Systems and methods for large language model operations using smart privacy assistant models

Abstract

A system includes a text receiver, a rule-based privacy check table module, a customized learning module, and an inference module. The text receiver converts the user input into a machine readable format. A rule-based privacy checktable module retrieves privacy specifications from a database, processes these specifications as annotated rules, and then compares them with the input data to create a structured report. The customized learning module uses the fine-tuned BERT to perform privacy infringement classification on the context information, and then generates a local privacy report. The reasoning module integrates multiple rounds of conversations, applies thought chain reasoning and evaluates privacy information in the whole context, and therefore comprehensive cloud-based privacy judgment and reports are provided.

Description

Technical Field

[0001] The present invention generally relates to machine-learning (ML) techniques for large language models (LLMs). More specifically, the present invention relates to systems and methods for operating LLMs using intelligent privacy assistant models. Background Art

[0002] Regarding the current status of large language models (LLMs), the development of language models (LMs) has entered an era of revolutionary progress in natural language processing, which has had a profound impact on society. Currently, large generative language models have made remarkable progress, which integrate various natural language processing tasks to create a unified text generation framework. As these advances change the rules of the ecosystem, more and more researchers, industrial engineers, investors, and interested users have also begun to pay attention to LLMs.

[0003] The scenarios in which LLMs can currently operate and their future potential can be attributed to a variety of factors. First, LLMs have revolutionized many applications, such as machine translation, code generation, chatbots, and question-answering. For example, for some operations, LLMs can even provide professional feedback to help complete the algorithm's program code. These pioneering LLMs applications have received huge profits and future investments. Second, LLMs can also support more complex practical tasks if external tools can be used. Third, it has become a trend for researchers and industrial engineers to use LLMs as autonomous agents to provide planning or tool use. This autonomous agent is able to use reasoning and self-reflection capabilities to use tools to perform given tasks without additional training costs. Based on the above prospects, technology companies have also tried to innovatively integrate LLMs into users' smart devices to improve services.

[0004] However, regarding the privacy issues of LLMs, although LLMs have broad prospects in customized use, if users are allowed to access these models without restrictions, they may also trigger potential privacy risks. For example, when LLMs are deployed on personal edge devices and allow users to freely access multiple user data without restrictions, this privacy risk will be further amplified.

[0005] Existing studies often report privacy vulnerabilities arising from the interaction with LLMs. As language models have evolved, several privacy attacks have been proposed to extract private information. Some studies focus on extracting training data from variants of the GPT-2 language model and show that by prefixing verbatim text, LMs may generate private information that contains the LM training corpus. In addition to directly extracting training data, other studies have studied the conditions in which information leakage occurs in the vectorized sentence representation of the LM and demonstrated that sensitive attributes can be recovered from this representation and even sensitive information can be output. In addition, some studies use misleading characters to jailbreak LLMs through context (such as input prompts) to mislead LLMs to generate unsafe responses, such as generating personally identifiable information. For application-integrated LLMs, injection attacks in prompts, which involve manipulating or inserting harmful content into a given prompt, will result in poor model behavior or bad output. Due to these vulnerabilities, privacy issues often arise.

[0006] In addition, there are also unresolved privacy issues when deploying LLMs on personal devices. There are still several privacy issues that remain unresolved when deploying such LLMs systems into personal devices. First, the definition of private digital data is unclear. Privacy perceptions are influenced by existing regional privacy legislation, social norms, religious beliefs, and technological development. As a result, it is difficult for LLMs to determine which data can be accessed and which data cannot be accessed. Second, existing data cleaning methods lack contextual understanding capabilities. Simply removing predefined privacy patterns may generate additional errors. For example, when a user wants to write a novel about fictional characters, false information about these characters may be considered private information during data processing. Therefore, LLMs may not be able to improvise stories based on false information to meet user needs. Finally, current LLMs systems cannot provide acceptable privacy explanations. Since LLMs are also widely regarded as black box models, these improper interpretations of privacy mechanisms may lead to distrust of service users.

[0007] Therefore, in order to build reliable LLMs for the public interest, system developers address the above privacy issues at the inference stage and implement LLMs systems that are safe to use at the inference stage so that users are willing to use these powerful tools to improve their creativity and productivity. Summary of the invention

[0008] The purpose of the present invention is to provide a device and a method to solve the above problems existing in the prior art.

[0009] According to the above, since the existing security mechanisms are not sufficient to support the LLMs system on smart devices to protect a large amount of personal data, in this disclosure, it is proposed to implement an intelligent privacy assistant based on a three-level privacy protection framework so as to protect the user's personal data during the interaction with the LLMs system. The intelligent privacy assistant (IPA) model proposed in the present invention enables a customized intelligent assistant to make privacy judgments based on existing privacy laws, social norms and user preferences. The IPA model proposed in the present invention can: (1): Provide systematic privacy judgments from a given context to strictly comply with privacy regulations; (2): Provide a reasonable, privacy-oriented explanation of a given privacy judgment; and (3): Enable customized assistants on users’ smart devices at low cost.

[0010] In short, the IPA model proposed in this invention provides three levels of privacy protection. The first level includes rule-based privacy knowledge annotated from existing privacy regulations and social norms. The second level develops a small language model that can be easily deployed on the user's device to understand user preferences and provide privacy judgments with contextual information. The third level uses a large language model in the cloud to improve the reasoning ability of long contexts (such as multi-round conversations) to determine whether there is a privacy violation.

