AI-Based Non-Intrusive Real-Time Context-Aware Advertising System and Control Method Thereof

Abstract

The present invention relates to a non-invasive (Zero-SDK) artificial intelligence advertising system that automatically determines an appropriate section for displaying advertisements and displays conditioned generative advertisements by analyzing multiple heterogeneous signals observable from the outside in a time-series manner, without modifying the internal code of real-time responsive interfaces such as game applications or conversational artificial intelligence (LLM; Large Language Model). This system consists of a data acquisition unit, a multimodal time series analysis unit, a generative artificial intelligence advertising engine, a display control unit, and a security module. The data acquisition unit time-series analyzes non-invasive external signals such as frame buffers, system load, network latency, and audio spectrum to determine game loading / transition sections or LLM response waiting sections (Cognitive Gap) as suitable sections for ad display. Generative AI ad engines use Diffusion or Transformer-based models to generate context-conditioned ads based on game worldviews, conversation topics, emotional tone, UI style, etc., and The display control unit non-invasively displays advertisements without modifying application code through official interfaces of the operating system and graphics API. The security module prevents unauthorized leakage by performing TEE (Trusted Execution Environment) or TPM-based encryption to decrypt and display advertising assets, and then immediately discarding them. The feedback module periodically optimizes ad quality through reinforcement learning. The present invention enables the implementation of advertising functions without SDK installation or engine modification, realizes context-adaptive advertising in various interface environments such as games, LLM, and streaming, and simultaneously overcomes the limitations of existing in-game and AI advertising technologies by providing a reliable advertising delivery structure based on hardware security.

Description

Technology Field

[0001] The present invention relates to a non-invasive advertising system and a control method thereof that uses artificial intelligence (AI) to automatically recognize cognitive gaps between user input and system responses occurring in real-time responsive interface environments such as games, conversational AI (LLM; Large Language Model), voice assistants, and streaming services, and displays advertising content without modifying the internal code of an application or inserting an SDK.

[0002] More specifically, it determines the appropriate section for ad display in real time by analyzing multiple modality data, such as screen rendering signals, system resource load (CPU, GPU, memory), network latency, and audio spectrum, in a time-series manner, and

[0003] Based on this, it relates to a technology in which a generative AI advertising engine automatically generates and displays advertising content that is naturally integrated into contexts such as game worldviews, conversation topics, and streaming scenes.

[0004] Here, 'SDK (Software Development Kit)' refers to development libraries, plugins, or modules that are directly integrated into the internal code of a specific game or application program to operate. The present invention is characterized by an SDK-free, non-invasive structure that controls the display of advertisements without involving access to the target program's internal memory or code injection, by utilizing only an external observation interface or screen synthesis interface provided by the operating system or graphical application programming interface (API) as officially documented in publicly available technical documents, without integrating such an SDK into the program.

[0005] The present invention can simultaneously secure high immersion and profitability without compromising the user experience (UX) by utilizing sections where user attention is maintained but delays occur due to system processing, such as large-scale language model-based conversational models, LLM (Large Language Model), response waiting sections, or game loading sections, as ad exposure points. Background Technology

[0006] The existing in-game advertising system was implemented with an SDK-based structure, requiring the insertion of advertising modules into the game code, which resulted in the following limitations.

[0007] (1) Technical instability

[0008] It has been reported that various errors, such as frame drops or crashes during loading, occur due to conflicts between the game engine's rendering loop and the network or graphics resources of the ad SDK. This is a common issue that can occur in ad SDK integration environments and may be affected by platform and engine configurations.

[0009] (2) Increase in operating costs

[0010] Since SDK integration requires quality verification by engine and platform, separate development costs are incurred for each individual game content. According to industry quotation reports and developer interviews, costs generally range from thousands to tens of thousands of dollars, which poses a burden particularly for small and medium-sized development companies.

[0011] (3) Problem of underutilizing loading sections

[0012] Loading sections where user input is impossible are suitable for ad exposure because they momentarily lower user tension without interrupting concentration; however, their actual utilization is difficult due to the limitations of existing SDK-based structures. This problem exists not only in games but also in large-scale language model-based conversational services (LLM).

[0013] For example, user attention is maintained even during response waiting periods commonly encountered when using LLM, but these periods are not utilized for displaying advertisements. Therefore, there is a growing need for technology that can detect loading or response waiting periods without modifying program code and display advertisements non-invasively. The problem to be solved

[0014] This invention aims to resolve the technical instability, increased operating costs, and the underutilization of advertisements during loading and response waiting periods caused by the structural limitations of existing SDK-based advertising systems. In particular, it aims to provide a completely non-invasive ad control technology that can stably display advertisements without modifying the internal structure of the program in real-time interface environments based on games and Large Language Models (LLM).

[0015] The specific objectives of the present invention are as follows.

[0016] (1) SDK-Free non-invasive structure realization

[0017] It enables ad display solely through the official external observation interfaces of operating system and graphics APIs, without modifying game code or engines, inserting SDKs, or hooking memory. This eliminates integration costs across engines and platforms and avoids security conflicts with DRM, anti-cheat systems, and the like.

[0018] (2) Context-based ad display appropriate segment detection

[0019] Through multimodal time-series analysis, it accurately detects game loading times or LLM response waiting periods to automatically determine ad display timing. This allows it to utilize waiting times in cognitive gaps as ad inventory without compromising the user experience.

[0020] (3) Provides advertisements that match the interface environment

[0021] By considering visual and contextual elements of screen color, lighting, object placement, or UI style, advertising images, animations, text, etc. are provided to naturally blend with the worldview or conversational environment.

[0022] (4) Delay-free ad display

[0023] By executing a predictive trigger request prior to the ad display time, it enables ads to be displayed without delay even in a real-time interface environment.

[0024] (5) Protection of advertising assets and prevention of unauthorized leakage

[0025] Advertising assets are decrypted, displayed, and discarded only within a secure execution environment, fundamentally preventing unauthorized copying, alteration, and external leakage.

[0026] (6) Stable ad display even in exceptional situations

[0027] In the event of generation delays, network failures, or prediction failures, ad assets are automatically replaced with pre-approved assets to ensure stable operation without service interruption.

[0028] (7) Operational Closed-Loop Based Performance Optimization

[0029] Ad impression and user response data are collected in a de-identified form to enable continuous optimization of ad quality, display conditions, and generation policies.

[0031] In summary, the present invention aims to fundamentally solve the problems of SDK-based advertising technology by providing a fully non-invasive advertising system that combines external observation-based AI judgment, conditional ad delivery, security display, and closed-loop optimization. means of solving the problem

[0032] The present invention includes the following configuration to solve the above problem.

[0033] (1) Data acquisition unit (100)

[0034] External signals such as screen rendering, frame rate, CPU / GPU usage, thread status, network packets, audio spectrograms, token generation rate, request length, and response delay are non-invasively collected during the game execution process or the response processing process of the conversational artificial intelligence (LLM), and the acquired signals are sorted into the form of a time-series data stream and transmitted to the multimodal time-series analysis unit. In particular, the data acquisition unit (100) operates only through the official interface provided by the operating system and graphics API (OpenGL, Vulkan, Metal, DirectX, etc.) without accessing game code or internal APIs.

[0035] (2) Multimodal time series analysis unit (110)

[0036] A time series signal received from a data acquisition unit (100) is input into a time series analysis model such as LSTM, Temporal CNN, or HMM to recognize a specific state of the game or LLM (loading, stage entry, pause, waiting for response, etc.). The analysis unit recognizes temporal change patterns and detects potential ad exposure intervals with high reliability by observing a “continuous interval of T seconds (T-5 seconds to T+0 seconds)” rather than static analysis of a single frame. At this time, the analysis window (T) can be dynamically adjusted by considering the average duration of the loading interval during the game (8.6 to 15 seconds) or the LLM response delay (1 to 5 seconds).

[0037] (3) Context-aware ad control unit (120)

[0038] Based on the signal transmitted from the multimodal time series analysis unit (110), the AI ​​model determines a loading section or an ad-possible section by comparing it with a pattern that has been repeatedly learned in advance. The control unit (120) can detect an ad-appropriate section with high reliability using only screen transitions, loading bar progress, UI changes, background sound changes, or LLM response delay signals. The context-aware ad control unit (120) preemptively performs an ad creation request (Predictive Trigger) a certain amount of time (1 to 2 seconds) before the detected loading start time or LLM response waiting start time, and displays the ad received from the generative AI ad engine (130) or the ad asset management unit in an overlay manner. Additionally, the control unit can pre-create and store ad content optimized for a specific genre or section in accordance with prior agreement with the advertiser or platform operator. At this time, the pre-created asset is produced in advance through the same conditioning algorithm of the generative AI ad engine (130), and is decoded and displayed through the same path as the real-time creation result. If the reliability of real-time generated results falls below the standard or generation delays occur, a balance between ad quality, display stability, and server load can be achieved by prioritizing and immediately displaying pre-generated assets.

