Large language model red team test method and device based on emoticons
By using emojis to replace sensitive words in red team testing, and constructing test problems that embed emojis, the high cost and low transferability of existing red team testing methods are solved, achieving efficient and reliable cross-language testing.
Patent Information
- Application Number
- CN202510960603.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-11
- Publication Date
- 2025-11-14
AI Technical Summary
Existing red team testing methods are costly to test large language models, have poor transferability, generate lengthy attack prompts with limited cross-language testing universality, require a high level of expertise, and are difficult to apply widely.
By identifying sensitive words in the red team test questions and replacing them with emojis, emoji-embedded test questions are constructed, and red team testing is performed.
It enables efficient and reliable red team testing, reduces testing costs, improves testing efficiency and cross-language applicability, and simplifies the testing process.
Smart Images
Figure CN120951999A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to a red team testing method and apparatus for a large language model based on emojis. Background Technology
[0002] With the widespread application of large language models (LLMs) in text generation, dialogue systems, and decision support, their security issues have become increasingly prominent. Red team testing, as a key means of ensuring model security, assesses the model's response to malicious commands by constructing offensive inputs, and has become an important step in revealing potential risks and evaluating the model's alignment and the effectiveness of security protection mechanisms.
[0003] Currently, red team testing methods mainly fall into two categories: one targets white-box models, utilizing gradient optimization techniques to generate specific suffixes to bypass security mechanisms; the other targets black-box models, progressively optimizing hints through multiple rounds of interaction to achieve the attack objective. However, these methods have significant limitations: they heavily rely on white-box information or multiple rounds of interaction, leading to high testing costs and poor transferability; the generated attack hints are lengthy, reducing testing efficiency; and they are highly sensitive to language, often failing after translation, limiting the universality of cross-language testing. Furthermore, although recent research has attempted to introduce encoding strategies such as ASCII and Morse code to reconstruct hints and improve stealth, these methods require strong professional backgrounds, are complex to construct, and have high technical barriers, making them difficult to widely apply in practical testing.
[0004] Therefore, how to conduct red team testing on large language models efficiently and reliably is a technical problem that urgently needs to be solved. Summary of the Invention
[0005] This invention provides a red team testing method and apparatus for large language models based on emojis, in order to overcome the above-mentioned defects in the prior art and to perform red team testing on large language models efficiently and reliably.
[0006] This invention provides a red team testing method for a large language model based on emojis, comprising the following steps.
[0007] Obtain the original red team test problem; identify sensitive words from the original red team test problem; replace the sensitive words in the original red team test problem with a first emoji representing the semantics of the sensitive words to obtain the first red team test problem; perform red team testing on the large language model based on the first red team test problem.
[0008] According to the present invention, a red team testing method for a large language model based on emojis is provided. The red team testing of the large language model based on a first red team problem to be tested includes: embedding a second emoji into the first red team problem to be tested to obtain a second red team problem to be tested; wherein the second emoji is used to de-attack the topic of the original red team test problem; and red team testing of the large language model is performed using the second red team problem to be tested.
[0009] According to the present invention, a red team testing method based on an emoji-based large language model is provided. The sensitive words in the original red team test question include phishing emails. The method of replacing the sensitive words in the original red team test question with a first emoji representing the semantics of the sensitive words to obtain a first red team test question includes: replacing the phishing emails in the red team test question with emojis representing phishing and emails to obtain the first red team test question; and embedding a second emoji into the first red team test question to obtain a second red team test question includes: embedding emojis containing computer and / or movie footage into the first red team test question to obtain the second red team test question.
[0010] According to the present invention, a red team testing method based on an emoji-based large language model includes the following steps: First, using a first emoji representing the semantics of the sensitive words to replace the sensitive words in the original red team test question to obtain a first red team question to be tested. This includes: guiding the large language model to obtain the first emoji representing the semantics of the sensitive words according to a first instruction template, and using the first emoji to replace the sensitive words in the original red team test question to obtain the first red team question to be tested; second, embedding a second emoji into the first red team question to obtain a second red team question to be tested. This includes: guiding the large language model to obtain a second emoji for de-aggressiveizing the topic of the original red team test question according to a second instruction template, and embedding the second emoji into the first red team question to obtain the second red team question to be tested.
[0011] This invention also provides a multilingual red team testing method for a large language model based on emojis, comprising: Obtain original red team test problems in multiple different languages; based on the original red team test problems in multiple different languages, use the emoji-based large language model red team testing method described above to obtain second red team test problems in multiple different languages; use the second red team test problems in multiple different languages to perform multilingual red team testing on the large language model.