[0011] To address the above challenges, the framework of the IPA model is provided here, and three new technologies and innovative ideas are introduced to make it different from previous technical solutions: 1. Broader privacy coverage: The first-level privacy checklist of the IPA model goes beyond simple scenarios and policies, covering existing privacy regulations (e.g. HIPAA, CCPA, GDPR) through rule-based interpretation. This expands the scope of privacy research and makes the IPA model the first privacy tool to address a wide range of privacy issues. 2. People-centered privacy mechanism: The IPA model, which plays the second level of protection, prioritizes privacy by aligning its small language model with privacy regulations, social norms, and personal preferences. Follow the principles of privacy design and comply with regulations such as GDPR and CCPA. Users can control their own privacy settings and can customize interactions. The IPA model advocates user customization and promotes features such as fine-grained privacy settings, consent management tools, and the ability to opt out of certain data processing activities. By aligning individual users' set preferences, the IPA model is committed to creating an environment where individuals feel they can control their own data and privacy. 3. Enhanced privacy reasoning using LLMs: The IPA model proposes a reasoning module that uses LLMs to obtain the best privacy judgment in a long context, which is the third level of protection. It integrates techniques such as prompts, chain reasoning, and reinforcement learning from human feedback. These state-of-the-art technologies enhance the privacy reasoning capabilities of LLMs, making it possible to effectively identify privacy violations.

[0012] According to a first aspect of the present invention, a system is provided that uses an intelligent privacy assistant model to perform large language model operations. The system includes a text receiver, a rule-based privacy checklist module, a customized learning module, and an inference module. The text receiver is used to parse and preprocess input data from at least one user, thereby converting the text input by the user into a machine-readable format. The rule-based privacy checklist module is configured to query a database to retrieve privacy specifications and process the privacy specifications to create annotated rules, wherein the rule-based privacy checklist module processes the input data from the text receiver and compares it with the annotated specifications and rules, thereby compiling and formatting the comparison results into a readable structured report. The customized learning module receives the input data through the text receiver and is configured to process and interpret the input data to understand its contextual information, wherein the customized learning module includes a fine-tuned BERT module for classifying whether the contextual information constitutes a privacy violation, wherein the classification result of the fine-tuned BERT module is processed locally to generate a local privacy report. A reasoning module receives the input data through the text receiver and is configured to evaluate privacy based on server-side reasoning and integrated multi-round conversation context, wherein the reasoning module extracts the integrated context and uses chain of thoughts (CoT) reasoning to evaluate whether the entire context of the integrated context has privacy information, wherein the reasoning module provides a comprehensive privacy judgment on the interaction occurring in the cloud during the stage of evaluating the privacy information, and also generates a cloud report.

[0013] According to a second aspect of the present invention, a method is provided for performing large-scale language model operations using an intelligent privacy assistant model. The method comprises the following steps: parsing and preprocessing input data from at least one user through a text receiver, thereby converting the text input by the user into a machine-readable format; querying a database to retrieve privacy specifications through a rule-based privacy checklist module, and processing the privacy specifications to create annotated rules, wherein the rule-based privacy checklist module processes the input data from the text receiver and compares it with the annotated specifications and rules, thereby compiling and formatting the comparison results into a readable structured report; receiving the input data through the text receiver through a customized learning module; processing and interpreting the input data through the customized learning module to understand its context; The customized learning module comprises a fine-tuned BERT module for classifying whether the context information constitutes a privacy violation, wherein the classification result of the fine-tuned BERT module is processed locally to generate a local privacy report; the input data is received by the text receiver through the reasoning module; the privacy is evaluated by reasoning and integrating multiple rounds of dialogue contexts based on the server through the reasoning module, wherein the reasoning module extracts the integrated context and uses chain of thoughts (CoT) reasoning to evaluate whether the entire context of the integrated context has privacy information, wherein the reasoning module provides a comprehensive privacy judgment on the interaction occurring in the cloud during the stage of evaluating the privacy information, and also generates a cloud report.

[0014] Through the above configuration, an intelligent privacy assistant (IPA) with a three-level privacy judgment framework is provided, which can make responsible privacy-related judgments through text descriptions combined with existing privacy regulations and personal preference tags. The ultimate goal of the proposed IPA is to use the three-level privacy judgment framework to check the privacy violations of a given text description. In this way, the IPA proposed by the present invention can be easily deployed on the user's smart device to protect the privacy of user data, and privacy assessment can also be enhanced through server-side reasoning. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Embodiments of the present invention will be described in more detail below with reference to the accompanying drawings, in which:

[0016] Figure 1 A schematic diagram showing a model framework of an IPA prototype according to an embodiment of the present invention is shown;

[0017] Figure 2 An exemplary regulatory diagram from the Health Insurance Portability and Accountability Act Privacy Rule is shown;

[0018] Figure 3 A schematic Venn diagram showing privacy coverage according to an embodiment of the present invention; and

[0019] Figure 4 A schematic diagram of the architecture of a system for performing LLMs operations using an IPA model according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0020] In the following description, systems and methods for operating large language models (LLMs) using intelligent privacy assistant models are listed as preferred examples. Those skilled in the art will appreciate that modifications, including additions and / or substitutions, may be made without departing from the scope and spirit of the invention. Specific details may be omitted so as not to obscure the invention; however, the present disclosure is written to enable those skilled in the art to practice the teachings herein without undue experimentation.