[0039] (4) Generative AI advertising engine (130)

[0040] By extracting visual elements of the loading screen (color tone, lighting, object placement, etc.) and game character and environment data into feature vectors, advertising assets that naturally blend into the game worldview are automatically generated using GAN, Diffusion, Transformer-based, or hybrid generative models. For example, highly immersive advertising presentations can be achieved by inserting scenes where game characters directly use advertised products on the loading screen, or by displaying brand logos in the form of reflected light that match the loading background color tone. Furthermore, in LLM response environments, instead of real-time generation, at least one short video ad, animated ad, static image ad, or text ad matching the conversation topic category can be automatically selected and displayed from the advertiser's pre-registered asset library. In this way, LLM response waiting periods (e.g., "Preparing response," "Searching") can be utilized as ad exposure intervals, thereby extending user waiting time into non-invasive ad inventory.

[0041] (5) Security module (140)

[0042] Advertising content is stored in an encrypted state, decrypted within the Trusted Execution Environment (TEE) only at the time of display, and immediately discarded, while the decryption key is never exposed externally. Additionally, invisible watermarks are embedded in the frequency domain of advertising assets during display, enabling tracking in the event of unauthorized copying or leakage. This fundamentally prevents the tampering or external leakage of advertising assets.

[0043] (6) Feedback and monitoring unit (150)

[0044] User response data, such as ad impressions, clicks, and viewing time, is collected and analyzed to periodically improve the performance of the multimodal time-series analysis unit and the generative AI advertising engine. Collected data is processed through anonymization and pseudonymization, and personal information is not transmitted externally.

[0045] (7) IP rights enforcement module (160)

[0046] It supports the tracking and execution of unauthorized reproduction and modified advertisements by recording ad display logs, watermark signatures, display history, etc., in a central database or permissioned ledger.

[0047] (8) Operations Management Department (170)

[0048] It manages server status, model version, ad exposure statistics, security events, Fallback, cloud path settings, etc. in an integrated manner. The operations management department (170) monitors the status of system components in real time and performs recovery procedures immediately when an abnormal state is detected. Effects of the invention

[0049] According to the present invention, the following technical and industrial effects can be achieved.

[0050] (1) Reduction in integration costs due to SDK-free non-invasive structure

[0051] By adopting an SDK-free, completely non-invasive structure, the present invention enables the direct application of an advertising system to various platforms (mobile, PC, console, etc.) without modifying game code or inserting SDK modules. Accordingly, it is possible to significantly reduce the costs associated with additional development personnel and technical verification, such as publisher cooperation, game engine compatibility verification, and quality testing, which were required in existing SDK integration methods.

[0052] It is generally known in the game development industry that SDK integration requires significant cost and time, and there have been reports of small and medium-sized game companies, in particular, abandoning ad monetization due to this burden.

[0053] On the other hand, the non-invasive structure of the present invention overcomes the technical limitations of each platform and engine and can be applied immediately without additional development work, thereby enabling cost-effective advertising monetization in a wide range of service environments.

[0054] Furthermore, the same structure is applicable to Large-Scale Language Model (LLM)-based services, allowing ad timing to be determined solely through UI state and response latency without accessing service code or internal APIs, thereby providing AI service providers with a new revenue stream.

[0055] (2) Expanding Ad Inventory

[0056] The present invention can convert game loading periods (average 8.6 to 15 seconds) or LLM response waiting periods (2 to 10 seconds), which could not be utilized with existing SDK methods, into non-invasive ad inventory.

[0057] Furthermore, AI-based multimodal time-series analysis automatically determines loading start and end timings, significantly improving the accuracy of ad impression timing control. This leads to increased ad inventory per session and total ad revenue, and enables a stable business structure by allowing parallel operation with paid subscription (non-ad) models.

[0058] (3) Improvement of ad quality by generative AI

[0059] The generative AI ad engine generates and displays ads while maintaining visual and contextual consistency by inferring the conversational mood—such as background colors, lighting, and object placement in game environments, and conversation topics, UI styles, and user tone in LLM environments—even without the need for minimal topic tags, section-based information, or text provided by the platform (e.g., calm ads for long questions and answers, and bright-toned ads for short, fast Q&A).

[0060] Accordingly, the following improvements in advertising performance are expected.

[0061] - Minimize loss of immersion

[0062] - Increase in Click-Through Rate (CTR) and Conversion Rate

[0063] - Increase in brand recall

[0064] - Reduced aversion to advertisements

[0065] Ad display methods are implemented using fade-in / fade-out, slides, and vector animations, naturally blending with the user experience.

[0066] (4) Strengthening protection of advertising assets

[0067] Based on a TEE-based security structure, advertising assets are accessible only at the time of decryption and are discarded immediately after display. Additionally, an invisible watermark is inserted during the display process, enabling tracking of leaks and tampering. This significantly enhances security against external leakage and unauthorized use.

[0068] (5) Stable Fallback Operation

[0069] Even in situations such as network latency, server failure, or generation failure, the non-generated ad mode is automatically switched, and pre-approved ad assets are displayed immediately. This ensures system stability and the continuity of ad revenue.

[0070] (6) Pre-learned generality

[0071] By generalizing loading patterns by game genre through pre-trained AI models, deployment time for new games can be reduced, and rapid application is possible in LLM and real-time interface environments based on UI latency patterns.

[0072] (7) Industrial scalability

[0073] The present invention can be extended to the following fields.

[0074] - Mobile, PC, Console Games

[0075] - Cloud gaming, game streaming (GeForce NOW, Xbox Cloud, etc.)

[0076] - Conversational AI services (GPT, Gemini, Copilot, etc.)

[0077] - Buffering section of video streaming

[0078] - Scene transition sections in AR / VR and the Metaverse

[0079] - Cloud application loading screen

[0080] Accordingly, the present invention contributes to the establishment of a new advertising ecosystem in which advertisers, service providers, and platform operators can jointly generate revenue, and can be utilized as a core technology for the next-generation real-time interface advertising market. Brief explanation of the drawing

[0081] An embodiment of the artificial intelligence-based non-invasive advertising system according to the present invention is briefly described below with reference to the attached drawings. Figure 1 is an overall system configuration diagram of the present invention. A three-layer structure consisting of a game terminal (client), an edge server, and a central server is illustrated. Each layer is composed of a data acquisition unit (100), a multimodal time series analysis unit (110), a context-aware ad control unit (120), a generative artificial intelligence ad engine (130), a security module (140), a feedback and monitoring unit (150), an IP rights enforcement module (160), and an operation management unit (170). FIG. 2 illustrates an external observational input structure of a data acquisition unit (100). Multiple modality input signals, such as screen rendering, frame rate, GPU and CPU load, network packets, and audio spectrograms, are non-invasively collected and arranged into a time-series data stream. The acquisition unit (100) uses only the official interfaces of the operating system and graphics API without accessing game code or internal memory. FIG. 3 illustrates the procedure for determining the appropriate advertising interval of a multimodal time series analysis unit (110). The collected time series data undergoes preprocessing, feature extraction, and time series fusion steps to determine the appropriate advertising intervals, such as loading, conversion, and pause, in real time. FIG. 4 illustrates a conditional generation structure of a generative AI advertising engine (130). The color, lighting, camera angle, behavioral pattern, etc. of a scene are converted into a condition vector and used as input to a generative AI model. Quality criteria are applied according to automatic evaluation based on SSIM and LPIPS. FIG. 5 illustrates the operation flow of the context-aware ad control unit (120) and the security module (140). A generative ad is called at the predicted ad exposure time, and the ad asset is decrypted and displayed within a trusted execution environment (TEE) or an equivalent security area, after which it is immediately discarded. An invisible watermark is inserted during display. Specific details for implementing the invention

[0082] 8.1 Overall System Overview

[0083] The present invention relates to an AI-based advertising system that automatically recognizes the inactive interval, or cognitive gap, between user input and system response occurring in real-time responsive interface environments such as games, conversational AI LLM, streaming services, augmented reality (AR), and virtual reality (VR), and displays advertisements in a non-invasive manner without modifying the internal code or engine of the application.

[0084] Since this system operates using only the external observation and synthesis interfaces officially provided by the operating system or graphical application programming interface, it is possible to display advertisements in a consistent manner across various platform environments, such as games, LLM, and streaming, without code modification or engine access.

[0085] The system can operate on at least one of a client terminal, an edge server, and a central server, and can dynamically distribute generative ad processing functions depending on network conditions or system load. The client terminal performs display processing functions based on a trusted execution environment for rapid responsiveness and security, the edge server minimizes latency by distributing generation processing and ad selection functions, and the central server can perform learning and policy management functions.

[0086] The system of the present invention is composed of the following main modules.