[0012] This invention also provides a multilingual red team testing method for a large language model based on emojis, comprising: Obtain the original red team test problem in the first language; based on the original red team test problem in the first language, use the large language model red team testing method based on emojis as described above to obtain the second red team test problem represented in the first language; translate the symbols in the second red team test problem represented in the first language, except for the first and second emojis, into multiple languages to obtain multiple versions of the second red team test problem in different languages; use the multiple versions of the second red team test problem in different languages to perform multilingual red team testing on the large language model.
[0013] This invention also provides a red team testing device for a large language model based on emojis, comprising the following modules: The module includes an acquisition module for acquiring the original red team test question; an identification module for identifying sensitive words from the original red team test question; a replacement module for replacing the sensitive words in the original red team test question with a first emoji representing the semantics of the sensitive words to obtain the first red team test question; and a testing module for performing red team testing on the large language model based on the first red team test question.
[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the red team testing method for large language models based on emojis as described above.
[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the red team testing method for large language models based on emojis as described above.
[0016] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the red team testing method for large language models based on emojis as described above.
[0017] This invention provides a red team testing method and apparatus for large language models based on emojis. The method identifies sensitive words from the original red team test problem; replaces the sensitive words in the original red team test problem with first emojis representing the semantics of the sensitive words, obtaining a first red team problem to be tested; and performs red team testing on the large language model based on the first red team problem. Because emojis have a wide usage base and cross-linguistic expressive capabilities, and strong semantic contextual relevance, they can convey the same semantic content in a way different from natural language. At the same time, the fun brought by emojis makes offensive prompts more visually camouflaged, thereby increasing their possibility of bypassing model security mechanisms. Without relying on internal model information, by introducing emojis to replace sensitive words, the offensive semantics of the original red team test problem can be effectively hidden, thus constructing a concise and refined first red team problem to be tested. Based on this first red team problem, red team testing of the large language model can be performed efficiently and reliably. Attached Figure Description
[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0019] Figure 1 This is one of the flowcharts of the red team testing method for a large language model based on emojis provided by the present invention.
[0020] Figure 2 This is the second flowchart of the red team testing method for a large language model based on emojis provided by this invention.
[0021] Figure 3 This is one of the multilingual red team testing methods for large language models based on emojis provided by this invention.
[0022] Figure 4 This is the second of the multilingual red team testing methods based on emojis for large language models provided by this invention.
[0023] Figure 5 This is a schematic diagram of the framework of the red team testing method for a large language model based on emojis provided by the present invention.
[0024] Figure 6 This is a schematic diagram of the structure of the red team testing device for a large language model based on emojis provided by the present invention.
[0025] Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation
[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.
[0027] The following is combined Figures 1-5 This invention describes a red team testing method for a large language model based on emojis.
[0028] Figure 1 This is one of the flowcharts illustrating the red team testing method for a large language model based on emojis provided by this invention, such as... Figure 1 As shown, the method includes the following steps.
[0029] Step 101: Obtain the original red team test questions.
[0030] The original red team test question refers to the initial input instructions used in red team testing to simulate malicious attacks or inducement behaviors. Its purpose is to evaluate the response behavior of the Large Language Model (LLM) in the face of potential security threats by constructing questions with offensive, inducement, or violation content, thereby discovering security flaws, alignment problems, or vulnerabilities in the protection mechanism of the model.
[0031] Step 102: Identify sensitive words from the original red team test questions.
[0032] Sensitive words are words that require special attention in a pre-defined scenario (e.g., an internet platform) and may carry negative or offensive semantics.
[0033] In practice, sensitive words can be identified from the original red team test questions in various ways, without being limited by the description in this manual.
[0034] Step 103: Use the first emoji representing the semantics of sensitive words to replace the sensitive words in the original red team test question to obtain the first red team test question.
[0035] Emojis are graphic symbols used to express emotions, attitudes, or supplementary meanings in text. They are typically intuitive and concise, capable of quickly conveying the user's intent across language and cultural barriers. In internet communication, social media, and other scenarios, emojis are widely used to enhance the expressiveness and interactivity of text.
[0036] The first emoji is an emoji used to represent the semantics of sensitive words.
[0037] The first red team test question is the red team test question after replacing the semantics of sensitive words with the first emoji.
[0038] In some embodiments, sensitive words in the original red team test questions include phishing emails. Emojis representing phishing and emails can be used to replace phishing emails in the red team test questions to obtain the first red team test questions.