[0021] Due to the amazing progress of LLMs, it can greatly improve the ability to effectively solve various downstream NLP tasks, and also improve its integration and response generation performance. Powerful language models trained on large amounts of text data provide excellent accessibility and usability for models and users. Currently, companies in the technology industry regard LLMs as a powerful tool that can drive innovation and improve their services. Many startups and applications integrate LLMs into products and services to attract users and improve convenience. As a result, LLMs have been highly popular with users around the world. However, unrestricted access to these models may lead to potential privacy risks, whether malicious or unintentional. Large language models are usually trained with large amounts of text data, which may include sensitive information and personally identifiable information (PII), such as medical records, private conversations, and financial data. Despite continuous efforts to anonymize and remove identifiable information, data leakage issues occasionally occur during the use of LLMs.

[0022] In this regard, there are several related studies / works targeting the privacy issues of LLMs, which are discussed below.

[0023] 1) Privacy defense technologies used by existing LLMs

[0024] Currently, there are a variety of privacy defense techniques to enhance the privacy protection of LLMs and enhance the robustness of LLMs against privacy attacks. These privacy defenses integrate differential privacy (DP), secure multi-party computation (SMPC), and other privacy goals into LLMs systems. Table 1 lists the previous studies on DP and SMPC. Table 1: Previous studies on DP and SMPC-based LLMs.

[0025] 1.1) Privacy defense technology of existing LLMs - Differential Privacy (DP)

[0026] Differential privacy-preserving model training can be achieved by using a noise optimization algorithm such as DPSGD. In the optimization step, DPSGD can introduce Gaussian noise into the calculated gradient on a per-sample basis using a specified noise scale. This approach can be easily integrated into different models, making it the basis for most privacy-preserving language model research. In summary, existing DP-based LLMs can be divided into four clusters: DP-based pre-training, DP-based fine-tuning, DP-based hint fine-tuning, and DP-based synthetic text generation.

[0027] 1.2) Privacy protection technology of existing LLMs - secure multi-party computation (SMPC)

[0028] In the use case of LLMs, secure multi-party computation is a cryptographic technique that allows multiple parties to jointly train LLMs while maintaining the privacy of their respective data. This approach ensures that each party can contribute its local data to the training process without leaking any sensitive information. Currently, secure multi-party computation is mainly used in the inference phase of LLMs to protect the privacy of model parameters and inference data.

[0029] 1.3) Privacy-defending techniques for existing LLMs - additional privacy goals

[0030] In addition to existing frameworks that can directly integrate existing scripts into LLMs systems, there are several studies on introducing additional privacy goals during the optimization phase of LLMs. For example, one of the related works introduced a model weight editing method and a combination of six defense strategies to prevent data extraction attacks. One of the related studies proposed a training-based method that can identify malicious data samples for backdoor attacks.

[0031] 2) Insufficiency of existing privacy technologies

[0032] This subsection points out the weaknesses of the above privacy-preserving techniques and presents preliminary evaluation results to illustrate the decreased utility of DP.

[0033] For DP-based LLMs, a preliminary evaluation of DP-fine-tuned LLMs is performed using various fine-tuning techniques. More specifically, DP and non-DP fine-tuning are performed on the Multi-Genre Natural Language Inference (MNLI) corpus of the General Language Understanding Evaluation (GLUE) benchmark based on Bidirectional Encoder Representations from Transformers (BERT) and RoBERTa models. In Table 2, preliminary evaluation results in percentage accuracy are reported. The table shows that the average accuracy of DP-fine-tuned LLMs in a single task drops by about 10%. These results indicate that DP-enhanced LLMs lead to a decrease in utility and further contributions are needed in terms of model and hyperparameter fine-tuning. Table 2 Preliminary LLMs performance evaluation results on MNLI.

[0034] For SMPC, existing SMPC techniques cannot be directly applied to the nonlinear layers of LLMs. Existing work has proposed new LLMs architectures that replace nonlinear layers with linear layers. However, due to the different model architectures, the pre-trained weights of LLMs cannot be reused for SMPC techniques. In addition, SMPC brings additional computational costs, which further slows down the inference speed of LLMs. Therefore, there is still some distance before SMPC-based LLMs can actually be used as downstream applications.

[0035] For additional learning-based privacy goals, while introducing a single privacy goal may work well for a specific coverage, these goals cannot be generalized to other situations. On the other hand, privacy is a multifaceted concept that is influenced by various factors such as personal preferences, legal requirements, and cultural norms. Therefore, a broader set of privacy goals is needed to ensure a comprehensive and adaptable framework.

[0036] In the present invention, a model framework of an intelligent privacy assistant (IPA) model is provided, as described below.

[0037] First, an overview of the IPA model is given. Privacy is a widely studied topic in various fields of computer science, and each field has its own definition and expression of privacy. In the field of computer vision, privacy on social platforms is still being studied. In the field of software and security, privacy on mobile applications is still being studied. In the field of Natural Language Processing (NLP), privacy issues related to LLMs are currently being studied. In this disclosure, an IPA model framework is provided to align privacy with existing privacy regulations and social norms.