[0087] 1 Data Acquisition Unit 100

[0088] 2 Multimodal Time Series Analysis Unit 110

[0089] 3 Context-aware ad controller 120

[0090] 4 Conditioning-based Generative AI Ad Engine 130

[0091] 5 Security Module 140

[0092] 6 Feedback and Monitoring Department 150 Select

[0093] 7 IP Rights Enforcement Module 160

[0094] 8 Operations Management and Distributed Deployment Division 170

[0095] The LLM response waiting period is in the range of 2 to 10 seconds on average, and the game loading period is longer, averaging 10 to 20 seconds, and an inactive state of up to 30 seconds or more may occur.

[0096] The present invention automatically reflects these environmental differences to recognize cognitive gaps in each situation as suitable intervals for ad display, and displays ads in a non-invasive manner without platform approval. Accordingly, the present invention provides a real-time closed-loop structure that simultaneously secures ad profitability and security without interfering with the user experience, and fundamentally resolves the issues of security conflicts, DRM and anti-cheat incompatibility, platform approval, and cost burden associated with SDK-based ad systems.

[0098] 8.2 Non-invasive Structure and Operating Principle

[0099] The core component of the present invention is an SDK-free, non-invasive structure. This means that for ad display, state information is received and ad display is controlled through an external observation interface or screen synthesis interface officially documented and provided by the operating system or graphical application programming interface (API), without directly accessing the source code, executable binaries, or internal memory within the application.

[0100] The present invention collects signals that do not contain personal information, such as CPU / GPU usage, frame rate of change, memory occupancy, network latency, and audio spectrum, through an external observation interface, and detects suitable sections for displaying advertisements.

[0101] This structure has the following advantages.

[0102] First, commercial distribution is possible without conflicts with anti-cheat DRM policies.

[0103] Second, universal applicability regardless of game updates and engine changes

[0104] Third, regulatory compliance is possible without accessing personal information or internal data.

[0105] Fourth, since original text is not collected even in an LLM environment, the protection of personal information is guaranteed, and

[0106] The flow of the non-invasive structure is as follows.

[0107] a. Collect system signals from the data acquisition unit (100).

[0108] b. Determination of the appropriate interval for advertising display in the multimodal time series analysis unit (110)

[0109] c. A creation request or a pre-asset call from the context-aware ad control unit (120).

[0110] D. Conditional generation or pre-asset decoding in the generative artificial intelligence advertising engine (130)

[0111] e. The security module (140) performs display within the trusted execution environment and immediate disposal.

[0112] The above process guarantees stability, reliability, and compatibility through an independent operation method isolated at the application, memory, and logic levels. Therefore, the present invention resolves the structural limitations of SDK-based advertising technology and realizes a universal artificial intelligence advertising technology applicable to any platform without internal approval.

[0114] 8.3 Multimodal Time Series Analysis Terms related to signal fusion methods and judgment errors used in this sub-specification are defined as follows: “Early fusion” refers to a method of combining multiple modality signals prior to feature extraction. “Late fusion” refers to a method of combining features extracted individually from each modality. “Hybrid fusion” refers to a form combining the above-mentioned early fusion and late fusion methods. Additionally, “false positive” refers to a case where an unsuitable section of an advertisement display is incorrectly judged as a suitable section, and

[0115] “False negative” refers to a case where the section suitable for displaying an advertisement is not determined, even though it is a suitable section. The multimodal time series analysis unit (110) of the present invention is a component that collects external observation signals provided by an operating system or a graphics API as time series data and automatically determines the section suitable for displaying an advertisement by analyzing time-modal reciprocal anomaly patterns. Since the analysis process is performed without access to the internal code, engine-level logic, or memory of the application program, the timing of the advertisement display can be detected with high reliability while maintaining a non-invasive SDK-Free structure.

[0116] (1) Signal acquisition and normalization

[0117] The collected signal consists of four modalities: time (V), system (S), network (N), and sound (A), and is sampled at intervals of 100 to 200 milliseconds and normalized. This is defined as the time series tensor T_t.

[0118] T_t = [V_t, S_t, N_t, A_t]

[0119] Here, T_t represents a form in which d-dimensional features are aligned along the time axis t. Data collection is performed non-invasively through performance monitoring APIs officially provided by the operating system.

[0120] (2) Definition and Quantitative Conditions of Co-occurring Anomaly Patterns In this specification, “co-occurring anomaly pattern” refers to a state in which predefined anomaly conditions, such as an increase in variance, a decrease in correlation coefficient, or a sudden change in the rate of change of time series, are simultaneously satisfied for two or more different state signals within a certain time interval.

[0121] In the present invention, an advertisement display candidate interval is defined as a case where two or more of the following conditions are simultaneously satisfied in multiple modalities during a certain time interval Δt (1 second to 3 seconds).

[0122] a. Condition for increasing variance

[0123] σ^2_i(t) / σ^2_i(t - Δt) > τ_var

[0124] Here, τ_var is set to a default value of 1.5, but can be adjusted in the range of 1.3 to 1.7 depending on the environment.

[0125] b. Conditions for reducing cross-correlation

[0126] ρ(i,j,t) < τ_corr

[0127] Here, τ_corr has a default value of 0.3 and can be adjusted in the range of 0.25 to 0.4.

[0128] c. Condition for stabilization of the rate of change of time

[0129] abs(dV / dt) < τ_dv

[0130] Here, abs(dV / dt) represents the average pixel value difference between time point (t) and (t - Δt_sample) or the L2 Norm change rate of the feature vector, and τ_dv is set to 0.1 by default.

[0131] The above threshold value is an exemplary standard for understanding the invention and does not limit the scope of the invention.

[0132] The above judgment conditions are based on statistical fusion and are a key element in preventing false positives and omissions in the appropriate advertising display ranges.

[0133] (3) Cross-modal attention-based early fusion

[0134] For each modality, Query, Key, and Value vectors are defined, and the importance between modals is calculated as an attention matrix with the following mathematical formula.

[0135] A_t(i,j) = Softmax((Q_t(i) * K_t(j)) / sqrt(d_k))

[0136] The fusion feature C_t is calculated as follows.

[0137] C_t(i) = Σ_j A_t(i,j) * V_t(j)

[0138] If the acoustic modality is silent for an extended period or contaminated by external noise, it is set to dynamically reduce its weight.

[0139] W_A_new = W_A * exp( - NoiseScore )

[0140] This early fusion structure improves upon the limitations of single-signal-based methods and enhances detection stability.

[0141] (4) Prediction of the probability of ad segment duration

[0142] The fusion feature C_t is input into a time series forecasting model (e.g., LSTM or Transformer), and the probability of the ad display fitting interval persistence is estimated as follows.

[0143] P_gap(t + T) = σ(W * C_t + b)

[0144] Here, when P_gap is 0.8 or higher, it is confirmed as an ad display suitable interval.

[0145] This step enables a preemptive trigger to the generative ad engine, minimizing ad generation delays.

[0146] (5) Effect

[0147] The multimodal time series analysis unit of the present invention provides the following effects.

[0148] a. Highly reliable detection of ad display sections while maintaining an SDK-free, non-invasive structure

[0149] b. Time-modal fusion-based detection significantly reduces false positives and omissions compared to single-signal based detection

[0150] c. Real-time applicable in various interface environments (games, LLM, streaming, etc.)

[0151] d. Low retraining burden results in excellent generalization performance even in new environments.

[0152] As such, this configuration reliably determines the appropriate section for ad display without accessing the internal structure of the application, making it highly applicable to industry as a core component of non-invasive advertising systems.

[0154] 8.4 Context-Aware Ad Controller

[0155] The advertising control unit receives the output value of the multimodal time series analysis unit in real time and

[0156] Determines the timing and conditions for ad display and controls the generation request.

[0157] (1) Predictive trigger-based control

[0158] An ad generation request is preemptively executed 0.5 to 3 seconds prior to the predicted time point. Prediction reliability is calculated by reflecting the prediction probability (P_gap) of the multimodal analysis unit and historical data; in this case, a value that quantitatively expresses the reliability of the prediction result or generation result can be defined as the confidence score. The confidence score can be calculated by combining one or more of the correlation between multiple status signals, the duration of joint anomaly patterns, the stability of statistical analysis indicators, and the output distribution characteristics of the prediction model. The confidence score can be used as a control criterion for determining whether the ad display time point is confirmed, whether the preemptive trigger is activated, and for selecting one of the real-time generation mode, pre-generated asset playback mode, or non-generated alternative mode. The lower the confidence, the shorter the preemptive request time is adjusted.

[0159] (2) Judgment of suitability of the display

[0160] The fit is calculated as a time-weighted integrated value that combines the stability indicators of multiple modalities based on exponential decay weights. If the fit satisfies the threshold (τ_fit, default value 0.75), ad display is approved.

[0161] (3) Creation request and parameter passing

[0162] Upon approval of display suitability, a generation request is sent to the generative ad engine as an asynchronous signal. The request parameters include visual feature vectors, environment state information, and advertiser constraints.

[0163] (4) Non-generating alternative mode control

[0164] If the generated result does not arrive within a certain time, if the quality reliability falls below the standard, or if there is prior agreement with the advertiser, it automatically or manually switches to a non-generated alternative mode. In this case, among the pre-approved assets, the asset requested by the advertiser or the asset with the highest cosine similarity to the HSV transformed value of the scene is selected.

