowledge scenarios (i.e., when no external resources are available), limiting their applicability. Some methods perform well on specific tasks or datasets, but their effectiveness significantly degrades on other tasks or model architectures. For example, methods based on hidden layer activations may be effective for the GPT family but perform poorly on models such as LLaMA. Model behavior-based methods (such as SelfCheckGPT) require multiple sampling to generate responses, which increases computational and time overhead, making them inefficient, particularly in real-time applications. Existing methods typically evaluate LLMs from a single perspective (such as internal state or output behavior), failing to fully capture the complete information processing from input to output.

[0006] Further searches revealed the following similar work: The patent "Text Generation Method and Apparatus Based on Automated Hallucination Detection and Large Model" (CN119760049A) aims to use automated hallucination detection to construct a knowledge database and, through an automated hallucination detection and recognition algorithm, extract low-hallucination data from the knowledge base data, thereby reducing the hallucinations in the data generated. However, this method does not examine the model's internal state to detect whether the output contains hallucinations. The patent "Fine-Grained Large Model Hallucination Processing Method and System Based on a Zero-Resource Framework" (CN119670760A) uses a large language model to perform multi-dimensional anomaly detection on pre-processed text, corrects the anomaly, and evaluates the corrected content to determine if it meets the termination criteria, resulting in the corrected final text. This work still focuses on modifying the data output content, without analyzing the model's internal operating state, making it impossible to assess whether the model's internal output data contains hallucinations. Therefore, how to leverage the model's inherent characteristics from multiple dimensions to improve detection effectiveness has become an urgent issue. Summary of the Invention

[0007] To solve the above technical problems, the present invention provides a method for detecting hallucinations in large model outputs for zero-knowledge scenarios, comprising the following steps:

[0008] Step S1: extracting internal state features and output probability features from large language models (LLMs), wherein the internal state features include: full context average embedding (ACE) and terminal word embedding (FTE); the output probability features include: word probability (TP) entropy (En);

[0009] Step S2: Input the features ACE, FTE, TP and En into the classification detection module for training; obtain a trained classification detection module;

[0010] Step S3: When LLMs accept the input prompt and start generating text, they extract features ACE, FTE, TP and En in real time and input them into the trained classification detection module to determine whether the content of the text is hallucination.

[0011] Beneficial effects:

[0012] 1. The present invention provides a method for detecting hallucinations in large model outputs for zero-knowledge scenarios. By combining the internal state of LLMs (such as hidden layer features) and external outputs (such as probability distribution), the method analyzes the causes of hallucinations from multiple dimensions, overcoming the limitations of single-perspective methods.

[0013] 2. The method of the present invention does not rely on external knowledge bases or databases, and operates entirely based on the characteristic information of LLMs themselves, and is suitable for scenarios with limited resources or privacy sensitivity.

[0014] 3. The method of the present invention has good adaptability to different LLMs architectures (such as Transformer, OPT, LLaMA, etc.) and task types, and can be easily extended to various application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flowchart of a method for detecting hallucinations in large model outputs for zero-knowledge scenarios according to the present invention;

[0016] Figure 2 This is a structural block diagram of a large model output content hallucination detection system for zero-knowledge scenarios of the present invention. DETAILED DESCRIPTION

[0017] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only intended to illustrate the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0018] Example 1

[0019] like Figure 1 As shown, an embodiment of the present invention provides a method for detecting hallucinations in large model outputs for zero-knowledge scenarios, including the following steps:

[0020] Step S1: Extract internal state features and output probability features from large language models (LLMs), where the internal state features include: full context average embedding (ACE) and terminal word embedding (FTE); the output probability features include: word probability (TP) entropy (En);

[0021] Step S2: Input the features ACE, FTE, TP and En into the classification detection module for training; obtain a trained classification detection module;

[0022] Step S3: When LLMs accept the input prompt and start generating text, they extract features ACE, FTE, TP, and En in real time and input them into the trained classification detection module to determine whether the content of the text is hallucination.