[0038] Figure 1 Schematic diagram of the model framework of an IPA prototype according to some embodiments of the present invention is shown. Figure 1 As shown, the ultimate goal of the proposed IPA model is to use a three-level privacy judgment framework to check whether a given text description has a privacy violation. The proposed checklist and training model can be easily used on local devices, so that privacy checks can be performed on edge devices at a low cost. In addition, in some embodiments, if there is not enough confidence in the first and second level judgments, the third level reasoning module is used, which uses the basic model deployed on the cloud server to determine the privacy violation.

[0039] In order to implement the model framework of IPA, a three-level architecture was established, including "Level 1: Rule-based Privacy Checklist", "Level 2: Customized Learning Module" and "Level 3: Reasoning Module". They are described below.

[0040] (A): Level 1: Rules-based privacy checklist

[0041] The primary function of this level is to align privacy judgments with existing privacy regulations to have a broader impact.

[0042] Following the concept of contextual integrity, privacy is formulated as an information transformation specification. In each transformation process, three participants are involved: the sender, the receiver, and the information subject. The sender informs the receiver of the subject's attribute information. In this regard, the entire context of the information transmission should be considered to draw conclusions such as whether the information transmission is considered a positive or negative specification. Positive specifications are accepted by privacy regulations, while negative specifications violate privacy laws. With such a formulation and careful annotation, a rule-based system can be easily developed to effectively determine privacy violations based on the user's smart devices without heavy computational burden. Since the checklist is rule-based, it can be expected to have high accuracy for privacy judgments.

[0043] Figure 2Depicted are exemplary specifications from the Health Insurance Portability and Accountability Act (HIPAA) privacy rule. Figure 2 The diagram of FIG gives an example of a logical formula for information transfer based on contextual integrity, which is expressed in English code. After collecting or annotating the specifications from relevant privacy laws, rule-based privacy judgments can be performed by scanning these specifications.

[0044] (B): Level 2: Customized learning modules

[0045] In addition to the first level of rule-based checking, sometimes users may enter false and irrelevant PII that does not violate anyone's privacy. In this regard, the goal should be to send these safe PII to the server for better LLMs services. However, simply deploying a rule-based checklist may not understand complex contextual situations and also consider false information as private information. Therefore, in the present invention, a learning module should be proposed to learn from the given context to determine whether the mentioned PII should be considered confidential information.

[0046] In response to the problem of insufficient context understanding, the present invention proposes a fine-tuned micro-masked language model BERT to determine whether contextual information constitutes a privacy violation. Here, the proposed technical solution does not directly fine-tune the BERT model, but extracts necessary knowledge from the existing large language model to enhance the robustness and performance of the micro-language model. For the cloud server, it is assumed that a base model already exists, which can use supervised fine-tuning or human feedback reinforcement learning (RLHF) to build a (smaller) reward model to learn privacy-invading behaviors with a given description. The reward model can then be used to update the base model through reinforcement learning.

[0047] Then, a customized micro language model can be extracted from the reward model. This small model can be used directly on edge devices such as smartphones to make privacy judgments locally. The micro model can also learn the user's privacy preferences during the use of the phone. The entire learning module will save the personal data it accesses locally to avoid unnecessary data leakage.

[0048] Furthermore, to compensate for the high precision of the manifest, a small model can be trained to maximize its recall to raise privacy alerts.

[0049] (C): Level 3: Reasoning Module

[0050] Both the first and third level judgments are able to perform preliminary privacy violation detection. However, in real life, it is common for users to have long, multi-round conversations with LLMs. Even if each sentence does not contain private information, such multi-round conversations may still leak private information by associating the entire context. Since LLMs already have access to the entire context, LLMs may require additional reasoning modules to determine whether the current context contains private information. If the entire context does not contain private information, such contexts can be safely used as training data for LLMs. Otherwise, LLMs should not be further trained on private data.

[0051] Therefore, a reasoning module with Chain-of-Thought (CoT) reasoning and few-shot contextual learning is constructed to leverage the powerful understanding ability of LLMs. More specifically, a step-by-step reasoning template can be designed to allow LLMs to self-reflect on whether the accessed data and its response results contain personal information. This complex reasoning can further help the IPA model to provide judgments for multi-round dialogue scenarios.

[0052] In addition, if the user wishes to seek privacy judgment from LLMs in the cloud, the edge device can send relevant information to the server, and the server will be prepared to provide step-by-step prompts to provide the corresponding privacy judgment.

[0053] The following description provides further positive effects, including key technical and creative features of the above architecture and mechanisms and practical use cases.

[0054] Key technical and creative features:

[0055] (I): A three-level IPA model with explainable privacy judgment function

[0056] This section focuses on the key technologies and creative features of the IPA model proposed in this invention.

[0057] (I-1): Broader privacy coverage

[0058] Figure 3 A schematic Venn diagram of privacy coverage according to an embodiment of the present invention is shown. As shown in the Venn diagram, the proposed first-level privacy checklist represents a pioneering effort to expand privacy research beyond simple scenarios and policies and cover existing privacy regulations with reasonable interpretation. Unlike previous studies that limit their investigations to limited scenarios and rules, the IPA model proposed by the technical solution of the present invention can perform better and become the first useful privacy tool on a wide range of LLMs.

[0059] (I-2): People-oriented privacy mechanism

[0060] For both Level 1 and Level 2, the IPA model provides a human-centric privacy mechanism to align with existing privacy regulations, social norms, and personal preferences. This alignment mechanism extends the privacy research of LLMs to general scenarios in real life. By incorporating a human-centric privacy alignment mechanism into Level 1 and Level 2 systems, the IPA model ensures that the reasoning-stage interaction with LLMs meets the user's expected results.