[0165] (5) Utilization of minimum platform information

[0166] The control unit utilizes the minimum state information (RTT, token generation speed, etc.) officially provided by the platform as input and is used only within the scope of privacy non-infringement. This maintains a non-invasive structure without modifying the platform.

[0167] (6) Learning and Optimization

[0168] Ad display results automatically optimize ad display timing by collecting anonymized quality metrics and user response logs and reflecting them in periodic model updates.

[0169] (7) Management of pre-generated advertising assets

[0170] The Ad Control Unit prioritizes and immediately displays pre-generated assets in cases where generation delays are anticipated, reliability is compromised, or there is a prior request and consultation with the advertiser. Pre-generated assets are processed through the same secure path as real-time generation.

[0171] (8) Effect

[0172] This control unit can reliably display high-quality advertisements through preemptive-based control and dynamic mode switching. Furthermore, it simultaneously ensures platform independence and ease of commercial service implementation through a non-invasive structure based on external observation data.

[0174] 8.5 Generative AI Ad Engine

[0175] The generative AI ad engine is a module that receives visual feature vectors, environmental state information, and ad constraints transmitted from the ad control unit as input, and automatically generates ads that are visually consistent with the worldview of a game or real-time interface, or reprocesses pre-generated assets.

[0176] (1) Definition and processing of conditional inputs

[0177] By integrating visual features (F), system status (L), and advertising constraints (B)

[0178] Construct the conditional vector Z_t.

[0179] Z_t = [F_t ? L_t ? B_t]

[0180] Z_t is injected into the intermediate layer of the generative model and transformed to maintain visual consistency with color tones, lighting, and interface emotional tones. In this specification, "injection into the intermediate layer" means providing a conditional vector to any hidden layer between the input layer and the output layer during the inference process of a generative artificial intelligence model. Such injection into the intermediate layer is intended to control the influence of the condition on the output result more directly than simple combination at the input stage.

[0181] (2) Inverse rendering-based physical property estimation

[0182] Using a deep learning-based inverse rendering model from the input image I(x,y)

[0183] Estimate the normal N(x,y), lighting direction L(x,y), and shadow consistency S(x,y).

[0184] ∇I(x,y) → {N(x,y), L(x,y), S(x,y)}

[0185] This is utilized as conditional information to directly control the normal, lighting, and shadow alignment of the ad pixels to be generated within the generative model.

[0186] (3) Generative model structure and conditioning injection

[0187] This engine includes Diffusion, Transformer, GAN, or a combination of these models, and

[0188] The conditional vector Z_t and physical properties (N,L,S) are injected as follows.

[0189] U'_t = U_t ? g(Z_t, N, L, S)

[0190] Here, g() is a function that transforms U_t to the same dimension.

[0191] (4) Physical and visual alignment cost function

[0192] The generated result G minimizes the following composite cost function.

[0193] Loss_total = lambda1 * Loss_SSIM(G, I)

[0194] + lambda2 * Loss_LPIPS(G, I)

[0195] + lambda3 * Loss_physics(N, L, S)

[0196] Loss_physics = a * ||N - N_G||^2

[0197] + b * ||L - L_G||^2

[0198] + c * ||S - S_G||^2

[0199] The above weight coefficients lambda1, lambda2, lambda3 and a, b, c are

[0200] It is adjustable according to learning objectives and environments, and the scope of the rights of the present invention is not limited to specific coefficient values.

[0201] (5) Switching between real-time generation and pre-generation modes

[0202] It generates assets in real-time based on predictive trigger conditions, and immediately displays pre-generated assets selected upon advertiser requests or in the event of asset quality degradation or delays. This ensures the continuity and stability of ad displays.

[0203] (6) Post-processing and security marking procedures

[0204] The generated advertisement undergoes rule-based correction immediately before display, followed by decryption, display, and immediate plaintext discarding within the security module's trusted execution environment. This security procedure takes place within a single trusted path.

[0205] (7) Effect

[0206] This engine enables natural ad insertion without compromising user immersion through physics-based world-view fusion. Furthermore, it maintains an SDK-free structure, allowing for immediate application in commercial environments.

[0208] 8.6 Security Modules and Trusted Execution Environments

[0209] The security module according to the present invention is configured to perform the entire process of decoding, displaying, and discarding advertising content within a single trusted path, and fundamentally prevents advertising assets from being leaked or tampered with externally.

[0210] The security module protects decryption keys and asset data through a hardware-based Trusted Execution Environment (TEE) or an equivalent level of isolated security container.

[0211] (1) Decryption and Key Management

[0212] Ad assets are stored in an encrypted state and decrypted within the TEE only at the time of ad display. Decryption keys are generated and stored in the TEE's internal security key storage, making external access impossible. Decrypted plaintext assets are immediately discarded after display and do not remain in memory. The encryption method uses AES-GCM or an authentication cipher with an equivalent security level, and unique keys and initialization values ​​are automatically generated for each session and are not reused.

[0213] (2) Display path protection

[0214] Decrypted assets are passed to the rendering pipeline only within the TEE or an equivalent security area, and are composited onto the screen via official APIs provided by the operating system. Modification of game or application source code, memory hooking, or unauthorized copying are not performed.

[0215] (3) Immediate disposal procedure

[0216] Immediately after the ad display is complete,

[0217] a. Plaintext data memory overwriting,

[0218] B. Destruction of decryption key and session parameters,

[0219] C. TEE Context Initialization,

[0220] It automatically performs this to prevent the possibility of recycling and leakage of plaintext data.

[0221] (4) Watermarking and enforcement of rights

[0222] At the time of decryption, an invisible watermark containing non-identifiable terminal information, display time, and advertiser ID is inserted. The watermark is inserted in an encrypted format in the frequency domain or alpha channel and can be utilized for tracking verification without being recognized externally. In this specification, “invisible watermark” refers to an invisible signal component inserted into the frequency domain or alpha channel of an image through cryptographic processing or transformation, and means information utilized for tracking or verification within a range that does not affect the visual perception quality of the advertising content upon external observation.

[0223] (5) Response to environments not supported by TEE

[0224] In terminals or virtualization environments where TEE is not supported, the memory isolation and integrity verification functions provided by the operating system

[0225] Examples: Hypervisor Enforced Code Integrity (VBS) on Windows, HVCI,

[0226] Linux's seccomp or namespace-based isolation

[0227] The same procedure is performed in an alternative security container configured using.

[0228] Decryption and display processing are performed only in temporary session memory, and data retention is prevented by simultaneously deleting snapshots and overwriting memory immediately upon session termination.

[0229] (6) Effect

[0230] This security module is

[0231] a. Securing a continuous trust path from decoding to display to discard,

[0232] B. Prevention of advertising asset leakage and tampering,

[0233] C. Support for rights enforcement through session-based tracking,

[0234] D. SDK-Free Non-invasive Structure Maintenance,

[0235] It achieves the desired effect simultaneously. As a result, it can be integrated and applied without conflict with DRM and anti-cheat systems in various environments, such as games, LLM, and streaming.

[0237] 8.7 Feedback and Advertising Effectiveness Monitoring Department

[0238] This module provides a closed-loop learning structure that periodically optimizes ad performance by collecting user responses and system states in an anonymized form after ad display.

[0239] (1) Collection of response indicators

[0240] - View Duration

[0241] - Visual attention (view center dwell time, etc.)

[0242] - Ad avoidance metrics (volume reduction, screen movement, etc.)

[0243] - User input delay (optional)

[0244] Note: Eye-hold time and visual attention are indirectly estimated through external observation interfaces or screen movement / manipulation signals, and all data is anonymized and no personal information or original text is collected.

[0245] (2) Creative Performance Evaluation

[0246] The collected response data is weighted statistically to calculate the ad performance score, Score_ad.

[0247] Score_ad = w1 × ClickRate + w2 × ViewRate + w3 × DwellTime

[0248] The above weights w1, w2, and w3 are the performance goals of the advertiser or platform.

[0249] It can be dynamically configured based on (e.g., click-through rate or eye retention).

[0250] (3) Reinforcement learning-based feedback reflection

[0251] If ad performance is below the standard,

[0252] Conditioning parameter (Z_t) in the direction of maximizing Score_ad,

[0253] Automatically optimizes display timing (P_gap) and ad control thresholds (τ_fit, etc.).

[0254] The policy depends on data availability and system computational load.

[0255] It is updated by selecting either an offline or online reinforcement learning method.

[0256] (4) Personal Information and Regulatory Compliance

[0257] All feedback data is stored in a de-identified and statistical form and complies with relevant regulations, including the GDPR and domestic Personal Information Protection Act. Even in an LLM environment, direct user content, such as query sentences or response text, is not collected.

[0258] (5) Effect

[0259] This module directly reflects the results of an ad display at the time of the next display to continuously optimize ad efficiency without compromising the user experience, and realizes a complete closed-loop improvement structure even in an SDK-free, non-invasive environment.