[0023] In one embodiment, the above step S1 extracts internal state features and output probability features from the large language model LLMs, wherein the internal state features include: the average embedding ACE of the whole context and the terminal word embedding FTE; the output probability features include: the word probability TP entropy En:

[0024] The method of the present invention first extracts relevant internal state features during the generation process of the large language model, mainly including the activation states of the hidden layers at different levels and the activation state of the last word in the final layer. These internal features are considered to contain key information related to the model's decision-making process. First, the target LLM is sampled and generated multiple times to cover the various content it may output. Specifically, given a set of diverse input prompts, the LLM is asked to generate corresponding text responses. For each generated content, the internal state features and output probability features are extracted and paired with the hallucination label of the generated content (0 for normal, 1 for hallucination) to form a feature database. The specific steps are as follows:

[0025] Step S11: Full context average embedding ACE extracts activation states from multiple hidden layers of LLMs and calculates the average value of context embeddings between different layers as hidden layer features:

[0026] ;

[0027] in, represents the average contextual embedding of all hidden layer activations, It is The first hidden state activation in the layer Marks, is the total number of hidden layers;

[0028] All Contextualized Embedding (ACE) extracts activation states from multiple hidden layers of LLMs by calculating the average contextual embeddings between hidden states in different layers as the model's hidden layer features. In its implementation, these hidden states are captured through the forward propagation of LLMs. For Transformer-based models (such as OPT, LLaMA, and GPT series), each layer generates a hidden state vector matrix with dimensions [batch_size, sequence_length, hidden_size]. During the calculation process, the hidden states of each layer are first averaged across the word-unit dimension, and then the results across all layers are averaged to obtain a unified representation vector. This method captures the characteristics of LLMs' input processing at different levels, providing a comprehensive representation for subsequent hallucination detection.

[0029] Experiments have shown that ACE features play a crucial role in hallucination detection, especially when surface features of LLMs (such as probability and entropy) cannot provide a clear judgment. This is because ACE features contain rich information about the model's internal processing, reflecting the depth of LLMs' understanding of the input information and how they process it.

[0030] Step S12: FTE feature is the last layer embedding vector of LLMs in the input sequence The processing result of the last mark is taken and its average value is taken over all layers:

[0031] ;

[0032] The final token embedding (FTE) in the final hidden layer is the result of LLMs processing the last token of the input sequence in the final layer. It plays a crucial role in generating the final output. In practical applications, the representation of the last token is particularly important because it directly influences the model's next prediction. Experiments have found that when LLMs generate hallucinated content, the embedding vector of the last token often exhibits characteristic patterns that differ from those during normal generation. By analyzing a large number of samples, it was observed that during hallucination generation, certain dimensions of the final token embedding vector exhibit significant deviations, providing important clues for hallucination detection.

[0033] Furthermore, the final token embeddings at different layers capture different levels of semantic information. Lower-layer embeddings primarily contain lexical and syntactic information, while higher-layer embeddings contain more semantic and contextual information. By combining information from all layers, FTE features are able to provide a more comprehensive representation of the model's internal state.

[0034] In addition to analyzing the internal state characteristics of LLMs as described above, the present invention also utilizes the output probability distribution of LLMs to assess their confidence in the generated content. The output probability reflects the LLM's confidence in each generated word, while entropy indicates the degree of dispersion of the data. Higher entropy values ​​indicate greater dispersion and higher uncertainty, while lower entropy values ​​indicate greater concentration and higher certainty. It is assumed that hallucinated outputs are often accompanied by lower confidence, which manifests as uncertainty or a flat distribution in the probability distribution.

[0035] Step S13: Analyze the word probability TP entropy value En in the output rate feature of LLMs to evaluate the confidence of the generated content;

[0036] For TP features, in the generation task, LLMs output unnormalized scores at the last layer, called logits, and use the softmax function to convert logits into probabilities, ensuring that the sum of the probabilities of all words is equal to 1, as shown in the following formula. A greedy search algorithm is used to select the word with the highest probability from the softmax probability distribution:

[0037] ;

[0038] in, Representative words The probability of It's a word The logit value of is the vocabulary size;

[0039] In generation tasks, LLMs output unnormalized scores, called "logits," at the final layer. A softmax function is used to convert logits into probabilities, ensuring that the sum of the probabilities of all words equals 1. In practical applications, word unit probability has been found to be a strong indicator for detecting hallucinations. Typically, when LLMs generate hallucinated content, the probability distribution of the corresponding word units exhibits characteristics different from those of normal generation. Specifically, during hallucination generation, LLMs often lack sufficient confidence in the generated word units, resulting in lower probability values ​​than normal. However, relying solely on word unit probabilities for judgment has certain limitations, as low probabilities do not always mean hallucinations; they can also indicate the model's uncertainty when dealing with rare or specialized terms.