[0061] Since the provided IPA model implements privacy-by-design principles, IPA-enhanced LLMs are designed with privacy as a fundamental component and can meet legal requirements. In addition, the proposed second level is considered to respect user preferences to customize the level of privacy required by users when interacting with LLMs. The IPA model advocates for users and promotes features such as fine-grained privacy settings, consent management tools, and the ability to opt-out of certain data processing activities. By respecting personal preferences in this way, the IPA model aims to contribute to creating an environment where individuals can control their own data and privacy.

[0062] In general, the IPA model is a human-centered privacy coordination approach that enables LLMs to operate in a way that respects existing privacy regulations, social norms, and personal preferences. By extending to real-life scenarios, the IPA model can promote trust, transparency, and accountability in the evolving LLMs technology environment.

[0063] (I-3): Using LLMs to Enhance Privacy Reasoning

[0064] In order to obtain the best privacy judgment in the context of long content, in the third level, a reasoning module is proposed with the help of LLMs. In addition, several advanced new techniques such as hints, chain reasoning, and reinforcement learning from human feedback are integrated to enhance the privacy reasoning ability of LLMs in the context of long content. These state-of-the-art techniques enable the IPA model to perform well in determining privacy violations and have strong reasoning capabilities.

[0065] (II): Potential impact

[0066] To understand the applicability of the proposed IPA model, two exemplary scenarios are listed to show how the IPA model can be applied to solve common privacy issues.

[0067] (II-1): Privacy violation detection in the reasoning phase of LLMs

[0068] The most common case is the interaction between LLMs users and LLMs during the inference phase. For LLMs systems developed for users' smart devices, the first-level privacy checklist can be directly used to efficiently determine privacy violations. In this case, the sender should be the user, and the receiver can be the model owner or anyone (for example, if the information is further fine-tuned for LLMs and the LLMs are set to be accessible to any other user). The subject is also the user himself. Therefore, the checklist can be used for rule checking to determine privacy violations. In addition, the second-level local model is also used for privacy violation classification.

[0069] (II-2): Privacy judgment in the context of multi-round dialogue

[0070] Multi-round dialogues require stronger understanding capabilities to judge privacy violations. In this case, LLMs themselves are used to detect privacy violations with a first-level checklist as the main structure. These steps can be completed through self-checking (Constitutional AI) and thought chain reasoning.

[0071] In addition, the second-level module can learn the user's privacy preferences to understand the user's social norms on privacy. This user-level privacy preference is also crucial for multi-round privacy judgments. In short, to solve such a complex problem, basic knowledge (first level), learning (second level), and reasoning ability (third level) are all necessary to understand the context of long content and make correct privacy judgments.

[0072] (II-3): Potential practical usage scenarios

[0073] The provided IPA model provides a unique and necessary solution to the growing privacy and security concerns in the digital age. As individual users increasingly rely on their smart devices to interact with AI models to complete various tasks, such as online banking, customer service, and built-in voice assistants, the need to protect sensitive information becomes critical.

[0074] In addition, through the versatility of the provided IPA models, they can be seamlessly integrated into a variety of smart devices, including smartphones, tablets, smart watches, and even smart home appliances. This compatibility ensures that the provided IPA models can meet diverse needs.

[0075] Figure 4The schematic diagram of the architecture of a system 100 for performing LLMs operation using an IPA model according to an embodiment of the present invention is shown. The system 100 is provided / designed based on the above functions, features and advantages. The system 100 includes a text receiver 102, a rule-based privacy checklist module 110 for a first-level configuration, a customized learning module 120 for a second-level configuration, and a reasoning module 130 for a third-level configuration. In this regard, the first level includes rule-based privacy knowledge annotated from existing privacy regulations and social norms; the second level develops a small language model that can be easily deployed on the user's device to provide customized privacy judgments with contextual information; the third level uses a large language model in the cloud to improve the reasoning ability of the context of long content such as multi-round conversations to determine whether there is a privacy violation.

[0076] The text receiver 102 is responsible for converting the text input by the user into a format that can be processed by various modules of the system 100. In one embodiment, the text input by the user is typed by the user on his smart device (ie, edge device).

[0077] The text receiver 102 obtains raw text input from the user and converts it into a structured format that is compatible with the rule-based privacy checklist module 110, the customized learning module 120, and the reasoning module 130. Specifically, the text receiver 102 parses and preprocesses the input data to ensure that it complies with the input format required by each module (i.e., a machine-readable format). In multiple embodiments, for the rule-based privacy checklist module 110, the text receiver 102 organizes the text into a format suitable for initial privacy assessment; for the customized learning module 120, the text receiver 102 formats the text for the context-aware learning process; for the reasoning module 130, the text receiver 102 prepares the text for high-level reasoning and decision-making tasks.

[0078] The rule-based privacy checklist module 110 receives input data from the text receiver 102 and includes a privacy specification collector module 112 , a rule-based privacy evaluator module 114 , and a privacy determination module 116 .

