Cooperative self-adaptive RTB advertisement compliance auditing system and method based on generative traceability

Abstract

The invention discloses a collaborative self-adaptive RTB advertisement compliance auditing system and method based on generative traceability, and relates to the field of digital advertisement and artificial intelligence. Comprising a data acquisition module which acquires multi-modal advertisement materials in real time from a supplier platform SSP, a demander platform DSP and an advertisement transaction platform ADX; and the SACE module is connected with the data acquisition module and is used for performing generative content tracing, causal anomaly detection, model training and edge deployment under privacy protection and energy consumption optimization on the multi-modal advertisement materials. According to the method, an SACE engine is taken as a core, generative content tracing is taken as a starting point, causal anomaly detection is fused, and a privacy protection model is trained under a federated learning framework; edge calculation ensures real-time execution, green optimization and dynamic adjustment of resources; and a self-adaptive feedback mechanism is connected with all links to form closed-loop optimization, so that efficient compliance auditing of the multi-modal AIGC materials is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field, and particularly relates to a collaborative adaptive RTB advertisement compliance auditing system and method based on generative provenance. BACKGROUND

[0002] The existing RTB advertisement auditing technology, such as CN106447366B, processes material features through a centralized AI model, but faces multiple challenges: centralized training easily leaks private data, violating the Data Security Law; auditing delay often exceeds 100 ms, which cannot meet the real-time needs of RTB; there is a lack of provenance verification for generative AI (AIGC) materials, which is easily affected by deep forgery and false bidding; energy consumption optimization is ignored, which does not meet the double carbon target. Related technologies involve federated learning or edge computing, but are mostly single applications, lacking integration of AIGC provenance and anomaly detection. The present application is designed by a human inventor and is not AI-generated, aiming to fill this gap. In 2025, with the rise of AIGC advertisements and the strengthening of regulations (such as AIGC labeling requirements), there is an urgent need for a system that integrates provenance, detection, privacy, real-time, and green mechanisms to improve compliance and efficiency.

[0003] To this end, the present application provides a collaborative adaptive RTB advertisement compliance auditing system and method based on generative provenance. SUMMARY

[0004] The present application aims to solve the shortcomings in the prior art and provides a collaborative adaptive RTB advertisement compliance auditing system and method based on generative provenance.

[0005] To achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0006] A collaborative adaptive RTB advertisement compliance auditing system and method based on generative provenance, comprising:

[0007] A data acquisition module that acquires multi-modal advertisement materials in real time from supply-side platforms (SSPs), demand-side platforms (DSPs), and advertisement exchange platforms (ADXs);

[0008] An SACE module connected to the data acquisition module, used for generative content provenance, causal anomaly detection, model training and edge deployment under privacy protection, and energy consumption optimization of the multi-modal advertisement materials;

[0009] The SACE module comprises:

[0010] The generative provenance and causal detection submodule extracts material features through a Transformer-based unified feature flow extraction technology, and verifies implicit watermarks and cross-modal fingerprints to confirm the source of the material, while a causal relationship graph is constructed based on a structured causal model (SCM) to detect abnormal behavior.

[0011] The federated learning and edge deployment submodule locally trains a provenance detection model on the edge node using the FedAvg algorithm and differential privacy technology, and only uploads encrypted gradients for aggregation.

[0012] The green optimization submodule predicts energy consumption through an LSTM network and dynamically allocates resources based on load.

[0013] Preferably, in the generative provenance and causal detection submodule, implicit watermark verification is used for source authentication of image and video materials, and cross-modal fingerprints are used to associate text descriptions with visual content for consistency verification and provenance.

[0014] Preferably, in the federated learning and edge deployment submodule, the differential privacy technology has an initial privacy budget ε set to 1.0-1.2, and when the high-concurrency scenario is greater than 200,000 per second, ε is adjusted to 1.5-1.8.

[0015] Preferably, in the federated learning and edge deployment submodule, the differential privacy technology has an initial privacy budget ε set to 1.0, and when the high-concurrency scenario is greater than 200,000 per second, ε is adjusted to 1.5.

