102 results about "Content filtering" patented technology

The invention provides a large model dynamic protection method and system based on a zero-trust architecture. The method comprises the steps that a security proxy gateway receives an access request; authenticating an initiating main body of the access request, collecting context information and transmitting the context information to a strategy decision point; the strategy decision point calculates a trust score in real time based on a dynamic trust evaluation model and performs real-time evaluation in combination with an access control strategy to generate a dynamic authorization judgment result; if the access is allowed, forwarding the access request to a large language model server, and performing input security filtering; the large language model server generates response content and performs output security filtering; and returning the final response subjected to the output security filtering to the initiating main body through the security proxy gateway. According to the method, a multi-layer protection framework is constructed, a dynamic trust evaluation model is introduced, and a content filtering layer is deployed, so that continuous permission verification, risk adaptive control and full-link content security protection are realized, and the service security of a large model is effectively guaranteed.

The invention discloses a large-model sensitive large-content filtering method in a learning scene, and the method comprises the steps: extracting an image-side visual coding feature and a text-side semantic coding feature of a learning problem from the learning problem inputted by a user through employing a neural network model based on deep learning; and multi-modal feature fusion is further carried out on the characters and the pictures to form more comprehensive and accurate question joint representation, and the joint representation can fully mine the deep relation between the characters and the pictures, so that sensitive or improper contents in complex input are effectively identified, and automatic and intelligent filtering of the compound learning questions is realized. The problems that a traditional single-mode filtering scheme is narrow in coverage, low in detection accuracy, low in manual auditing efficiency and the like are solved, the recognition capacity of a platform for sensitive content is improved, and then a firm and reliable safety guarantee is provided for an intelligent education platform.

A genetic algorithm is implemented to generate prompts that evade content filters of generative artificial intelligence (AI) systems. The genetic algorithm applies grammar operations to mutate candidate prompts, communicates the candidate prompts to generative AI systems, and selects candidate prompts that successfully evade content filters according to corresponding responses. A disambiguation model that corrects grammar in prompts is tested on the selected prompts to determine if grammar is properly corrected. Once tested, the disambiguation model is deployed in an ensemble with a classifier that outputs verdicts for prompts with grammar corrected by the disambiguation model.

The invention discloses an intelligent customer opinion processing system based on multiple agents. A receiving module receives complaint information of a user and standardizes the complaint information. And the task distribution module realizes splitting of customer complaint tasks and intelligent agent adaptation for standardized customer complaint data through a multi-intelligent agent iterative learning framework. And the subtask execution module executes the received customer complaint task. And the feedback optimization module collects full-link indexes in real time, detects abnormal events by adopting a sliding time window, and optimizes the construction of a decision tree in the sub-task execution module. And the content filtering module is used for carrying out post-processing on output texts of the consultation agent and the risk identification agent by using a text filtering and optimizing framework based on an off-line large model. And the result storage module adopts a multi-level storage strategy to record a full-link operation log. According to the method, the problems of low efficiency, insufficient accuracy, weak dynamic adaptive capacity and the like of a traditional customer complaint processing method can be effectively solved.

Context-aware content filtering adapted for a knowledge domain is provided. In certain examples, a classification confidence score by a classifier indicates a level of confidence that a prompt from a user is associated with the knowledge domain. The classification confidence score is compared with a threshold. When the score is below a threshold, a violation notice is provided to the user without submitting the prompt to a generative artificial intelligence (GAI) model. When the classification confidence score is above the threshold, the prompt is further processed to determine, according to rules, whether the prompt should be submitted to the GAI model. In various examples, the rules are applied to contextual information, safety score information, and intent information derived from the prompt.

The invention belongs to the technical field of computer information security, particularly discloses a content security protection method and device based on semantic consistency, and aims to solve the problem that high-level cue word injection attacks are difficult to effectively recognize in the prior art. Carrying out real-time interception on dominant illegal texts through a content filtering module; quantizing generation rationality differences of prompt words between attack and normal language models by using a word vector confusion degree calculation module; evaluating the logic coherence of the sentence structure of the prompt word through a statement semantic consistency judgment module; integrating the multi-dimensional feature data and carrying out risk classification by adopting a machine learning algorithm; and routing the suspected attack request to a value fine tuning model for processing according to a judgment result. According to the technical scheme, high-precision recognition and response to complex cue word injection attacks are achieved, and the content safety protection capacity and compliance guarantee level of a large model in an interaction scene are remarkably improved.