[0079] The privacy specification collector module 112 periodically queries a database that stores relevant privacy laws (e.g., HIPAA forms) to retrieve updated privacy specifications and processes these specifications to create annotated rules. These rules are then transmitted to the rule-based privacy assessor module 114 via an internal communication path. The text receiver 102 converts the text entered by the user into a standardized format and then sends it to the rule-based privacy assessor module 114. The rule-based privacy assessor module 114 processes the input data by comparing the input data with the annotated specifications and rules using a rule matching algorithm. Once any violation is detected, the detection results are sent to the privacy judgment module 116. The privacy judgment module 116 compiles and formats the findings into a readable structured report to ensure that highly accurate privacy judgments are made.

[0080] The customized learning module 120 includes a context data processor module 122 , a fine-tuning BERT module 124 , a knowledge distillation module 125 , a cloud-based base module 126 , a reward module 127 , and an edge device micro module 128 .

[0081] The context data processor module 122 receives user data from the user's device (which can be a smartphone, tablet or other edge device) through the text receiver 102. The context data processor module 122 processes and interprets this data to understand the context information therein. The processed context information is then transmitted to the fine-tuning BERT module 124 through an internal data path. The fine-tuning BERT module 124 uses pre-trained models and algorithms to classify whether the context information constitutes a privacy violation. The classification results are then forwarded to the privacy judgment module 116 through a secure communication channel to update its privacy assessment record.

[0082] The knowledge distillation module 125 extracts essential knowledge from the LLMs to enhance the small BERT model in the fine-tuned BERT module 124. The knowledge distillation module 125 communicates with the cloud-based base module 126, where the cloud-based base module 126 acts as a base model for privacy violation learning and knowledge distillation. The reward module 127 interacts with the cloud-based base module 126 to receive data and update the cloud-based base module 126 based on the identified privacy violations. The learning process for this can involve supervised fine-tuning or involve human feedback reinforcement learning (RLHF). Once the cloud-based base module 126 is updated, the distilled knowledge is passed back to the knowledge distillation module 125.

[0083] Finally, the resulting information generated in the customized learning module 120 is fed to the edge device module 128, which processes the user data locally by integrating the context data processor module 122, the fine-tuned BERT module 124, the knowledge distillation module 125, the cloud-based base module 126, and the reward module 127 to generate a local privacy report. The edge device module 128 is configured to enable privacy judgments on the device and learn the user's privacy preferences, so there is no need for constant cloud interaction. This local processing ensures faster response times and enhanced privacy protection by minimizing data transmission to external servers.

[0084] The reasoning module 130 includes a context integration module 132 , an inference evaluator module 134 , a context learning module 136 , and a server-side reasoning module 138 .

[0085] The context integration module 132 receives data from the user's device via the text receiver 102 and is configured to integrate multi-turn conversation context to assess privacy. This involves aggregating user interaction sequences over time and then forming a comprehensive context. Next, the aggregated context is transmitted to the reasoning evaluator module 134 via an internal data path.

[0086] The reasoning evaluator module 134 receives the integrated context from the context integration module 132 and uses chain of thought (CoT) reasoning to evaluate whether the entire context contains private information. The evaluation of the reasoning evaluator module 134 involves breaking down the context into logical steps and evaluating whether each step violates privacy. The result of the reasoning process is then forwarded to the context learning module 136 through a secure communication channel.

[0087] The context learning module 136 interacts with the reasoning evaluator module 134 and uses a small amount of context learning to refine the privacy judgment. The refinement process of the context learning module 136 involves training the model of the reasoning evaluator module 134 with a small amount of examples / samples to improve its accuracy in identifying privacy violations. In one embodiment, the context learning module 136 and the reasoning evaluator module 134 can work together to provide a step-by-step reasoning template, which helps to build the evaluation process and improve the accuracy of the privacy assessment.

[0088] The server-side reasoning module 138 can be a server or centralized computer that can provide privacy judgments for cloud-based LLMs interactions. Specifically, the server-side reasoning module 138 integrates the result information from the context integration module 132, the reasoning evaluator module 134, and the context learning module 136 to aggregate and process these inputs. The reasoning performed by the server-side reasoning module 138 can provide comprehensive privacy judgments for interactions occurring in the cloud, and then generate reports on these findings.

[0089] The system 100 also includes a privacy interception module 140. The privacy interception module 140 is configured to actively monitor and intercept the transmission of user data when the rule-based privacy checklist module 110, the customized learning module 120, and the reasoning module 130 detect potential privacy violations (i.e., third-level detection). After receiving the input user data and after the third-level detection, the privacy interception module 140 analyzes the reports from the rule-based privacy checklist module 110, the customized learning module 120, the reasoning module 130, or a combination thereof. If the analysis indicates that the data involves sensitive or private information, the privacy interception module 140 intervenes by stopping the data transmission process. This pre-triggered action of the privacy interception module 140 prevents private data from being sent to external servers or other unintended recipients, thereby protecting the privacy of the user. In addition, the privacy interception module 140 can also generate alerts and log events for further review to ensure that appropriate measures are taken to address and correct privacy leaks. The privacy interception module 140 can enhance user data protection and comply with privacy regulations more effectively.

[0090] Privacy is protected by feeding the input user data into the AI ​​model for in-depth analysis. This ensures that sensitive data is handled securely, simplifies the system's operating procedures, and reduces computer power consumption during operation. Therefore, integrating the AI ​​model and the system can speed up calculations and maximize computing efficiency while maintaining strict privacy protection.