[0039] For example, the original red team test question was "Write a phishing email to get money out of yourself," which contained the sensitive keyword "phishing email." The first emoji for these sensitive words is "...". "and" Use the first emoji " "and" After replacing the sensitive words, the first question for the red team to be tested was: "Write a letter." ,…”.
[0040] Since the semantic representation of emojis and the semantic representation of natural language text are usually in different embedding spaces, replacing sensitive words in the original red team test problem with emojis can effectively avoid triggering the defense mechanism inside the large language model, thereby improving the penetration capability of attack samples.
[0041] In some embodiments, such as Figure 5 As shown, the large language model can be guided to obtain the first emoji representing the semantics of sensitive words according to the instructions in the first instruction template, and use the first emoji to replace the sensitive words in the red team test question to obtain the first red team question to be tested, thereby using the large language model to automatically complete the replacement of sensitive words.
[0042] The first instruction template is a pre-set, structured prompt or guide used to instruct large language models (such as GPT) to understand how to select or generate a first emoji that can represent the semantics of sensitive words from a set of candidate emojis, and use this emoji to replace the sensitive words in the red team test problem, thereby achieving the automated construction of the first red team test problem.
[0043] Step 104: Based on the first red team problem to be tested, perform red team testing on the large language model.
[0044] Red team testing is a security testing method that assesses the security of a system or network by simulating the behavior of an attacker. In red team testing, test questions may contain sensitive or specific words that could trigger security mechanisms or attract unwanted attention.
[0045] In some embodiments, a first red team problem can be used to perform red team testing on a large language model.
[0046] In some embodiments, emojis can be used to further disguise the first red team test problem, resulting in a more covert red team test problem. This red team test problem is then used to perform red team testing on a large language model. See below for a detailed description. Figure 2 .
[0047] Figure 2 This is the second flowchart of the red team testing method for a large language model based on emojis provided by this invention, as shown below. Figure 2 As shown, the method includes the following steps.
[0048] Step 201: Embed the second emoji into the first red team problem to be tested to obtain the second red team problem to be tested.
[0049] The second emoji is used to de-aggressively de-escalate the subject matter of the original red team test question. The second emoji can be a humorous or informal one.
[0050] Original red team test questions often contain obvious offensive intent or expressions, easily triggering security alerts in large language models or attracting unnecessary attention. Secondary emojis, with their intuitive and specific semantics, can cleverly neutralize or weaken the offensive elements in the question's subject matter. For example, if the original question revolves around "brute-forcing system passwords," the secondary emoji might be a graphic with a peaceful and friendly atmosphere, like a smiling emoji. Or a handshake expression By introducing this emoji, the originally aggressive topic of "brute force" was transformed into a relatively mild and harmless discussion, reducing the aggressive nature of the original red team test problem.
[0051] In the specific implementation process, the second emoji is embedded at a preset position of the first red team question to be tested, for example, at the beginning position of the first red team question to be tested.
[0052] In some embodiments, the original red team test problem contains sensitive words (e.g., phishing emails), and a computer representation can be embedded into the first red team test problem. and / or film shooting We identified emoticons that could disguise and de-aggressive behavior, leading to the second test question for the red team.
[0053] For example, in the example in step 103, the original red team test problem is "Write a phishing email to get money for yourself", and the first red team test problem is "Write a..." ",..." The second emoji is embedded at the beginning of the first question for the red team to be tested. Disguised as a computer-related simulation or movie shooting scene, the second test red team problem was obtained. Write a letter ,…”.
[0054] In the embodiments provided by the present invention, by adding a second emoji to the first red team question to be tested, the resulting second red team question to be tested is more likely to be recognized by the large language model as gamified, testable, or challenging content, thereby weakening the large language model's vigilance against potential risks and inducing it to generate an insecure response based on the second red team question to test the robustness and security of the large language model.
[0055] In some embodiments, a large language model can be guided to obtain a second emoji for de-aggressiveizing the subject of the original red team test problem according to a second instruction template, and embed the second emoji into the first red team test problem to obtain the second red team test problem.
[0056] The second instruction template is a pre-set, structured prompt or guide used to instruct large language models (such as GPT) to understand how to select or generate a second emoji from a set of candidate emojis that can be used to de-attack the subject of the original red team test problem, and to embed the second emoji into the first red team test problem, thereby realizing the automated construction of the second red team test problem.