[0261] 8.8 IP Rights Enforcement Module

[0262] This module is a component that manages the display process and history of advertising assets with integrity to detect illegal copying or unauthorized use and technically enforce IP rights.

[0263] (1) Ad display log records

[0264] The following items are recorded in real time during the ad display process.

[0265] - Ad asset identifier (encrypted form)

[0266] - Display timestamp

[0267] - Display environment information (client or server)

[0268] - Security Module Status Information

[0269] Since the above data is processed within the same trusted execution environment as the security module, the possibility of external tampering can be blocked.

[0270] (2) Inserting an invisible watermark

[0271] An invisible watermark containing the following information is inserted at the time of advertising asset decryption.

[0272] - Session ID (anonymized) or terminal non-identification hash

[0273] - Display time and exposure cycle

[0274] - Advertiser ID or campaign information

[0275] Watermarks are embedded using frequency domain or alpha channel-based encryption methods

[0276] It cannot be perceived on the user screen.

[0277] (3) Detection of counterfeiting and reuse

[0278] When externally captured or duplicated advertising frames are collected, the watermark is extracted and its legitimacy is verified through the following procedure.

[0279] - Dedicated watermark decryption network (deep learning-based) or

[0280] - Frequency domain analysis techniques (e.g., DCT transform-based)

[0281] This allows for the determination of unauthorized reuse, ad replacement, and illegal copying.

[0282] (4) Deletion integrity verification

[0283] After the display is complete, the following is automatically performed during the disposal phase.

[0284] ① Memory overwriting (zeroing) of plaintext assets

[0285] ② Hash generation and separate storage of discarded assets

[0286] ③ Context Initialization and Key Destruction

[0287] This completely blocks the possibility of asset extraction even after disposal.

[0288] (5) Ledger integration (optional)

[0289] Ad display history can be linked with one or more of the following storage technologies.

[0290] - Central server DB (encryption table and hash chain-based tamper-proof structure)

[0291] - Distributed ledger technology (e.g., permissioned blockchain)

[0292] The structure provides the following effects.

[0293] - Prevention of historical data forgery

[0294] - Transparent advertising settlement (accounting documentation)

[0295] - Increased possibility of third-party audits

[0296] (6) Effect

[0297] This module provides the following technical effects.

[0298] - Prevention of leakage, forgery, and reuse of advertising assets

[0299] - Ensuring the integrity and reliability of the display history

[0300] - Lawful enforcement possible with platform approval

[0301] In addition, it is linked with a non-invasive security display structure to protect advertiser IP and brand integrity, and

[0302] It realizes full rights enforcement even under an SDK-Free structure.

[0304] 8.9 Feasibility and Commercial Implementation Examples

[0305] The present invention provides an advertising system that can be implemented in current commercial hardware and operating environments by combining a non-invasive (SDK-Free) structure with a conditional generative algorithm. This embodiment is distinguished from simple preloading cache technology and dynamically controls the timing of ad display using minimal state information (API latency, processing time, etc.) provided by the platform.

[0306] (1) Preloading-based real-time ad implementation

[0307] The ad control unit can prepare pre-generated assets by considering the characteristics of sections or genres where repeated exposure is predicted. Pre-generated assets are produced by applying the same conditional generation algorithm and are adopted only when they satisfy SSIM 0.85 or higher and LPIPS 0.15 or lower. Since rendering occurs immediately after decoding at the time of display, delay-free display is possible regardless of GPU load and network conditions.

[0308] (2) Non-invasive control utilizing status information provided by the platform

[0309] The interval is determined using only state information officially provided in each environment, such as games, LLM, and streaming. Examples include LLM response latency (x-processing-ms), GPU load, frame rate, RTT, etc. Since it does not access raw user data, it does not contain personal information. This means that the timing of ad display can be determined solely through external observation, without the need for SDK insertion.

[0310] (3) Hardware and software configuration

[0311] The multimodal time series analysis unit can be composed of TensorFlow / PyTorch-based LSTMs and Transformers. The generative advertising engine can utilize Diffusion, GAN, or hybrid models that have been lightweighted through quantization or pruning techniques. The TEE-based security module can be implemented using ARM TrustZone, Intel SGX, etc., and in environments that do not support TEE, an alternative security container based on memory isolation is used.

[0312] Display delay in example environment:

[0313] - Real-time generation mode: Within approximately 2 seconds (example)

[0314] - Preloading mode: Within approximately 0.1 seconds (example)

[0315] Note: The above delay figures (e.g., real-time generation about 2 seconds, preloading about 0.1 seconds) are merely examples of implementation to demonstrate the technical effects of the present invention and do not limit the scope of interpretation of the claims to the above figures.

[0316] (4) Generalizability through transfer learning

[0317] The present invention learns common signal patterns for each platform using a transfer learning method, thereby maintaining, for example, an accuracy of 90% or higher in determining ad suitability without separate retraining even in new environments; this is a technical effect achieved by a multimodal fusion and prediction trigger structure. Note: The above accuracy figures are merely examples of implementation and do not limit the scope of the rights of the present invention to specific figures.

[0318] (5) Commercialization application method

[0319] This system can be distributed across client, edge server, and cloud environments. Network load and response latency can be minimized through distributed storage of preloaded assets and an edge server-based predictive decision method. This process does not require modification of internal application code or the insertion of SDKs.

[0320] (6) Conclusion

[0321] Therefore, the present invention is technically fully implementable in the current real-time responsive service environment and satisfies all requirements for ad quality (visual consistency), non-invasiveness, security, and real-time capabilities. Furthermore, it has a structure that can naturally expand to a real-time full-screen generation method as computational performance improves.

[0323] 8.10 Platform-specific Implementation Examples

[0324] The present invention can be implemented with the same "non-invasive structure" in various operating system environments such as Android, Windows, macOS, and iOS. In particular, since the iOS environment has special restrictions in which access to rendering buffers and the display of external overlays by other applications are restricted according to sandbox and UI protection policies, the present invention includes the following practical alternative implementation methods.

[0325] (1) Cloud rendering-based structure

[0326] Cloud-rendered games or real-time interfaces run on the server, and ad content is composited into the frame buffer on the server side. Since iOS devices receive and display the content via a streaming method (WebRTC or a similar structure), ads can be displayed without violating sandbox restrictions. Ad assets are decrypted and displayed within the server's trusted execution environment, after which they are immediately discarded, ensuring that no plaintext assets remain on the device.

[0327] (2) Overlay structure within the same application: The ad display module is inserted into the Render Pass or SceneGraph stage within the same application. Since the ad is rendered within the same process without external access, it does not compromise iOS app signature and sandbox integrity. The ad is displayed only within a composite buffer (such as CALayer-based) officially permitted by the operating system.

[0328] (3) Method of using the official API

[0329] Performance metrics and network status can be collected with user consent through interfaces officially provided by the operating system, such as ReplayKit, MetricKit, and Network Framework. The collected data includes only beacon context performance signals, without directly extracting or manipulating screen pixels or original UI. Therefore, a "non-invasive structure" can be maintained without conflicting with platform policies.

[0330] (4) Technical Conclusion

[0331] The present invention enables the display of advertisements in all major operating system environments by selectively using one layer among a server, cloud, edge, or within the same application, even if access restrictions exist depending on the operating system. This is a specific embodiment of the "formal interface-based non-invasive structure defined in Claim 14."

[0332] (5) Security processing

[0333] Ad assets are decrypted, marked, and discarded in a TEE or equivalent security area, and decryption keys are managed only in an area inaccessible to the outside. After marking is complete, plaintext assets are immediately deleted, thereby preventing asset leakage and tampering. This corresponds to the "secure marking pipeline of Claim 4".

[0334] (6) Example of application in a cloud gaming environment

[0335] Cloud gaming or AI streaming services are already provided based on iOS Safari or in-app WebRTC. In these environments, this system can perform server-side or hybrid synthesis to non-invasively detect loading, transition, and pause sections and display advertisements immediately.

[0336] (7) DRM and anti-cheat system compatibility

[0337] In a commercial game environment, a DRM or anti-cheat system may operate simultaneously. The system may optionally perform whitelist registration and process authentication provided by a publisher or platform, and this operates within the scope of the official interface of the operating system. This embodiment corresponds to the "handshake procedure of Claim 17" and prevents policy conflicts while maintaining an SDK-Free structure.

[0339] 8.11 Example of applying ads using game loading sections

[0340] According to one embodiment of the present invention, a section where direct user input is temporarily restricted, such as a game loading or scene transition section, can be utilized as a suitable section for displaying advertisements. The section is primarily based on the joint abnormal pattern determination results of visual signals, system signals, network signals, and audio signals detected by a multimodal time-series analysis unit, and refers to external user experience analysis results as a supplementary indicator.