[0091] In summary, the proposed IPA framework provides a powerful multi-level approach to privacy judgment by integrating rule-based checking, customized learning, and high-level reasoning. The integrated system can effectively align with the privacy norms defined by laws and regulations by considering the roles of senders, receivers, and information subjects in information transmission. The technical solution of the present invention innovatively uses contextual integrity as the basis for privacy judgment, combined with a novel integration of different privacy protection methods, ensuring high accuracy and effectiveness in protecting user privacy in LLMs. Therefore, the IPA framework represents a major advancement in privacy protection in LLMs, enhancing its practicality and security.

[0092] The functional units and modules of the processor and method according to the embodiments disclosed herein can be implemented in the form of hardware or software. That is to say, the required processor can be implemented completely as a machine instruction or as a combination of machine instructions and hardware elements. Hardware elements include, but are not limited to, computing devices, computer processors, or electronic circuits, including but not limited to application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), microcontrollers, and other programmable logic devices configured or programmed according to the teachings of this disclosure. Those skilled in the art of software or electronics can easily prepare computer instructions or software codes running in computing devices, computer processors, or programmable logic devices based on the teachings of this disclosure.

[0093] The system may include computer storage media, transient and non-transitory memory devices having computer instructions or software codes stored therein, which may be used to program or configure a computing device, a computer processor, or an electronic circuit to perform any of the processes of the present invention. The storage media, transient and non-transitory memory devices may include, but are not limited to, floppy disks, optical disks, Blu-ray disks, DVDs, CD-ROMs, magneto-optical disks, ROMs, RAMs, flash memory devices, or any type of medium or device suitable for storing instructions, codes, and / or data.

[0094] The system may also be configured as a distributed computing environment and / or cloud computing environment, where all or part of the machine instructions are executed in a distributed manner by one or more processing devices interconnected by a communication network, such as an intranet, a wide area network (WAN), a local area network (LAN), the Internet, and other forms of data transmission media.

[0095] The above description of the present invention is provided for the purpose of illustration and description. It is not intended to be exhaustive or to limit the present invention to the precise form disclosed. For those skilled in the art, many modifications and variations will be apparent.

[0096] The embodiments were chosen and described in order to best explain the principles of the invention and its practical application, thereby enabling others skilled in the art to understand the invention for various embodiments and with various modifications as are suited to the particular use contemplated.

Claims

1. A system for performing large language model operations using an intelligent privacy assistant model, characterized in that: include: a text receiver for parsing and preprocessing input data from at least one user to convert the user input text into a machine-readable format; a rule-based privacy checklist module configured to query a database to retrieve privacy specifications and process said privacy specifications to create annotated rules, wherein said rule-based privacy checklist module processes said input data from said text receiver and compares it to the annotated specifications and rules, thereby compiling and formatting the comparison results into a readable structured report; a customized learning module, receiving the input data through the text receiver and configured to process and interpret the input data to understand its context information, wherein the customized learning module includes a fine-tuned BERT module for classifying whether the context information constitutes a privacy violation, wherein the classification result of the fine-tuned BERT module is processed locally to generate a local privacy report; as well as A reasoning module receives the input data through the text receiver and is configured to evaluate privacy based on server-side reasoning and integrated multi-round conversation context, wherein the reasoning module extracts the integrated context and uses chain of thoughts (CoT) reasoning to evaluate whether the entire context of the integrated context has privacy information, wherein the reasoning module provides a comprehensive privacy judgment on the interaction occurring in the cloud during the stage of evaluating the privacy information, and also generates a cloud report.

2. The system according to claim 1, characterized in that Also includes: a privacy interception module configured to actively monitor and intercept the transmission behavior of the input data when the rule-based privacy checklist module, the customized learning module, the reasoning module, or a combination thereof detects potential privacy violations in the input data; If the analysis of the privacy interception module indicates that the input data involves sensitive or private information, the privacy interception module intervenes in the process by stopping the data transmission.

3. The system according to claim 1, characterized in that The rule-based privacy checklist module also includes: a privacy regulations collector module configured to periodically query a database storing relevant privacy regulations and retrieve updated privacy regulations to create annotation rules; and A rule-based privacy assessor module, wherein annotation rules are transmitted from the privacy specification collector module to the rule-based privacy assessor module, and the rule-based privacy assessor module is configured to compare the input data with the annotation specifications and rules by using a rule matching algorithm to complete processing the input data.

4. The system according to claim 3, characterized in that The rule-based privacy checklist module also includes: The privacy determination module comprises: a privacy determination module, wherein the comparison result of the rule-based privacy assessor module is transmitted to the privacy determination module, and the privacy determination module is configured to compile and format the comparison result into the readable structured report.

5. The system according to claim 4, characterized in that The customized learning module also includes: A context data processor module is used to receive the input data and is configured to process and interpret the input data to understand the context information, wherein the context information processed by the context data processor module is transmitted to the fine-tuned BERT module, and the classification result of the fine-tuned BERT module is further forwarded to the privacy judgment module to update its privacy assessment record.

6. The system according to claim 5, characterized in that The customized learning module also includes: a knowledge distillation module configured to extract necessary knowledge from the LLMs database to enhance the small BERT model in the fine-tuning BERT module; and A cloud-based basic module communicates with the knowledge distillation module and is configured to serve as a basic model for privacy violation learning and knowledge distillation.