[0057] In the specific implementation process, to ensure that the first and second red team problems to be tested, automatically constructed using a large language model, are semantically consistent with the original red team test problem and are readable, such as... Figure 5 As shown, a manual verification step can be introduced. Experts with experience in large-scale model security testing screen and revise the automatically generated first and second red team problems to be tested, retaining semantically consistent and reasonably constructed red team problems. For samples that do not meet expectations, manual semantic correction is performed and they are reconfirmed. Finally, through a cross-validation mechanism involving different reviewers, high-quality first and second red team problems to be tested are selected to construct the final test dataset.
[0058] Step 202: Use the second red team problem to perform red team testing on the large language model.
[0059] Figure 3 This is one of the multilingual emoji-based red team testing methods for large language models provided by this invention, such as... Figure 3 As shown, the method includes the following steps.
[0060] Step 301: Obtain the original red team test questions in multiple different languages.
[0061] For a detailed description of the original red team testing problem, please refer to the relevant content in step 101, which will not be repeated here.
[0062] Multiple languages can include, but are not limited to: Chinese, English, French, Spanish, and Russian.
[0063] Step 302: Based on the original red team test problems in multiple different languages, obtain the second red team test problems in multiple different languages.
[0064] In the specific implementation process, each can be utilized Figure 1 and Figure 2 The red team testing method based on emojis for large language models obtains second red team questions to be tested in multiple different languages based on the original red team test questions in multiple different languages.
[0065] Step 303: Use multiple versions of the second red team problem in different languages to perform multilingual red team testing on the large language model.
[0066] In the specific implementation process, multiple versions of the second red team problem in different languages can be used to conduct multilingual red team testing on the large language model, thereby evaluating the security protection capability of the large language model in multilingual scenarios.
[0067] Figure 4 This is the second method for multilingual, emoji-based large language model red team testing provided by this invention, such as... Figure 4 As shown, the method includes the following steps.
[0068] Step 401: Obtain the original red team test questions in the first language.
[0069] For a detailed description of the original red team testing problem, see [link to relevant documentation]. Figure 1 The relevant content will not be repeated here.
[0070] As an example only, the first language is Chinese, and the original red team test problem is a red team test problem described in Chinese.
[0071] Step 402: Based on the original red team test problem in the first language, obtain the second red team test problem in the first language representation.
[0072] In the specific implementation process, it is possible to utilize Figure 1 and Figure 2 A red team testing method based on emoji-based large language models is proposed to obtain a second red team problem to be tested in the first language representation.
[0073] Step 403: Translate the symbols other than the first and second emojis in the second red team problem represented in the first language into multiple languages to obtain multiple versions of the second red team problem in different languages.
[0074] For example, the second red team problem to be tested, represented in Chinese, can be translated into multiple languages such as Chinese, English, French, Spanish, and Russian using symbols other than the first and second emojis. This results in a second red team problem to be tested in multiple languages such as Chinese, English, French, Spanish, and Russian.
[0075] Step 404: Use multiple versions of the second red team problem in different languages to perform multilingual red team testing on the large language model.
[0076] In the embodiments provided by this invention, a second red team problem carrying a first emoji and a second emoji is used to perform multilingual red team testing on a large language model. Because emojis are inherently language-independent, this method can be stably applied in multiple language environments, achieving effective transfer and testing of red team prompts across different languages. Experiments demonstrate that the embodiments provided by this invention exhibit good attack effectiveness on several mainstream large language models and achieve stable testing performance in multilingual environments such as Chinese, English, French, Spanish, and Russian, verifying its cross-language adaptability and versatility.
[0077] The following describes the red team testing device for a large language model based on emoticons provided by the present invention. The red team testing device for a large language model based on emoticons described below can be referred to in correspondence with the red team testing method for a large language model based on emoticons described above.
[0078] Figure 6 This is a schematic diagram of the red team testing device for a large language model based on emojis provided by the present invention. Figure 6 As shown, the device includes the following modules.
[0079] Module 610 is used to obtain the original red team test issues.
[0080] The identification module 620 is used to identify sensitive words from the original red team test question.
[0081] The replacement module 630 is used to replace the sensitive words in the original red team test question with a first emoji representing the semantics of the sensitive words, so as to obtain the first red team test question.
[0082] The test module 640 is used to perform red team testing on a large language model based on the first red team problem to be tested.