The invention discloses a multi-mode private illegal content retrieval method and system based on semantic granularity perception, and belongs to the technical field of image-text multi-mode content retrieval. In order to solve the technical problem that private and illegal contents are difficult to effectively retrieve through explicit text description, the method comprises the following steps of: constructing an image identifier consisting of clustering, matrix decomposition, quantization and unique coding, and realizing generative learning of an image and the image identifier by utilizing a multi-modal large model; and constructing a soft label based on the similarity between the image and the text representation, and outputting a corresponding image identifier through text-driven generative retrieval learning. According to the method, the illegal image content with semantic matching can be retrieved from the private text description, and the method is suitable for scenes such as illegal content examination, risk identification and content filtering.

A dedicated interactive device designed for children comprises an embedded electronic module uniquely paired to a single child via a secure identifier. The device enables conversational and educational interactions, applying a multi-layer artificial intelligence filtering system to block sensitive content. Interaction data are encrypted and stored locally when offline, with secure synchronization to the cloud upon reconnection. A cloud-based parental dashboard allows management of one or multiple child profiles, language and voice preferences, review of time-stamped interaction logs, and configuration of usage controls. The system features privacy-by-design principles, including data minimization, strong encryption, authentication, and regulatory compliance. Optional components, such as cameras or displays, are integrated under the same safety and privacy framework. The platform unifies secure device-child pairing, robust AI filtering, comprehensive parental control, and enhanced protection of children's data in a single system.

Techniques for filtering out undesirable generative machine learned model (or LLM) output are discussed herein. A system may receive a subset of an LLM output. That is, the system may stream the LLM output to a user device by receiving one or more tokens from the LLM and outputting such token(s) to a user device. However, prior to outputting the token(s) to the user device, the system may determine whether the token(s) include undesirable content that is to be blocked. The system may use synchronous blocking components (e.g., blocks the undesirable token(s) before such token(s) get output to the user device) and / or asynchronous blocking components (e.g., blocks the undesirable token(s) after the token(s) have been output to the user device) to filter out undesirable content. The synchronous and / or asynchronous blocking components may be designed to block one or more undesirable topics such as hateful speech, profanity, bias, toxicity, factualness, etc.

The invention discloses a method and device for pushing recruitment platform content based on user fatigue, and the method comprises the steps: building a historical behavior baseline for the pushing content of different business types of a recruitment platform for a target user, and enabling the historical behavior baseline to be obtained through the rolling calculation of the historical behavior data of the target user based on a time decay algorithm; for any service type, calculating a comprehensive fatigue score of the target user to the push content of the service type based on the historical behavior baseline and the interactive behavior data of the target user to the push content of the service type; the comprehensive fatigue degree score is compared with a threshold value interval, a pushing strategy corresponding to the interval to which the comprehensive fatigue degree score belongs is executed according to the comparison result, and the pushing strategy comprises normal pushing, frequency reduction pushing, content filtering, pushing channel switching, content transformation and / or silence period entering. According to the application, accurate pushing of the content published by the recruitment platform can be realized, the waste of pushing resources is reduced, and the user satisfaction is improved.

The invention provides a SaaS-based multi-party system data pushing method and system and a storage medium, and the method comprises the steps: obtaining a differentiated pushing rule pre-configured by a target tenant, the differentiated pushing rule comprising a field selection rule for a target interface, a data filtering rule and a data format conversion rule; obtaining original data to be sent to a target tenant, and performing field screening on the original data based on a field selection rule to obtain first intermediate data; performing content filtering on the first intermediate data based on a data filtering rule to obtain second intermediate data; performing format conversion on the second intermediate data based on a data format conversion rule to obtain to-be-transmitted data meeting the customized format requirement of the target tenant; and sending the to-be-transmitted data to a receiving system corresponding to the target tenant. According to the invention, multi-party system data pushing can be carried out safely and efficiently at low cost.

The application relates to the technical field of artificial intelligence, in particular to a multi-modal human-computer interaction content filtering method, device and medium. The method comprises the following steps: collecting multi-modal interaction data; obtaining a current application scene, calculating single invalidity scores of various data, obtaining joint invalidity scores in combination with multi-modal weights corresponding to the scene, and judging whether the joint invalidity scores are lower than a preset threshold. If yes, scene-specific violation judgment weights are called, multi-modal violation features are extracted from the data, and current fusion features are obtained by combining the weights; the latest N (N>1) historical fusion features are obtained, the current fusion features are combined to generate modified fusion features, the modified fusion features are input into a pre-trained deep learning model, and a risk confidence is obtained. When the risk confidence exceeds a preset threshold, an interaction instruction transmission link is immediately cut off, and multi-modal data is encrypted and stored as evidence, so that accurate filtering and high-risk prevention and control of multi-modal interaction content are realized. The application can effectively improve the judgment accuracy.