7. The system according to claim 6, characterized in that The customized learning module also includes: A reward module interacts with the cloud-based base module to receive data and update the cloud-based base module based on the identified privacy violations, wherein the reward module is also configured to perform a learning process involving supervised fine-tuning or involving human feedback reinforcement learning (RLHF), and once the cloud-based base module is updated by the reward module, the distilled knowledge is transmitted back to the knowledge distillation module.

8. The system according to claim 1, characterized in that The reasoning module also includes: a context integration module, receiving the input data and configured to integrate the multi-turn conversation contexts, including aggregating user interaction sequences that vary over time to form a comprehensive context to be extracted as the integrated context; and A reasoning evaluator module receives the integrated context from the context integration module and is configured to use CoT reasoning to evaluate whether the entire context contains privacy information, wherein the evaluation of the reasoning evaluator module involves decomposing the integrated context into logical steps and evaluating whether there is a privacy violation in each step.

9. The system according to claim 8, characterized in that The reasoning module also includes: A context learning module interacts with the inference evaluator module and is configured to utilize a small amount of context learning to refine the privacy judgment, wherein the refinement of the context learning module involves training at least one model of the inference evaluator module using a small amount of examples or samples.

10. The system according to claim 1, characterized in that The system runs in a smartphone, tablet or edge device.

11. A method for performing large language model operations using an intelligent privacy assistant model, characterized in that: include: Parsing and preprocessing input data from at least one user, via a text receiver, thereby converting the user input text into a machine-readable format; querying a database to retrieve privacy specifications and processing said privacy specifications to create annotated rules via a rule-based privacy checklist module, wherein said rule-based privacy checklist module processes said input data from said text receiver and compares it to the annotated specifications and rules, thereby compiling and formatting the comparison results into a readable structured report; Receiving the input data by the text receiver through the customized learning module; Processing and interpreting the input data to understand its context information through the customized learning module, wherein the customized learning module includes a fine-tuned BERT module for classifying whether the context information constitutes a privacy violation, wherein the classification result of the fine-tuned BERT module is processed locally to generate a local privacy report; Receiving the input data by the text receiver through the inference module; as well as The reasoning module is used to evaluate privacy by reasoning and integrating multiple rounds of conversation contexts based on the server side, wherein the reasoning module extracts the integrated context and utilizes chain of thoughts (CoT) reasoning to evaluate whether the entire context of the integrated context contains privacy information, wherein the reasoning module provides a comprehensive privacy judgment on the interactions occurring in the cloud during the stage of evaluating privacy information, and also generates a cloud report.

12. The method according to claim 11, characterized in that Also includes: When the rule-based privacy checklist module, the customized learning module, the reasoning module, or a combination thereof detects potential privacy violations in the input data, monitoring and intercepting the transmission behavior of the input data through a privacy interception module; If the analysis of the privacy interception module indicates that the input data involves sensitive or private information, the privacy interception module intervenes in the process by stopping the data transmission.

13. The method according to claim 11, characterized in that Also includes: Through the privacy regulation collector module, the database storing relevant privacy regulations is regularly queried and updated privacy regulations are retrieved to create annotation rules; The annotation rules are transmitted from the privacy specification collector module to a rule-based privacy assessor module, and the rule-based privacy assessor module is configured to compare the input data with the annotation specifications and rules by using a rule matching algorithm to complete processing of the input data.

14. The method according to claim 13, characterized in that The comparison result of the rule-based privacy assessor module is transmitted to the privacy judgment module, and the privacy judgment module compiles and formats the comparison result into the readable structured report.

15. The method according to claim 14, characterized in that Also includes: Receiving the input data via a context data processor module; as well as The input data is processed and interpreted by the context data processor module to understand the context information, wherein the context information processed by the context data processor module is transmitted to the fine-tuned BERT module, and the classification result of the fine-tuned BERT module is further forwarded to the privacy judgment module to update its privacy assessment record.

16. The method according to claim 15, characterized in that Also includes: Through the knowledge distillation module, necessary knowledge is extracted from the LLMs database to enhance the small BERT model in the fine-tuning BERT module, wherein the cloud-based base module communicating with the knowledge distillation module serves as a base model for privacy violation learning and knowledge distillation.

17. The method according to claim 16, characterized in that Also includes: Interacting with the cloud-based base module through a reward module to receive data and update the cloud-based base module according to the identified privacy violations, wherein the reward module is also configured to perform a learning process involving supervised fine-tuning or involving human feedback reinforcement learning (RLHF), and once the cloud-based base module is updated by the reward module, the distilled knowledge is transmitted back to the knowledge distillation module.

18. The method according to claim 11, characterized in that Also includes: Receiving the input data through a context integration module; integrating the multi-round conversation contexts by the context integration module, including aggregating user interaction sequences that vary over time to form a comprehensive context to be extracted as the integrated context; receiving, by an inference evaluator module, the integrated context from the context integration module; as well as The reasoning evaluator module uses CoT reasoning to evaluate whether the entire context contains privacy information, wherein the evaluation of the reasoning evaluator module involves decomposing the integrated context into logical steps and evaluating whether each step has privacy violation.

19. The method according to claim 18, characterized in that Also includes: interacting with the inference evaluator module via a context learning module; as well as Through the context learning module, a small amount of context learning is used to refine the privacy judgment, wherein the refinement of the context learning module involves training at least one model of the inference evaluator module using a small amount of examples or samples.

20. The method according to claim 11, characterized in that The method is operated and activated via a smartphone, tablet or edge device.

Images