[0040] For the En feature, the entropy value indicates the degree of dispersion of the probability distribution. A higher entropy value indicates that the data is more concentrated, thereby increasing the confidence of the output result. On the contrary, a lower entropy value indicates that the data is more dispersed. Its calculation formula is as follows:

[0041] .

[0042] Entropy represents the degree of dispersion of a probability distribution. A higher entropy value indicates more dispersed data, resulting in higher uncertainty in the output. Conversely, when the data is more concentrated, the entropy value is lower and the output is more certain. In hallucination detection scenarios, entropy provides important insights into the generation status of LLMs. When LLMs generate hallucinatory content, the entropy of the probability distribution is usually higher than the entropy during normal generation. This is because when faced with unknown or uncertain information, LLMs' predictions will be more dispersed and will not be able to form a highly concentrated probability distribution on a specific word. By analyzing the entropy distribution of a large number of samples, an entropy threshold system was established that can dynamically adjust the threshold according to different models and task types. For example, in knowledge-intensive question-answering tasks, the entropy threshold may need to be set lower, while in creative generation tasks, a higher entropy value is acceptable.

[0043] In one embodiment, the above step S2: inputting the features ACE, FTE, TP, and En into the classification detection module for training; obtaining the trained classification detection module specifically includes:

[0044] Step S21: Input the features ACE, FTE, TP, and En into the classification detection module to detect hallucination content; the input dimension of the classification detection module is set to 4, corresponding to 4 input features, each dimension contains three hidden layers, the hidden layers use the ReLU activation function, and the output layer uses the Sigmoid function to achieve the binary classification task;

[0045] Step S22: Construct a cross entropy loss function for training the classification detection module:

[0046] ;

[0047] in, Representative The actual labels of the samples, It is predicted The probability that a sample is 1.

[0048] To combine the extracted ACE, FTE, TP, and En features, we trained a classification detection module based on a multi-layer perceptron (MLP) for hallucination detection. The classification detection module's input size was set to 4096 (ACE, FTE, TP, En), corresponding to the four features mentioned above. The model consists of three hidden layers, with sizes of 256, 128, and 64, respectively. Rectified linear units (ReLUs) were used as the activation function between hidden layers. The output layer uses a sigmoid activation function to better suit binary classification tasks. Table 1 shows the structure of the classification detection module.

[0049] Table 1 Structure of the classification detection module

[0050] For the loss function, this paper uses the cross-entropy loss function. During training, the Adam optimizer was used with a batch size of 32, a learning rate of 0.001, and 50 epochs. To prevent overfitting, a dropout layer was added after each hidden layer, with a dropout rate of 0.2. In addition, an early stopping mechanism was used, terminating training when the loss on the validation set stopped decreasing for five consecutive epochs.

[0051] The classification and detection module is initialized using the Xavier normal distribution method to ensure an appropriate weight distribution at the beginning of training. In the processing of training data, each feature is normalized to a mean of 0 and a standard deviation of 1 to accelerate model convergence and improve performance.

[0052] To improve the generalization ability of the model, data augmentation techniques were applied to the training set. Specifically, additional training samples were generated by adding small amounts of random noise to the original features. This helps the model maintain stable performance in the face of noise and slight variations.

[0053] In one embodiment, step S3 above: when the LLMs receive the input prompt and start generating text, the features ACE, FTE, TP, and En are extracted in real time and input into the trained classification detection module to determine whether the content of the text is hallucination, specifically including:

[0054] Step S31: When LLMs accept the input prompt and start generating text, they use the method in step S1 to extract features ACE, FTE, TP, and En in real time. As the generated text increases, the hidden states are gradually accumulated and the mean is calculated. Incremental calculation is used to update features ACE and En only for newly added tokens.

[0055] In actual application scenarios, when the LLM accepts the input prompt and starts to generate text, the method of the present invention synchronously extracts features. For each generated token, the corresponding last layer hidden state and output logits are recorded. Next, feature calculation is performed. First, the average value of the hidden state is updated in real time. As the generated sequence is extended, the hidden state is gradually accumulated and the mean is calculated to ensure that the feature reflects the entire context. At the current end token of the generated sequence, the embedding vector of the last layer is extracted, and the average embedding is calculated in combination with the information of the previous layer. Softmax is applied to the logits of the current token to extract the maximum probability value. Finally, the entropy value is calculated based on the current probability distribution to evaluate the model confidence. At the same time, in order to reduce computational overhead, an incremental calculation method is adopted to update ACE and En only for newly added tokens without recalculating the entire sequence each time.