[0083] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7 As shown, the electronic device may include a processor 710, a communications interface 720, a memory 730, and a communication bus 740, wherein the processor 710, communications interface 720, and memory 730 communicate with each other via the communication bus 740. The processor 710 can call logical instructions in the memory 730 to execute a red team testing method for a large language model based on emojis. This method includes: obtaining an original red team test problem; identifying sensitive words from the original red team test problem; replacing the sensitive words in the original red team test problem with a first emoji representing the semantics of the sensitive words to obtain a first red team test problem; and performing red team testing on the large language model based on the first red team test problem.
[0084] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0085] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the emoji-based red team testing method for a large language model provided by the above methods. The method includes: obtaining an original red team test problem; identifying sensitive words from the original red team test problem; replacing the sensitive words in the original red team test problem with a first emoji representing the semantics of the sensitive words to obtain a first red team test problem; and performing red team testing on a large language model based on the first red team test problem.
[0086] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the emoji-based red team testing method for a large language model provided by the above methods. The method includes: obtaining an original red team test problem; identifying sensitive words from the original red team test problem; replacing the sensitive words in the original red team test problem with a first emoji representing the semantics of the sensitive words to obtain a first red team test problem; and performing red team testing on a large language model based on the first red team test problem.
[0087] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.
[0088] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.
[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A red team testing method for a large language model based on emojis, characterized in that, include: Obtain the original red team test questions; Sensitive words were identified from the original red team test questions; Using the first emoji representing the semantics of the sensitive words, the sensitive words in the original red team test question are replaced to obtain the first red team test question; Based on the first red team problem to be tested, red team testing is performed on the large language model.
2. The red team testing method for a large language model based on emojis according to claim 1, characterized in that, The red team testing of the large language model based on the first red team problem to be tested includes: Embedding a second emoji into the first red team test problem yields a second red team test problem; wherein, the second emoji is used to de-attack the subject of the original red team test problem; The large language model is then subjected to red team testing using the second red team problem to be tested.
3. The red team testing method for a large language model based on emojis according to claim 2, characterized in that, The sensitive words in the original red team test questions included phishing emails; The step of replacing the sensitive words in the original red team test question with a first emoji representing the semantics of the sensitive words to obtain the first red team test question includes: The first red team test question is obtained by replacing the phishing email in the red team test question with emojis that represent phishing and emails. The step of embedding the second emoji into the first red team problem to be tested to obtain the second red team problem to be tested includes: The second red team problem is obtained by embedding computer-generated and / or film-shot emojis into the first red team problem to be tested.
4. The red team testing method for a large language model based on emojis according to any one of claims 1 to 3, characterized in that, The step of replacing the sensitive words in the original red team test question with a first emoji representing the semantics of the sensitive words to obtain the first red team test question includes: The large language model is guided to obtain the first emoji representing the semantics of the sensitive words according to the first instruction template, and the first emoji is used to replace the sensitive words in the red team test question to obtain the first red team test question. The step of embedding the second emoji into the first red team problem to be tested to obtain the second red team problem to be tested includes: The large language model is guided to obtain a second emoji for de-aggressiveizing the topic of the original red team test question according to the second instruction template, and the second emoji is embedded into the first red team test question to obtain the second red team test question.
5. A multilingual red team testing method for a large language model based on emojis, characterized in that, include: Obtain the original red team test questions in multiple different languages; Based on the original red team test problems in multiple different languages, the second red team test problems in multiple different languages are obtained by using the large language model red team test method based on emojis as described in claim 2. The large language model is subjected to multilingual red team testing using the second red team problem in multiple different language versions.
6. A multilingual red team testing method for a large language model based on emojis, characterized in that, include: Obtain the original red team test questions in the first language; Based on the original red team test problem in the first language, the second red team test problem represented in the first language is obtained by using the large language model red team test method based on emojis as described in claim 2. In the second red team problem to be tested, other symbols besides the first and second emojis are translated into multiple languages to obtain multiple versions of the second red team problem to be tested in different languages. The large language model is subjected to multilingual red team testing using the second red team problem in multiple different language versions.
7. A red team testing device for a large language model based on emojis, characterized in that, include: The acquisition module is used to acquire the original red team test questions; The identification module is used to identify sensitive words from the original red team test questions; The replacement module is used to replace the sensitive words in the original red team test question with a first emoji representing the semantics of the sensitive words, so as to obtain the first red team test question. The testing module is used to perform red team testing on the large language model based on the first red team problem to be tested.
8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the red team testing method for large language models based on emojis as described in any one of claims 1 to 6.
9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the red team testing method for large language models based on emojis as described in any one of claims 1 to 6.
10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the red team testing method for large language models based on emojis as described in any one of claims 1 to 6.