[0341] (1) Temporal definition of the loading period

[0342] The average duration of a game loading period ranges from approximately 5 to 30 seconds depending on the engine structure and genre. The present invention is based on loading status information determined by a multimodal time-series analysis unit and can define a suitable period for displaying advertisements by considering the results of user experience analysis (e.g., 5 to 20 seconds). This period is a time when user attention is maintained while operation is interrupted, and advertisements can be displayed within a range that does not cause cognitive fatigue.

[0343] (2) Ad insertion method

[0344] The ad control unit preemptively issues an ad generation request 0.5 to 3 seconds before the start of the predicted loading period. The generative AI ad engine uses features such as color palettes, lighting direction, and object placement as conditional inputs to generate ads that are visually consistent with the game worldview. Ads are displayed by naturally blending with the loading gauge or background objects, and the display method is automatically selected based on one or more of the following conditions.

[0345] a. When the estimated loading length is short: Fade-in / Fade-out method

[0346] B. Cases of high ad complexity: Texture replacement type or object fusion type

[0347] C. Where advertiser brand guide constraints exist: Static image-centered display

[0348] This enables natural context fusion.

[0349] (3) Pre-generated asset management

[0350] Since loading sections are highly repeatable, the ad control unit can pre-generate and store ad assets optimized for specific scenes. Pre-generated assets are produced using the same conditioning algorithm as real-time generation, enabling immediate decoding and display when a loading section occurs. Furthermore, if the reliability score of the generated result is determined to be below a certain threshold, pre-generated assets are prioritized to ensure ad quality and display stability.

[0351] (4) Actual application example

[0352] For example, in a fantasy RPG game, while a loading gauge is displayed, the character moves in sync with the bar, and the brand logo can be naturally exposed by fusing it into the character's equipment. Upon completion of loading, the brand slogan appears as the advertisement gradually fades out, and since gameplay continues without lag, users do not perceive the advertisement as an external element.

[0353] (5) Technical effects

[0354] This embodiment is differentiated from existing technologies in that it determines the loading section and generates and displays advertisements using only external observation signals, without the need for SDK insertion. Furthermore, by operating real-time generation paths and pre-generated assets in parallel, it is possible to simultaneously ensure ad quality, system stability, and user experience.

[0356] 8.12 Example of Ad Application Using Large-Scale Language Model (LLM) Lag Periods

[0357] In addition to games, the present invention can utilize the response delay interval between the completion of user input and the generation of a response by the model as a suitable interval for displaying advertisements in a conversational AI environment based on a Large Language Model (LLM). This interval corresponds to a "cognitive gap" where visual attention is temporarily dispersed while the user's concentration is maintained, and the multimodal time-series analysis unit detects this by combining external non-invasive signals such as GPU computational load, API response delay, and token streaming speed.

[0358] (1) Temporal definition of the LLM response waiting period

[0359] The response latency of a large-scale language model varies depending on server load, the scale of model parameters, and network RTT. The present invention is based primarily on the joint anomaly pattern determination results of a multimodal time series analysis unit and, by referring to publicly available service performance measurement data (e.g., LLM response latency analysis data), defines a cognitive gap of 2 to 10 seconds and configures it to be expandable up to a maximum of 12 seconds depending on the environment. The above time figures are examples of a range that does not impair the user experience and do not limit the scope of the claims to specific figures.

[0360] (2) Ad Trigger Detection Structure In this embodiment, a 'non-contextual signal' refers to state information at the metadata level provided by a platform or interface, which does not directly include semantic content such as the original text of a user's query or response text. The non-contextual signal may be generated at the control or orchestration layer of the platform and includes state information that can be received in the form of a data object or event having a certain field structure, and is characterized by being composed only of numerical values ​​and state values ​​at the infrastructure level without semantic analysis of the conversation content.

[0361] The analysis unit determines that if two or more of the following beacon context signals satisfy the threshold condition, it is a suitable section for displaying advertisements.

[0362] ① Patterns of surge or fluctuation in GPU / CPU usage

[0363] ② Changes in API request size or number of input tokens

[0364] ③ Network RTT or response metadata delay value

[0365] ④ Irregularity in token generation rate (increase in Δtoken / sec)

[0366] ⑤ UI “Preparing Response” or “Loading” event status

[0367] The above signal is an external, non-identifiable signal unrelated to LLM internal conversation data; in web-based UIs, it can be obtained through JavaScript DOM event detection with user consent, while in application-based UIs, it is measured non-invasively using only official OS performance counters.

[0368] (3) Ad generation and conditional methods

[0369] The ad control unit executes ad generation requests 0.5 to 3 seconds prior to the predicted cognitive niche start time, and the generative AI ad engine uses minimal information such as UI state, request length, response latency, category tags, and GPU utilization as conditional vectors. Personal information, such as user queries, model response text, or internal prompts, is not collected. The generated results are configured to maintain consistency with the color tone, mood, and platform brand style of the corresponding UI.

[0370] (4) Parallel operation of pre-generated assets

[0371] Since the LLM response interval is short and the repetition frequency is high, the ad control unit can store pre-approved high-quality ad assets in a preloaded form. If the real-time generated result falls below the quality standard or the generation time exceeds a threshold time (e.g., 3 seconds), a pre-generated asset is selected first and displayed immediately. The threshold time can be adjusted according to the implementation environment and does not limit the scope of the claims.

[0372] (5) Security and privacy protection

[0373] This embodiment controls advertisements using only external signals provided by the platform, and does not collect user conversation content or internal data of the model in any form. The decryption, display, and disposal of advertising assets are performed only within a TEE or equivalent security container, and display logs are stored only in an anonymized form to satisfy privacy regulations (GDPR, CCPA, PIPA, etc.).

[0374] (6) Actual application example

[0375] While the user waits for a response after entering a car-related question, the "car" context can be determined from category tags and UI status signals. A car silhouette naturally appears in the loading indicator area, and after a brand slogan is displayed for 3 to 5 seconds, it fades out simultaneously with the actual response being output. Users perceive the advertisement as part of the interaction flow and do not feel a sense of incongruity.

[0376] (7) Technical effects

[0377] This embodiment realizes a new non-invasive advertising format even in existing text-based LLM interfaces by utilizing response waiting time as an advertising channel. Furthermore, since no internal data is used, legal and ethical risks are minimized, and stable and high-quality ad display is achieved by combining real-time generated and pre-generated advertisements.

[0379] 8.13 Differentiation from Prior Art

[0380] The present invention is characterized by a closed-loop system that predicts suitable sections for ad display and generates and displays ads using only external observation signals officially provided by an operating system or graphics API, without modifying the source code of an application, inserting an SDK, or accessing runtime memory, and performs decryption, display, and immediate disposal in an integrated manner within a security module. This is essentially distinct from existing SDK-based or static overlay methods.

[0381] (1) Summary of Key Differentiating Points

[0382] a. Realizes an SDK-free, non-invasive structure without infiltrating the program.

[0383] B. Predicts and determines the appropriate section for ad display by fusion-analyzing screen, system, network, and audio signals in a time-series manner.

[0384] c. Generates advertisements that naturally fuse with the worldview by conditioning visual characteristics such as scene color, lighting, and layout.

[0385] D. Decrypt, display, and discard advertisements in a TEE or equivalent security environment.

[0386] E. It is equipped with prediction-based preemptive generation, pre-generated asset management, and fallback-based live service stability.

[0387] (2) Differences from cited patent document 1

[0388] Cited Invention 1 is a structure that displays advertisements by receiving internal game engine events via an SDK method. Since it must be directly integrated into the game code or engine, the implementation costs and maintenance burdens per platform are high, and it cannot operate if the publisher does not display the advertisement events. On the other hand, the present invention predicts and determines the appropriate section for displaying advertisements based solely on external observation signals from the operating system or graphics API, thereby enabling independent operation without internal program access. Furthermore, the present invention combines conditional generative advertising with TEE security procedures, thereby simultaneously achieving real-time worldview consistency and asset protection.

[0389] (3) Differences from cited patent document 2

[0390] Cited Invention 2 is centered on overlay insertion for general video or streaming content and corresponds to post-analysis or single-signal-based processing. On the other hand, the present invention performs preemptive ad generation requests and conditional generative ad insertion based on the state of a real-time running game, interactive AI, etc., and realizes a closed-loop structure including a secure display pipeline. Therefore, it is fundamentally distinguished in technical purpose and processing method.

[0391] (4) Summary of differences from other prior art

[0392] a. Existing technologies are based on SDK insertion or memory access, so they conflict with program security modules or have high update dependencies. Since the present invention is based on external observation, it is independent of the engine and platform.

[0393] B. In existing technologies, ad timing is fixed or displayed based on a single signal, resulting in a disconnect in the user experience. The present invention enables preemptive control based on multimodal prediction.

[0394] C. While existing technologies lack adequate advertising security processing or leakage prevention functions, the present invention enables TEE-based decryption, display, and disposal, as well as watermark-based tracking.