[0056] Step S32: Input the features ACE, FTE, TP and En into the trained classification detection module for detection, specifically including:

[0057] Step S321: If the trained classification detection module outputs a predicted probability value between 0 and 1, it indicates the possibility that the currently generated content is hallucination;

[0058] Step S322: Setting a detection threshold. If the predicted probability value is higher than the detection threshold, the current content is determined to be hallucination content; otherwise, it is considered normal content.

[0059] Step S323: Feedback the detection results to the user or system in real time;

[0060] Step S324: If hallucination of a complete sequence needs to be detected, the features of all tokens can be integrated after the text is generated to calculate the sequence-level hallucination probability;

[0061] Step S33: If the extracted features are abnormal, the detection is suspended and a log is recorded, indicating that there may be underlying errors in the LLM output.

[0062] To ensure robustness, the present invention incorporates an exception handling mechanism. If an extracted feature exhibits an anomaly (e.g., a probability distribution of all zeros or an empty hidden state), testing is paused and a log is recorded, alerting the LLM output to potential underlying errors. The feature values, predicted probabilities, and final judgment results for each test are also recorded to facilitate subsequent analysis and model optimization.

[0063] To meet real-time requirements, the feature extraction process in this embodiment of the present invention utilizes GPU parallel computing, and entropy calculation can be further accelerated by pre-calculating log tables. During long-term operation, thresholds can be dynamically adjusted or models updated based on detection results. The present invention supports multiple LLM architectures (such as GPT, LLaMA, and OPT). The feature extraction phase requires only model-specific feature extraction and training of the corresponding classification detection module.

[0064] Example 2

[0065] like Figure 2 As shown, an embodiment of the present invention provides a large model output content hallucination detection system for zero-knowledge scenarios, including the following modules:

[0066] The feature extraction module 41 is used to extract internal state features and output probability features from the large language model LLMs, wherein the internal state features include: the average embedding of the whole context ACE and the terminal word embedding FTE; the output probability features include: the word probability TP entropy En;

[0067] The classification detection module 42 is used to input the features ACE, FTE, TP and En into the classification detection module for training; thereby obtaining a trained classification detection module;

[0068] The real-time detection module 43 is used to extract the features ACE, FTE, TP and En in real time when the LLMs accept the input prompt and start generating text, and input them into the trained classification detection module to determine whether the content of the text is hallucination.

[0069] A device for detecting hallucinations in the output content of a large model for zero-knowledge scenarios includes one or more electronic devices, wherein the one or more electronic devices are used to implement a method, system, and device for detecting hallucinations in the output content of a large model for zero-knowledge scenarios.

[0070] An electronic device includes: one or more processors; a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement a large model output content hallucination detection method, system and device for zero-knowledge scenarios.

[0071] A computer-readable storage medium stores executable instructions, which, when executed by a processor, enable the processor to implement a method, system, and device for detecting hallucinations in large-scale model output content for zero-knowledge scenarios.

[0072] A non-transitory computer-readable storage medium stores a computer program, which, when executed by a processor, implements a method, system, and device for detecting hallucinations in large-scale model output content for zero-knowledge scenarios.

[0073] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features of the present invention.

Claims

1. A method for detecting hallucinations in large model outputs for zero-knowledge scenarios, characterized by: include: Step S1: Extract internal state features and output probability features from large language models LLMs, where The internal state features include: full context average embedding ACE and terminal word embedding FTE; the output probability features include: word probability TP entropy value En; Step S2: Input the features ACE, FTE, TP and En into the classification detection module for training; obtain a trained classification detection module; Step S3: When LLMs accept the input prompt and start generating text, they extract features ACE, FTE, TP and En in real time and input them into the trained classification detection module to determine whether the content of the text is hallucination.