The invention discloses a video analysis task unloading method and device based on mixed granularity content filtering under edge computing and a storage medium, and the method comprises the steps: 1, dividing each video frame image into a plurality of sub-blocks, extracting a candidate target region of each sub-block, dividing each sub-block of each frame of image into a target sparse block and a target dense block; step 2, for the target sparse blocks obtained in the step 1, compactly arranging all candidate target areas of each target sparse block on a plurality of virtual canvases with fixed sizes to form canvas blocks, and forming a block to be unloaded by each canvas block and a target dense block; step 3, generating an unloading scheduling strategy by using a deep reinforcement learning strategy network trained based on a near-end strategy optimization algorithm, and distributing the blocks to be unloaded to each edge node in parallel according to the unloading scheduling strategy; the equipment and the storage medium are used for implementing the method. According to the method, the problems of calculation redundancy and resource bottleneck of existing edge video analysis are solved.

According to the embodiments of the present disclosure, provided are a method and apparatus for content recommendation, and a device, a medium and a program product. The method comprises: using a plurality of content filtering strategies to determine a plurality of groups of candidate recommended content from a content library, wherein the plurality of content filtering strategies are respectively based on different content ranking criteria; using a trained machine learning model to determine a recommendation score of each piece of candidate recommended content in the plurality of groups of candidate recommended content with respect to a target user on the basis of user reference information of the target user and content reference information of the plurality of groups of candidate recommended content; and on the basis of the recommendation score corresponding to each piece of candidate recommended content in the plurality of groups of candidate recommended content, determining a group of target recommended content from among the plurality of groups of candidate recommended content, so as to provide the group of target recommended content to the target user. Thus, the accuracy and efficiency of content recommendation can be improved.

The invention relates to a vehicle sensitive word instruction rejection method and device.The method comprises the steps that text instruction data of a target user for a current vehicle are collected, keyword matching is conducted on the text instruction data through a sensitive word rule engine, and whether sensitive words exist in the text instruction data or not is judged; if the target sensitive word exists, generating a sensitive word rejection instruction to perform sensitive word instruction reminding on the target user; otherwise, inputting the text instruction data into the sensitive word rejection model, outputting a semantic analysis result to judge whether the text instruction data is a target invalid word or not, and outputting an invalid word rejection instruction to perform invalid instruction reminding on the target user under the condition that the text instruction data is the target invalid word. According to the method, the user input is subjected to content filtering and validity judgment so as to ensure the safety and processing efficiency of the subsequent service, and through keyword filtering and semantic analysis, the recognition service can quickly and accurately recognize the input which needs to be rejected, and resource waste is avoided.

Methods and a system for personalizing self-checkout terminals based on customer loyalty information includes receiving customer preferences during registration, generating themed personas with specific color schemes, graphics, and animations relevant to the customer's registered interests, and dynamically applying personalized interface themes when customers are identified at self-checkout terminals. The system maintains cached personas tagged with attributes like interests, religious preferences, and dietary restrictions, enabling sub-second personalization by matching customer preferences to pre-generated themes. The system avoids showing inappropriate content by filtering themes based on offensive content, collects user feedback on displayed themes through a rating mechanism, and employs time-delay factors to prevent theme repetition. New personas and corresponding themes are periodically generated to maintain customer engagement while keeping processing costs manageable.