[0016] A collaborative adaptive RTB advertising compliance auditing method based on generative provenance, comprising the following steps:

[0017] S1: Data collection: real-time collection of multi-modal advertising materials from supply-side platforms (SSPs), demand-side platforms (DSPs), and advertising exchange platforms (ADXs);

[0018] S2: Generative provenance and causal detection: extracting features of the multi-modal advertising materials through a Transformer-based unified feature flow extraction technology, and verifying implicit watermarks and cross-modal fingerprints to confirm the source of the material, while a causal relationship graph is constructed based on a structured causal model (SCM) to detect abnormal behavior;

[0019] S3: Federated learning and edge auditing: locally training a provenance detection model on the edge node using the FedAvg algorithm and differential privacy technology, and only uploading encrypted gradients for aggregation.

[0020] S4: Green optimization: predicting energy consumption through an LSTM network, and dynamically allocating computing resources according to the traceability and detection load to shield high-energy-consumption abnormal materials;

[0021] S5: Adaptive feedback optimization, using the reinforcement learning Q-learning algorithm to adjust the traceability threshold and resource allocation weight based on the output result of step S2, forming a closed-loop optimization, and feeding back the abnormal signal to the traceability and green optimization steps.

[0022] Preferably: in the S1 step, the multi-modal advertising material includes text, images, videos, AR content and AIGC generated content.

[0023] Preferably: in the S2 step, the model expression of the SCM is: P(Y / do(X))=Σ Z P(Y / X,Z)P(Z), wherein X is the material feature, Y is the abnormal index, and Z is the intermediate variable.

[0024] Preferably: in the S3 step, the aggregation model is: wherein w is the model parameter, n k is the data amount of node k, and n is the total data amount.

[0025] Preferably: in the S4 step, the LSTM state update model is: h t =tanh(W xhxt +W hh h t-1 +b h ). Wherein h t is the hidden state at the current time t, W xh and W hh are weight matrices, x t is the input vector at time t, and b h is the bias term.

[0026] Preferably: in the S5 step, the Q-learning updates the threshold value Q(s, a)=Q(S, a)+a[r+λmax α′ Q(s′, a′)-Q(s, a)], wherein s is the state, a is the action, r is the reward function, and λ is the discount factor.

[0027] In the S5 step, the reward function of the Q-learning algorithm is:

[0028] The present application has the following advantages:

[0029] 1. This invention takes the SACE engine as its core, starts with generative content tracing (implicit watermarking and cross-modal fingerprinting), integrates causal anomaly detection (structured causal model SCM), trains a privacy-preserving model under a federated learning framework, ensures real-time execution through edge computing, dynamically adjusts resources through green optimization, and connects each link through an adaptive feedback mechanism to form a closed-loop optimization, thereby achieving efficient and compliant review of multimodal AIGC materials. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the overall architecture of a collaborative adaptive RTB advertising compliance review system based on generative tracing proposed in this invention;

[0031] Figure 2 This invention presents a closed-loop flowchart of source tracing and causality detection for a collaborative adaptive RTB advertising compliance review method based on generative source tracing.

[0032] Figure 3 This is a pseudocode flowchart of the adaptive feedback optimization of a collaborative adaptive RTB advertising compliance review method based on generative tracing proposed in this invention. Detailed Implementation

[0033] The technical solution of the present invention will be further described in detail below with reference to specific embodiments.

[0034] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "setting" should be interpreted broadly. For example, they can refer to a fixed connection or setting, a detachable connection or setting, or an integral connection or setting. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0035] Example 1:

[0036] A collaborative adaptive RTB advertising compliance auditing system and method based on generative tracing includes:

[0037] The data acquisition module collects multimodal advertising creatives in real time from the Supply-Side Platform (SSP), the Demand-Side Platform (DSP), and the Ad Exchange (ADX).

[0038] The SACE module, connected to the data acquisition module, is used to perform generative content tracing, causal anomaly detection, privacy-protected model training and edge deployment, and energy consumption optimization on the multimodal advertising materials.