[0395] (5) Difficulty in combining

[0396] Prior art technologies focus on SDK-based control and post-analysis overlay environments, and there are no examples of combining multimodal time-series analysis, conditional generation, and secure closed-loop systems into a single structure. Since the sequential structure of the present invention achieves both performance and security simultaneously through the interdependence between modules rather than a simple combination of functions, it cannot be easily derived solely by combining prior art technologies.

[0397] (6) Conclusion

[0398] The present invention is distinguished from existing technologies in terms of technical purpose, configuration, and processing method, and in particular satisfies novelty and inventive step through the simultaneous implementation of an SDK-free non-invasive structure, multimodal prediction, conditional generative advertising, and a TEE-based secure closed loop. Accordingly, Claim 1 and the dependent claims are subject to protection under patent law.

[0400] - Cited Literature

[0401] Patent Document 1: US20220126206A1 (Frameplay)

[0402] Patent Document 2: US11356746B2 (GumGum)

[0403] Patent Document 3: US11260299B2 (Anzu)

[0404] Patent Document 4: US20240095778A1 (Bidstack)

[0405] Patent Document 5: US20070072676A1 (Google)

[0406] Patent Document 6: GB2576435A (AdInMo)

[0407] Patent Document 7: KR10-1478172B1 (IGAWorks)

[0408] Patent Document 8: KR2019-0128591A (Buzzvil)

[0409] Patent Document 9: KR20100088096A (TV viewership measurement system)

[0410] Patent Document 10: JP2017198320A (UI state-based information overlay system) Industrial applicability

[0412] The present invention relates to an integrated system based on an “SDK-Free non-invasive” structure that analyzes and predicts multi-modality signals in a time series using only the official external observation interface of an operating system or graphics API, without modifying program code or hooking memory, and decodes, displays, and discards conditioned generative AI advertisements within a trusted execution environment (TEE, etc.). Since this structure can be repeatedly applied across multiple digital content industries, such as games, conversational AI (LLM), streaming, AR / VR, and cloud, it clearly satisfies the industrial applicability requirement under Article 29, Paragraph 1 of the Patent Act.

[0413] (1) Game industry sector

[0414] The present invention can be directly applied to mobile, PC, console, and cloud gaming environments, as well as general-purpose engines such as Unity and Unreal. It detects and predicts suitable ad display sections, such as loading and transitions, in real time using only external observation signals without installing SDKs or modifying engine builds, and provides advertisements conditioned on the worldview via overlay or server synthesis, thereby enabling rapid implementation in both new and existing titles.

[0415] The present invention provides the following industrial advantages.

[0416] A. Significant reduction in integration and verification costs: The SDK-Free structure significantly reduces the burden of game-specific customization and QA.

[0417] B. Stable Inventory Expansion: Loading, switching, and pause sections can be converted into a new inventory, allowing for additional exposure per session.

[0418] C. UX Protection and Brand Safety: Ensures visual consistency through conditional generation and prevents asset leakage through TEE-based closed-loop security.

[0419] D. Operational Compatibility: It can be safely deployed in commercial service environments by combining an optional whitelist handshake procedure to prevent DRM / anti-cheat conflicts with a non-invasive marking path.

[0420] (2) Conversational Artificial Intelligence (LLM) and Streaming Industry

[0421] In an LLM environment, "cognitive gaps" are non-invasively detected using beacon context signals—such as GPU / CPU load, API round-trip latency, token generation speed, and UI state signals—in the interval between user input and the initiation of a model response. The category and display intensity are then automatically determined within the minimum state information range provided by the platform. Since semantic content, such as conversational text and voice, is not accessed during this process, it is advantageous for privacy and platform policy compliance. In a streaming environment, natural insertion timing is calculated by combining signals such as frame rate, buffering events, network latency, and audio spectrum, enabling the automatic generation and display of advertisements conditioned on the scene's color tone, lighting, and style.

[0422] From an industrial perspective, the present invention enables various business models, such as monetization of the free user segment, parallel operation with a premium (ad-free) model, and multi-media shared creative operations (asset sharing among games, LLM, and streaming).

[0423] (3) AR·VR and Metaverse Industry

[0424] Multimodal time-series judgment and condition generation can be applied in the same way to periods of concentration relaxation, such as scene transitions, teleports, and focus transitions in AR and VR content. By utilizing spatial characteristics such as field of view, lighting, and object distance as condition vectors, immersive advertising is naturally integrated. Since TEE-based decryption, display, and discarding procedures can be implemented by utilizing the hardware security modules (TrustZone, Secure Enclave, etc.) of major devices such as Quest and Vision Pro, immediate commercialization is possible in XR and metaverse environments.

[0425] (4) Cloud and platform industry sectors

[0426] The components of the present invention can be distributed across the client, edge, and cloud, and server synthesis, client overlay, or hybrid paths can be selectively operated depending on service conditions. In the case of cloud gaming and streaming, the structure transmits video to the terminal after ad synthesis and security processing (TEE) on the server side, thereby simultaneously reducing latency and the risk of asset exposure. This architecture can be extended to various commercialization models, such as SaaS (Ad Context Determination / Conditional Generation API), PaaS (Platform as a Service), and on-premise (dedicated to large publishers).

[0427] (5) Economic effects

[0428] A. Development / QA Cost Reduction: Substantially reduces integration and verification costs per title through an SDK-Free structure (eliminating integration cost items commonly cited in industry reports and developer community data).

[0429] B. Increased Inventory and Improved Efficiency: By converting loading, transition, and pause sections into inventory, additional exposure per session is realized, and improved engagement metrics are expected by reducing visual resistance through conditional generation. Since specific figures may vary depending on the implementation environment, genre, and creative, they are presented at the level of representative embodiments in this specification and will be optimized through A / B verification for commercial distribution.

[0430] C. Security Effect: Structurally suppresses the risk of asset leakage, forgery, and alteration through TEE-based decryption, marking, immediate destruction, and dynamic watermarking.

[0431] D. Market scalability: Diversified revenue sources can be secured by simultaneously creating new inventory such as in-game ads, LLM response waiting, streaming buffering, and XR scene transitions.

[0432] (6) Conformity in terms of regulation, policy, and personal information protection

[0433] The present invention receives and utilizes only the minimum state information officially provided by the platform as a beacon context signal and does not process semantic content such as conversational text or voice. Logs are subjected to anonymization and pseudonymization procedures, and data retention periods, purpose restrictions, and access controls are designed to comply with regional privacy regulations. Advertisement display is performed only through official routes of the operating system and graphics APIs to avoid conflicts with DRM and anti-cheat policies (a whitelist handshake procedure is optionally performed if necessary).

[0434] (7) Feasibility in terms of production, distribution, and operation

[0435] A. Production: The model can be deployed after pre-training with public, synthetic, or publisher-provided data and performing genre-specific fine-tuning. The conditioning / judgment module can be configured with lightweighting, quantization, on-device inference, or edge serving.

[0436] B. Deployment: Due to its SDK-Free nature, the burden of client updates is low, and operation is possible through rolling updates on the edge / cloud side.

[0437] C. Operations: Enhance availability through operational automation, including preemptive triggers, pre-generated assets, non-generated fallback, distributed deployment, and anti-cheat compatible procedures, and monitor and tune performance, security, and quality metrics via dashboards.

[0438] (8) Overall evaluation

[0439] This invention integrates SDK-free external observation, multimodal time-series judgment and prediction, conditional generative advertising, TEE-based display and discarding, and fallback, distribution, and anti-cheat compatible operations into a single closed-loop system. This structure simultaneously satisfies technical reproducibility (cross-platform portability, stable management of prediction and display latency), economic effectiveness (reduced integration costs, expanded inventory), and industrial scalability (games, LLM, streaming, XR, and cloud), thereby demonstrating high industrial applicability. Furthermore, by supporting the parallel operation of pre-generated assets and real-time generation, it provides a practical balance between quality, latency, and cost. Therefore, this invention has a significant impact as a core, practical AI advertising technology that can be immediately utilized by various industry entities (publishers, platforms, advertisers, and cloud providers).