To easily introduce a generation AI in a company and to perform optimum prompt engineering and filtering of illegal contents.SOLUTION: The system includes a user interface, a content filter, a prompt engineering infrastructure, an LLM engine, and a ConfigCenter. The user interface receives a user operation. The content filter filters content generated based on an operation received by the user interface. The prompt engineering infrastructure optimizes the prompts based on the content filtered by the content filter. The LLM engine generates an output based on the optimized prompt. ConfigCenter manages the configuration of the user interface, content filters, prompt engineering infrastructure, and LLM engine elements.SELECTED DRAWING: Figure 1

The application discloses a private protocol transmission security management system and method based on a one-way communication device, and relates to the technical field of network security. The method comprises the following steps: extracting the communication characteristics of a private protocol and performing normalization processing, and constructing a private protocol feature library; deploying a protocol analysis engine on the low-security side of a one-way optical shutter, performing deep detection on the out-domain data stream, identifying the private protocol, and extracting the current session protocol characteristics; matching the extracted characteristics with the feature library, and performing compliance verification and content filtering according to a security policy after successful matching; the verified data is encapsulated into a transmission format allowed to pass through the one-way optical shutter, and is sent to the high-security side through the one-way optical shutter; the received data is reorganized, and integrity and consistency verification is performed on the high-security side, and after the verification is passed, an audit log is recorded and forwarded to a target system; the protocol behavior and the session state are monitored in real time, and a security alarm is triggered and the transmission is blocked for unknown protocols or abnormal behaviors.

The invention discloses an anti-serial connection method and system for network bidding. The method comprises the following steps: implementing information isolation on bidding related parties to ensure that any party cannot know identities and accurate quantity information of other participants, and displaying details, a total demand scale and a current price of an object to a buyer, a seller and a bystander; in the bidding process, a new competitor is dynamically introduced; anonymization control and content filtering are carried out on communication between the buyer, the seller, the bystander and the seller; monitoring a collaborative quotation behavior and triggering disposal; and the key data block chain is subjected to evidence storage, and a visual report is generated for inspection after bidding is finished. The method can effectively prevent series connection and activate competition.

Computer-implemented methods for a topical and contextual content filter for large language model chatbots. Aspects include receiving a user query of a user from a chatbot application. Aspects also include receiving a previous chatbot application conversation associated with the user. Aspects further include determining that the user query of the user from the chatbot application is not a continuation of the previous chatbot application conversation associated with the user. Aspects also include identifying a short list of possible topics using the user query and a topic database. Aspects include determining that the user query of the user from the chatbot application is not associated with the short list of possible topics. Aspects further include transmitting a notification to the chatbot application comprising a rejection of the user query.

On a device with an internet protocol network connection, an application is provided that allows a user to watch internet protocol television on the device. While the user is watching a first live internet protocol television program via the application on the device, an accurate current list of second through Nth live internet protocol television programs that are available to be watched by the user is displayed. The device could include, for example, a gaming console, a media streamer, or the like.

The application discloses a content filtering method and system based on role portraits and a storage medium. The method comprises the following steps: a client establishes a first role portrait library for a user role, and sends the first role portrait library to a server; the server learns each received first role portrait library, and generates a second role portrait library for each role; when the client receives conversation information sent by the user role, it is judged whether the user role matches the first role portrait library; if yes, the first role portrait library is used to filter the content of the conversation information; otherwise, the client sends the conversation information to the server, the server matches a corresponding second role portrait library according to the user role, and uses the second role portrait library to filter the content of the conversation information. The application solves the problems of the prior art, such as rigid filtering rules, incapability of combining with context, and uneven load of the client and the server, and realizes intelligent, accurate and adaptive content filtering.

Systems and methods for dynamically generating event cards from message streams is provided. A system receives a message of a sequence of messages from a content source, determine, from the message, the game condition of the game, compare the game condition with a previous game condition and determines events that occurred in the game based on the message and the comparison. The system can determine, for each event of the one or more events, an event type and individual player contributions of one or more players and assign, for each event of the one or more events, to each player that contributed in the event, a points allocation corresponding to the player's contribution in the event based on the points assignment policy and the event type. The system can generate event cards and transmit the event cards to emote devices according to a content filtering policy.

This invention relates to a method, apparatus, device, and storage medium for generating pet companion voice based on voiceprint modeling. The method includes: acquiring the voiceprint features of a target user and constructing a personalized timbre model based on a quality score; acquiring pet behavior data collected by a pet companion terminal, identifying the pet's behavior categories and the intensity of the behavior; determining a basic tone template based on the behavior categories, and dynamically adjusting the speech parameters in the basic tone template according to a preset mapping relationship based on the behavior intensity to obtain target tone parameters; and synthesizing target companion voice based on the target tone parameters, the personalized timbre model, and target text matching the behavior category, and sending it to the pet companion terminal for playback. This invention achieves emotional and adaptive automated pet companionship, and ensures modeling accuracy and usage security through a recording quality scoring mechanism and a sensitive content filtering mechanism.