[0039] The SACE module includes:

[0040] The Generative Source Tracing and Causal Detection Submodule extracts material features using Transformer-based Unified Feature Flow Extraction technology and verifies implicit watermarks and cross-modal fingerprints to confirm the source of the material. At the same time, it constructs a causal relationship graph based on the Structured Causal Model (SCM) to detect abnormal behavior.

[0041] The Federated Learning and Edge Deployment submodule trains the source detection model locally at the edge node using the FedAvg algorithm and differential privacy technology, and only uploads encrypted gradients for aggregation.

[0042] The green optimization submodule uses an LSTM network to predict energy consumption and dynamically allocate resources based on the load.

[0043] In the generative tracing and causal detection submodule, implicit watermark verification is used for source authentication of image and video materials, cross-modal fingerprinting is used to associate text descriptions with visual content for consistency verification and tracing, and the abnormal behaviors include false bidding and CTR spikes.

[0044] The differential privacy technology used in the federated learning and edge deployment submodule has an initial privacy budget ε of 1.0. When the high concurrency scenario exceeds 200,000 records / second, ε is adjusted to 1.5.

[0045] Example 2:

[0046] A collaborative adaptive RTB advertising compliance auditing system and method based on generative tracing includes:

[0047] The data acquisition module collects multimodal advertising creatives in real time from the Supply-Side Platform (SSP), the Demand-Side Platform (DSP), and the Ad Exchange (ADX).

[0048] The SACE module, connected to the data acquisition module, is used to perform generative content tracing, causal anomaly detection, privacy-protected model training and edge deployment, and energy consumption optimization on the multimodal advertising materials.

[0049] The SACE module includes:

[0050] The Generative Source Tracing and Causal Detection Submodule extracts material features using Transformer-based Unified Feature Flow Extraction technology and verifies implicit watermarks and cross-modal fingerprints to confirm the source of the material. At the same time, it constructs a causal relationship graph based on the Structured Causal Model (SCM) to detect abnormal behavior.

[0051] The Federated Learning and Edge Deployment submodule trains the source detection model locally at the edge node using the FedAvg algorithm and differential privacy technology, and only uploads encrypted gradients for aggregation.

[0052] The green optimization submodule uses an LSTM network to predict energy consumption and dynamically allocate resources based on the load.

[0053] Preferably, in the generative tracing and causal detection submodule, implicit watermark verification is used for source authentication of image and video materials, cross-modal fingerprinting is used to associate text descriptions with visual content for consistency verification and tracing, and the abnormal behaviors include false bidding and CTR spikes.

[0054] Preferably, the privacy budget ε of the differential privacy technology used in the federated learning and edge deployment submodule is initially set to 1.1. When the high concurrency scenario is greater than 200,000 records / second, ε is adjusted to 1.6.

[0055] Example 3:

[0056] A collaborative adaptive RTB advertising compliance auditing system and method based on generative tracing includes:

[0057] The data acquisition module collects multimodal advertising creatives in real time from the Supply-Side Platform (SSP), the Demand-Side Platform (DSP), and the Ad Exchange (ADX).

[0058] The SACE module, connected to the data acquisition module, is used to perform generative content tracing, causal anomaly detection, privacy-protected model training and edge deployment, and energy consumption optimization on the multimodal advertising materials.

[0059] The SACE module includes:

[0060] The Generative Source Tracing and Causal Detection Submodule extracts material features using Transformer-based Unified Feature Flow Extraction technology and verifies implicit watermarks and cross-modal fingerprints to confirm the source of the material. At the same time, it constructs a causal relationship graph based on the Structured Causal Model (SCM) to detect abnormal behavior.

[0061] The Federated Learning and Edge Deployment submodule trains the source detection model locally at the edge node using the FedAvg algorithm and differential privacy technology, and only uploads encrypted gradients for aggregation.

[0062] The green optimization submodule uses an LSTM network to predict energy consumption and dynamically allocate resources based on the load.