[0440] No content Explanation of the symbols

[0441] 100: Data acquisition unit 110: Multimodal Time Series Analysis Unit 120: Context-aware ad controller 130: Generative AI Ad Engine 140: Security Module 150: Feedback and Monitoring Department 160: IP Rights Enforcement Module 170: Operations Management Department

Claims

Claim 1 In a non-invasive artificial intelligence advertising system configured to determine an appropriate section for displaying an advertisement by chronologically analyzing two or more state signals belonging to different categories in a real-time responsive interface environment including games, conversational artificial intelligence, streaming, augmented reality (AR), or virtual reality (VR), without modifying the source code or executable binary of a program or directly inserting or hooking it into runtime memory, and using only an external observation interface or screen synthesis interface that does not involve accessing the internal memory of a target program or code injection, where an operating system or graphical application programming interface (API) is provided as officially documented in a publicly available technical document, the system comprises: a processor; a memory; and when stored in the memory and executed by the processor, a a. Multimodal time series analysis processing that aligns two or more status signals belonging to different categories among visual information, system resource status information, network status information, or sound information received through the external observation interface or screen synthesis interface based on a time axis, calculates a statistical analysis indicator including one or more of variance, correlation coefficient, or time series change rate from the status signals, and determines a time interval in which the statistical analysis indicator satisfies a preset condition as an appropriate interval for ad display; b. Context-aware ad control processing that controls the timing of ad display based on the determined appropriate interval for ad display; c. Condition-based generative AI ad generation processing that uses encrypted and stored ad assets to generate ad content suitable for the visual context corresponding to the appropriate interval for ad display or to select said ad assets; and d.A non-invasive artificial intelligence advertising system comprising one or more program modules configured to perform secure display processing, wherein the above-mentioned advertising assets are logically isolated at the hardware or operating system level and access by external processes is blocked in a Trusted Execution Environment or an equivalent isolated execution environment, and the decrypted plaintext data is immediately discarded after the display is completed; and characterized by performing non-invasive advertising display based on the results of time-series relationship analysis between the state signals without direct observation of the original content data or user identification information. Claim 2 A method for non-invasively controlling advertisements in a real-time responsive interface environment, performed by a computing device including a processor, comprising: a. a step in which the computing device collects two or more status signals belonging to different categories among visual information, system resource status information, network status information, or acoustic information through an external observation interface or a screen synthesis interface, which does not involve access to the internal memory of a target program or code injection, wherein an operating system or a graphical application programming interface (API) is provided and officially documented in a publicly available technical document; b. a step in which the computing device aligns the collected status signals based on a time axis, calculates a statistical analysis indicator including one or more of variance, a correlation coefficient, or a time-series change rate from the status signals, and determines a time interval in which the statistical analysis indicator satisfies a preset condition as a suitable interval for displaying advertisements; c. a step in which the computing device predicts the start time of advertisement display based on the determined suitable interval for displaying advertisements, and performs a processing request for advertisement creation or advertisement selection prior to the predicted time; d. The computing device extracts visual features corresponding to one or more of the color distribution, lighting pattern, or object placement information of the screen and uses them as conditional inputs for a generative artificial intelligence model trained to generate an image or video by receiving conditional inputs, thereby generating visually consistent advertising content or selecting pre-generated advertising assets based on similarity with said visual features; and e.A non-invasive advertising control method characterized by comprising the step of: the computing device decrypting the advertising content or advertising asset in a Trusted Execution Environment or an equivalent isolated execution environment and displaying it on a screen, and immediately deleting the decrypted plaintext data from memory after the display is completed; and controlling the display of an advertisement non-invasively based on the analysis of the time-series relationship between the state signals without modifying the source code of the program or accessing runtime memory. Claim 3 A preemptive trigger control device for minimizing ad display delay, comprising: a processor; a memory; and a program module configured to perform, when stored in the memory and executed by the processor, a. a prediction processing that predicts the start time of an ad display suitable interval from a multimodal time series signal; b. a scheduling processing that determines the request issuance time so that an ad creation request or selection request is completed before the start time by comparing the time difference between the predicted start time and the current time and the estimated processing time required for ad creation or selection; and c. a mode selection processing that selects one of a real-time creation mode, a pre-generated asset playback mode, or a non-generating alternative ad display mode based on a system state including one or more of the reliability of the prediction, computational load, network latency, or resource utilization. Claim 4 A secure display pipeline device that ensures a single trusted path from the decryption of an advertising asset to its display and disposal, comprising: a. a decryptor configured to store the advertising asset in an encrypted state and decrypt it only in a Trusted Execution Environment isolated at the hardware level or an equivalent isolated execution environment; b. a displayer that logically isolates the display path so that the decrypted plaintext data does not deviate from the protected display path within the isolated execution environment; and c. a disposal device configured to immediately dispose of the plaintext data immediately after display and to verify the integrity of the disposal result using a hash or a Message Authentication Code (MAC). Claim 5 A non-invasive advertising system that controls the display of advertisements using minimal state information that does not include original content data or user identification information received through an application programming interface (API) provided by a platform in a real-time responsive interface environment including a conversational artificial intelligence environment, comprising: a processor; memory; and one or more program modules configured to perform the following when stored in the memory and executed by the processor: a. receiving processing that receives the minimal state information which includes one or more of a user interface state, a request message length, a response generation delay time, a token generation speed, or a conversation topic category tag, and which does not include original text of the conversation content or user personal identification information; b. determination processing that analyzes the time-series change relationship of the minimal state information to determine an appropriate interval for displaying advertisements and determines an ad display time corresponding to the appropriate interval for displaying advertisements; and c. ad display control processing that controls the screen display of the advertisement content based on the determined ad display time; wherein the system is characterized by non-invasively controlling the display of advertisements using only the minimal state information without direct access to original data of the conversation content or user identification information. Claim 6 A system according to claim 1, wherein the multimodal time series analysis processing is configured to reduce false positives and false negatives in ad display section determination compared to single signal analysis by combining state signals of two or more different categories using one or more of early fusion, late fusion, or a hybrid fusion method combining the same. Claim 7 A system according to claim 1, wherein the multimodal time series analysis processing is configured to determine a time interval as an ad display suitable interval when the variance or correlation coefficient values ​​calculated for a plurality of state signals within a reference time interval each satisfy a preset threshold condition. Claim 8 The method of claim 2, wherein the time series analysis step comprises determining an ad display appropriate interval by observing a past continuous time interval based on the current time point using a time-dependent artificial intelligence model or probabilistic model including one or more of LSTM (Long Short-Term Memory), Temporal CNN, Transformer, or Hidden Markov Model (HMM). Claim 9 A system according to claim 1, wherein the advertisement generation process extracts at least one of a screen color distribution, lighting pattern, or object arrangement as a feature vector, and injects the feature vector into an intermediate layer during the generation inference process of an artificial intelligence model to enforce visual consistency. Claim 10 A system of claim 9, wherein the advertisement generation processing utilizes a deep neural network-based generative artificial intelligence model and is configured to perform at least one of character behavior integration, environment placement, lighting and color matching, or animation synchronization. Claim 11 The device of claim 3, wherein the mode selection processing is configured to switch between a real-time generation mode, a pre-generated asset playback mode, and a non-generated alternative ad display mode automatically or according to a set policy based on the reliability score of the ad content generation result and the system state. Claim 12 An apparatus of claim 11, wherein in the non-generating alternative ad display mode, the apparatus is configured to select and display the ad asset having the highest similarity among pre-approved ad assets based on an indicator quantifying visual similarity with the current screen. Claim 13 The device of claim 4, wherein the indicator inserts an invisible watermark including a session identifier or timestamp when displaying an advertisement, and the discarder is configured to verify the deletion integrity of plaintext data using a hash or a message authentication code (MAC). Claim 14 A system of claim 1, wherein the advertising content is rendered using at least one of a screen replication path, a screen capture path, or a screen synthesis path of an operating system, or an overlay rendering path officially provided by a graphics application programming interface (API), and configured to operate non-invasively without modifying the source code of a program, patching the executable binary, accessing runtime memory, or inserting an external SDK. Claim 15 A method according to claim 2, further comprising the step of the computing device dynamically adjusting the ad rendering load so that the frame rate degradation caused by ad rendering does not exceed a preset user experience threshold. Claim 16 A system according to claim 1, wherein the multimodal time series analysis processing determines a screen loading, resource cleanup, cache flush, map reload, or user inactivity period as a suitable period for displaying an advertisement. Claim 17 A system according to claim 5, wherein the minimum state information does not include the original text content of a conversation or screen and is a non-context signal provided in an anonymized or tokenized form. Claim 18 A non-invasive artificial intelligence advertising system according to claim 1, wherein a handshake procedure including process identifier or signature-based whitelist registration is performed to prevent conflict with an anti-cheat module, and configured to operate within the scope of an officially documented interface without program logic hooking. Claim 19 The apparatus of claim 11, wherein the pre-generated asset is generated and stored in advance by a condition input-based generation process, and configured to be decrypted and displayed at the time of display through a security path identical to the real-time generation path. Claim 20 An apparatus of claim 3, wherein the prediction processing is characterized by determining an ad display suitable interval by learning a pattern in which simultaneous anomalies appear within a reference time interval in two or more of GPU usage, frame change rate, disk, or network I / O patterns. Claim 21 A system according to claim 1, configured to use an authentication cryptographic method for security display, manage a decryption key in a key store within a trusted execution environment, and immediately destroy plaintext data and a decryption key after the advertisement display is completed. Claim 22 A non-invasive ad control method according to claim 2, comprising the step of performing a handshake procedure including a process identifier or signature-based whitelist registration to prevent conflict with an anti-cheat module, and further comprising the step of performing the ad display within the scope of an interface officially provided by an operating system or a graphical application programming interface (API) without program logic hooking. Claim 23 A non-invasive ad control method according to claim 2, further comprising the step of using an authentication cryptographic method for security display, managing a decryption key in a key store within a trusted execution environment, and immediately destroying plaintext data and a decryption key after the ad display is completed.

Images