The invention relates to the technical field of data processing, and discloses a personalized interaction method, device and equipment based on user embedding and a medium, which can be applied to the financial and medical fields, and the method comprises the following steps: carrying out embedded generation processing on a user historical data set to obtain a user embedded vector set; performing joint construction processing, and performing personalized parameter regulation and control processing in a preset large language model to obtain an initial personalized interaction result set; and carrying out content filtering processing on the initial personalized interaction result set, and then carrying out joint modeling updating processing on the initial personalized interaction result set and the task description text to obtain a personalized interaction result. In the invention, aiming at the problem of low business context accuracy caused by the fact that multiple sets of model parameters need to be maintained according to classification in the existing personalized intelligent interaction, the user embedding vector set can be adjusted through the target personalized interaction result set obtained by calculation, so that the multiple sets of model parameters do not need to be maintained according to classification, and the user experience is improved. And the accuracy of the business context is improved.

Techniques for filtering out undesirable generative machine learned model (or LLM) output are discussed herein. A system may receive a subset of an LLM output. That is, the system may stream the LLM output to a user device by receiving one or more tokens from the LLM and outputting such token(s) to a user device. However, prior to outputting the token(s) to the user device, the system may determine whether the token(s) include undesirable content that is to be blocked. The system may use synchronous blocking components (e.g., blocks the undesirable token(s) before such token(s) get output to the user device) and / or asynchronous blocking components (e.g., blocks the undesirable token(s) after the token(s) have been output to the user device) to filter out undesirable content. The synchronous and / or asynchronous blocking components may be designed to block one or more undesirable topics such as hateful speech, profanity, bias, toxicity, factualness, etc.

The invention is applicable to the technical field of artificial intelligence, and provides an intelligent toy control method based on semantic risk detection, which comprises the following steps: acquiring voice data of a target user, and performing voice-to-text processing on the voice data to obtain corresponding target text data; performing semantic risk analysis on the target text data through a preset semantic risk detection model, and outputting a semantic risk analysis result corresponding to the target text data and a corresponding risk level tag; and based on the semantic risk analysis result and the corresponding risk level label, determining a target safety response strategy in a preset safety response strategy library, and controlling the intelligent toy to execute the target safety response strategy. The problems that an existing method depends on a universal semantic recognition and content filtering algorithm, under the condition that language expression is not standard, semantic misjudgment or missed judgment is prone to occurring, a real-time detection and intervention mechanism for voice behaviors is lacked, and potential risk behaviors cannot be actively recognized in the interaction process are solved.

The application discloses an optimization method and system for realizing software function expansion based on DLL and a computing device. The DLL module is integrated in target software of a client in a hot plug manner. The method comprises the following steps: in response to a request of inputting external data through a data input interface of the target software, capturing the external data through a first interceptor; verifying the external data, obtaining a verification result and returning the verification result to the target software, so that the target software processes the external data according to the verification result; in response to a request of outputting to-be-output content through a data output interface of the target software, capturing the to-be-output content through a second interceptor; performing content filtering and security detection on the to-be-output content, obtaining target output content and returning the target output content to the target software, so that the target output content is displayed on the client. The application can flexibly and quickly develop and deploy plug-ins without modifying software source codes, and then realize flexible expansion and rapid iteration of software functions through the plug-ins.

The application discloses a kind of multi-hop question and answer model construction methods based on table-text hybrid data, the model constructed by text content filtering module, table row content screening module and fusion data extraction question and answer module are composed.Text content filtering module is used to screen the text sentence content related to question;Table row content screening module is used to construct table row-text fusion data and further screen out the content related to question;Fusion data extraction question and answer module is used to extract answer in table row content-text fusion data.The model constructed by the application is screened by content, greatly reduces redundant information, makes text content effectively supplement table content, accelerates model training speed, improves the accuracy of table-text hybrid data question and answer.

The invention discloses a collaborative robot task optimization method based on multi-agent selective knowledge sharing, and belongs to the technical field of multi-agent reinforcement learning and industrial robot intelligent control. According to the method, for tasks such as splicing and carrying of multiple cooperative mechanical arms in an industrial manufacturing scene, only a success path and an effective adjustment strategy after failure are shared through a three-stage mechanism of knowledge source evaluation, content screening and personalized fusion; and correlation screening and weighted fusion are performed on the knowledge according to the task context and the local experience of the robot, so that blind simulation is avoided, and the personalized skills of the mechanical arms are reserved.