[0063] Preferably, in the generative tracing and causal detection submodule, implicit watermark verification is used for source authentication of image and video materials, cross-modal fingerprinting is used to associate text descriptions with visual content for consistency verification and tracing, and the abnormal behaviors include false bidding and CTR spikes.

[0064] Preferably, the privacy budget ε of the differential privacy technology used in the federated learning and edge deployment submodule is initially set to 1.2. When the high concurrency scenario is greater than 200,000 records / second, ε is adjusted to 1.8.

[0065] Example 4:

[0066] A collaborative adaptive RTB advertising compliance review method based on generative tracing includes the following steps:

[0067] S1: Data Acquisition: Real-time acquisition of multimodal advertising creatives from the Supply-Side Platform (SSP), Demand-Side Platform (DSP), and Ad Exchange (ADX);

[0068] S2: Generative source tracing and causal detection: Features of the multimodal advertising materials are extracted using Transformer-based unified feature flow extraction technology, and implicit watermarks and cross-modal fingerprints are verified to confirm the source of the materials. At the same time, a causal relationship graph is constructed based on the structured causal model (SCM) to detect abnormal behavior.

[0069] S3: Federated Learning and Edge Verification: At edge nodes, the FedAvg algorithm and differential privacy technology are used to train the source detection model locally, and only encrypted gradients are uploaded for aggregation;

[0070] S4: Green Optimization: Predict energy consumption through LSTM network and dynamically allocate computing resources based on source tracing and detection load, and shield high-energy-consuming abnormal materials;

[0071] S5: Adaptive feedback optimization. The reinforcement learning Q-learning algorithm is used to adjust the source tracing threshold and resource allocation weight based on the output of step S2 to form a closed-loop optimization. Abnormal signals are fed back to the source tracing and green optimization steps.

[0072] In step S1, the multimodal advertising materials include text, images, videos, AR content, and AIGC-generated content.

[0073] In step S2, the model expression for SCM is: P(Y / do(X))=Σ Z P(Y / X, Z)P(Z), where X is the material feature, Y is the anomaly index, and Z is the mediating variable.

[0074] In step S3, the aggregation model is as follows: Where w are the model parameters, n kLet n be the amount of data at node k, and n be the total amount of data.

[0075] In step S4, the LSTM state update model is as follows: Where h t W represents the hidden state at time t. xh and W hh Let x be the weight matrix. t Let b be the input vector at time t. h This is a bias term.

[0076] In step S5, the Q-learning update threshold Q(s, α) = Q(S, α) + α[r + λmax] α′ Q(s′,α′)-Q(s,α)], where s is the state, α is the action, r is the reward function, and λ is the discount factor.

[0077] In step S5, the reward function of the Q-learning algorithm is:

[0078] This invention, with the SACE engine at its core, takes generative content tracing (implicit watermarking and cross-modal fingerprinting) as its starting point, integrates causal anomaly detection (structured causal model SCM), and trains a privacy-preserving model under a federated learning framework; edge computing ensures real-time execution, green optimization dynamically adjusts resources; and an adaptive feedback mechanism connects each link to form a closed-loop optimization, achieving efficient and compliant review of multimodal AIGC materials.

[0079] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A collaborative adaptive RTB advertising compliance review system based on generative traceability, characterized in that, include: The data acquisition module collects multimodal advertising creatives in real time from the Supply-Side Platform (SSP), the Demand-Side Platform (DSP), and the Ad Exchange (ADX). The SACE module, connected to the data acquisition module, is used to perform generative content tracing, causal anomaly detection, privacy-protected model training and edge deployment, and energy consumption optimization on the multimodal advertising materials. The SACE module includes: The Generative Source Tracing and Causal Detection Submodule extracts material features using Transformer-based Unified Feature Flow Extraction technology and verifies implicit watermarks and cross-modal fingerprints to confirm the source of the material. At the same time, it constructs a causal relationship graph based on the Structured Causal Model (SCM) to detect abnormal behavior. The Federated Learning and Edge Deployment submodule trains the source detection model locally at the edge node using the FedAvg algorithm and differential privacy technology, and only uploads encrypted gradients for aggregation. The green optimization submodule uses an LSTM network to predict energy consumption and dynamically allocate resources based on the load.