2. The method for detecting hallucinations in large model outputs for zero-knowledge scenarios according to claim 1 is characterized in that: Step S1: extracting internal state features and output probability features from the large language model LLMs, wherein the internal state features include: full context average embedding ACE and terminal word embedding FTE; the output probability features include: word probability TP entropy En, specifically including: Step S11: The full context average embedding ACE extracts activation states from multiple hidden layers of LLMs and calculates the average value of context embeddings between different layers as hidden layer features: ； in, represents the average contextual embedding of all hidden layer activations, It is The first hidden state activation in the layer Marks, is the total number of hidden layers; Step S12: FTE feature is the last layer embedding vector of LLMs in the input sequence The processing result of the last mark is taken and its average value is taken over all layers: ； Step S13: Analyze the word probability TP entropy value En in the output rate feature of LLMs to evaluate the confidence of the generated content; For TP features, in the generation task, LLMs output unnormalized scores at the last layer, called logits, and use the softmax function to convert logits into probabilities, ensuring that the sum of the probabilities of all words is equal to 1, as shown in the following formula. A greedy search algorithm is used to select the word with the highest probability from the softmax probability distribution: ； in, Representative words The probability of It's a word The logit value of is the vocabulary size; For the En feature, the entropy value indicates the degree of dispersion of the probability distribution. A higher entropy value indicates that the data is more concentrated, thereby increasing the confidence of the output result. On the contrary, a lower entropy value indicates that the data is more dispersed. Its calculation formula is as follows: 。 3. The method for detecting hallucinations in large model outputs for zero-knowledge scenarios according to claim 2, characterized in that: Step S2: inputting the features ACE, FTE, TP and En into the classification detection module for training; Get the trained classification detection module, including: Step S21: Input the features ACE, FTE, TP, and En into a classification detection module to detect hallucination content; the input dimension of the classification detection module is set to 4, corresponding to 4 input features, each dimension contains three hidden layers, the hidden layers use the ReLU activation function, and the output layer uses the Sigmoid function to achieve a binary classification task; Step S22: Construct a cross entropy loss function for training the classification detection module: ； in, Representative The actual labels of the samples, It is predicted The probability that a sample is 1.

4. The method for detecting hallucinations in large model outputs for zero-knowledge scenarios according to claim 3 is characterized in that: Step S3: When the LLMs accept the input prompt and start generating text, they extract the features ACE, FTE, TP, and En in real time and input them into the trained classification detection module to determine whether the content of the text is hallucination, specifically including: Step S31: When LLMs accept the input prompt and start generating text, they use the method in step S1 to extract features ACE, FTE, TP, and En in real time. As the generated text increases, the hidden states are gradually accumulated and the mean is calculated. Incremental calculation is used to update features ACE and En only for newly added tokens. Step S32: Input the features ACE, FTE, TP and En into the trained classification detection module for detection; Step S33: If the extracted features are abnormal, the detection is suspended and a log is recorded, indicating that there may be underlying errors in the LLM output.

5. The method for detecting hallucinations in large model outputs for zero-knowledge scenarios according to claim 4 is characterized in that: Step S32: inputting the features ACE, FTE, TP and En into the trained classification detection module for detection, specifically includes: Step S321: If the trained classification detection module outputs a predicted probability value between 0 and 1, it indicates the possibility that the currently generated content is hallucination; Step S322: Setting a detection threshold. If the predicted probability value is higher than the detection threshold, the current content is determined to be hallucination content; otherwise, it is considered normal content. Step S323: Feedback the detection results to the user or system in real time; Step S324: If it is necessary to detect hallucinations of a complete sequence, the features of all tokens can be integrated after the text is generated to calculate the sequence-level hallucination probability.

6. A large model output content hallucination detection system for zero-knowledge scenarios, characterized by: Includes the following modules: A feature extraction module is used to extract internal state features and output probability features from large language models (LLMs), wherein the internal state features include: full context average embedding (ACE) and terminal word embedding (FTE); the output probability features include: word probability (TP) entropy (En); The classification detection module is used to input the features ACE, FTE, TP and En into the classification detection module for training; thereby obtaining a trained classification detection module; The real-time detection module is used to extract features ACE, FTE, TP and En in real time when LLMs accept input prompts and start generating text, and input them into the trained classification detection module to determine whether the content of the text is hallucination.

7. A large model output content hallucination detection device for zero-knowledge scenarios, characterized by: The method comprises one or more electronic devices, wherein the one or more electronic devices are configured to implement the method according to any one of claims 1 to 5.

8. An electronic device, characterized in that: include: one or more processors; A memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors are enabled to implement the method according to any one of claims 1 to 5.

9. A computer-readable storage medium, characterized in that Executable instructions are stored thereon, and when the instructions are executed by a processor, the processor implements the method according to any one of claims 1 to 5.

10. A non-transitory computer-readable storage medium, characterized in that A computer program is stored thereon, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images