2. The collaborative adaptive RTB advertising compliance review system based on generative tracing as described in claim 1, characterized in that, In the generative tracing and causal detection submodule, implicit watermark verification is used for source authentication of image and video materials, cross-modal fingerprinting is used to associate text descriptions with visual content for consistency verification and tracing, and the abnormal behaviors include false bidding and CTR spikes.

3. The collaborative adaptive RTB advertising compliance review system based on generative tracing as described in claim 1, characterized in that, The differential privacy technology used in the federated learning and edge deployment submodule has an initial privacy budget ε of 1.0-1.

2. When the high concurrency scenario exceeds 200,000 records / second, ε is adjusted to 1.5-1.

8.

4. The collaborative adaptive RTB advertising compliance review system based on generative tracing as described in claim 1, characterized in that, The differential privacy technology used in the federated learning and edge deployment submodule has an initial privacy budget ε of 1.

0. When the high concurrency scenario exceeds 200,000 records / second, ε is adjusted to 1.

5.

5. A collaborative adaptive RTB advertising compliance review method based on generative traceability, implemented based on the collaborative adaptive RTB advertising compliance review system based on generative traceability as described in any one of claims 1-4, characterized in that, Includes the following steps: S1: Data Acquisition: Real-time acquisition of multimodal advertising creatives from the Supply-Side Platform (SSP), Demand-Side Platform (DSP), and Ad Exchange (ADX); S2: Generative source tracing and causal detection: Features of the multimodal advertising materials are extracted using Transformer-based unified feature flow extraction technology, and implicit watermarks and cross-modal fingerprints are verified to confirm the source of the materials. At the same time, a causal relationship graph is constructed based on the structured causal model (SCM) to detect abnormal behavior. S3: Federated Learning and Edge Verification: At edge nodes, the FedAvg algorithm and differential privacy technology are used to train the source detection model locally, and only encrypted gradients are uploaded for aggregation; S4: Green Optimization: Predict energy consumption through LSTM network and dynamically allocate computing resources based on source tracing and detection load, and shield high-energy-consuming abnormal materials; S5: Adaptive feedback optimization. The reinforcement learning Q-learning algorithm is used to adjust the source tracing threshold and resource allocation weight based on the output of step S2 to form a closed-loop optimization. Abnormal signals are fed back to the source tracing and green optimization steps.

6. The collaborative adaptive RTB advertising compliance review method based on generative tracing as described in claim 5, characterized in that, In step S1, the multimodal advertising materials include text, images, videos, AR content, and AIGC-generated content.

7. The collaborative adaptive RTB advertising compliance review method based on generative tracing as described in claim 5, characterized in that, In step S2, the model expression for SCM is: P(Y / do(X))=∑ Z P(Y / X, Z)P(Z), where X is the material feature, Y is the anomaly index, and Z is the mediating variable.

8. The collaborative adaptive RTB advertising compliance review method based on generative tracing as described in claim 5, characterized in that, In step S3, the aggregation model is as follows: Where w are the model parameters, n k Let n be the amount of data at node k, and n be the total amount of data.

9. A collaborative adaptive RTB advertising compliance review method based on generative tracing as described in claim 5, characterized in that, In step S4, the LSTM state update model is as follows: Where h t W represents the hidden state at time t. xh and W hh Let x be the weight matrix. t Let b be the input vector at time t. h This is a bias term.

10. A collaborative adaptive RTB advertising compliance review method based on generative tracing as described in claim 5, characterized in that, In step S5, the Q-learning update threshold Q(s, α) = Q(S, α) + α[r + λmax] α′ Q(s′,α′)-Q(s,α)], where s is the state, α is the action, r is the reward function, and λ is the discount factor; In step S5, the reward function of the Q-learning algorithm is:

Images