CN-121982454-A - AIGC detection model continuous evolution method and system based on anti-reflux

CN121982454ACN 121982454 ACN121982454 ACN 121982454ACN-121982454-A

Abstract

The invention provides a method and a system for continuously evolving a AIGC detection model based on anti-reflux, wherein the method comprises the steps of screening a high-confidence erroneous judgment sample and a low-confidence fuzzy sample from a detection result according to prediction confidence coefficient through a pre-constructed reflux channel, enabling the high-confidence erroneous judgment sample and the low-confidence fuzzy sample to be refluxed to a sample library, respectively classifying the high-confidence erroneous judgment sample and the low-confidence fuzzy sample into a real image sample set or a generated image sample set according to a prediction label, selecting the real image sample and the generated image sample from the sample library, extracting a target area through semantic segmentation, performing cross-sample recombination, generating a mixed enhancement sample with local semantic inconsistency, adding anti-disturbance to the mixed enhancement sample, generating an anti-disturbance sample, injecting the anti-disturbance sample into a dynamic anti-sample library, continuously learning based on the dynamic anti-sample library, and performing iterative optimization on a AIGC detection model, so that the AIGC detection model continuously evolves in the detection process.

Inventors

Sha Yuzhe
LIU RUI
Xing Zijuan
CHEN XIANYI
ZHOU YUTING
Wen Shihang

Assignees

南京信息工程大学

Dates

Publication Date: 20260505
Application Date: 20260330

Claims (10)

1. The AIGC detection model continuous evolution method based on backflow resistance is characterized by comprising the following steps of: AIGC the detection model inputs a sample to be detected, and outputs a detection result, wherein the detection result comprises a real image or a prediction label of a generated image and a corresponding prediction confidence; The pre-constructed backflow channel screens out high-confidence erroneous judgment samples and low-confidence fuzzy samples from detection results according to the prediction confidence coefficient, and the high-confidence erroneous judgment samples and the low-confidence fuzzy samples are used as valuable samples to be backflow to a sample library, and are respectively classified into a real image sample set or an image sample set is generated according to a prediction label; Selecting a real image sample from a sample library, generating an image sample, extracting an object region through semantic segmentation, and performing cross-sample recombination to generate a mixed enhancement sample with local semantic inconsistency; Adding opposite disturbance to the mixed enhanced sample to generate an opposite sample, and injecting the opposite sample into a dynamic opposite sample library; And (3) based on continuous learning of the dynamic challenge sample library, carrying out iterative optimization on the AIGC detection model, so that the AIGC detection model continuously evolves in the detection process.
2. The method for sustained evolution of a model based on AIGC detection against reflow of claim 1, wherein the value samples are: ; ; Wherein, the Representing a high confidence misjudgment sample set, Representing a sample to be tested and, Representing the detection model versus the sample Is used to determine the degree of confidence in the prediction, Representing a preset high confidence threshold Representing the predicted tag(s), Representing the real label confirmed by external feedback verification; ; Wherein, the Represents a low confidence fuzzy sample set, delta represents a preset low confidence threshold, And Respectively represent the first And (b) Sample by each detection module Is determined by the determination result of (2).
3. The method of claim 1, wherein extracting the object region by semantic segmentation and performing cross-sample recombination to generate the hybrid enhanced samples with local semantic inconsistencies comprises: Let the real image sample be Its object mask is The extraction of the object region is expressed as: ; Wherein, the Representing element-by-element multiplication operations; let the generated image sample be Its object mask is The extraction of the background region is expressed as: ; Wherein, the Representing the same full size as the mask A matrix; The hybrid enhanced samples are expressed as: 。
4. The method for generating a challenge sample by adding a challenge disturbance to the mixed enhanced sample according to claim 3, wherein the method comprises iteratively generating a challenge sample by a projection gradient descent method: ; Wherein, the Representing a challenge sample, A hybrid enhanced sample is represented and is shown, The projection operation is represented by a number of steps, Expressed in terms of Is a center, Is of radius A norm sphere is provided with a plurality of points, The iteration step size is represented as, The number of iterations is indicated and, Representing a loss function For input image Is used for the gradient of (a), The loss function is represented by a function of the loss, The target tag is represented by a number of tags, Representing a sign function; ; Wherein, the Indicating that the disturbance is counteracted, Representing a pixel value clipping operation for limiting the pixel value to a valid range of 0 to 1.
5. The method for sustained evolution of a model based on AIGC detection against reflow of claim 1, wherein the method for managing a dynamic challenge sample library comprises: constructing a layered storage structure comprising an original sample layer, a mixed enhancement layer and an anti-disturbance layer; calculating the challenge score of the sample; determining a sample warehousing priority based on the challenge score; sample deduplication is performed based on the feature similarity; when the sample pool capacity reaches the upper limit, sample elimination is performed based on timeliness and antagonism effects.
6. The method of claim 5, wherein the method of calculating challenge scores for the samples comprises: Calculating the probability of correct detection of the sample according to the detection model: ; Wherein, the Representing a challenge score of the challenge, The probability of correct detection of the detection model is represented, and the higher the challenge score is, the harder the sample is correctly detected, and the higher the warehousing priority is; the feature similarity is calculated through cosine similarity: ; Wherein the method comprises the steps of The feature vector representing the new sample is presented, Representing the eigenvectors of the nth existing sample in the sample library, & representing the vector dot product, Representing a vector norm; And when the feature similarity exceeds a preset similarity threshold, determining that the sample is a repeated sample, and removing the repeated sample.
7. The method for sustained evolution of a model based on AIGC detection against reflow of claim 4, wherein the method for sustained learning based on dynamic challenge sample library comprises: Monitoring the number of newly added samples of the dynamic challenge sample library and the performance of a verification set of the detection model; triggering a continuous learning process when the number of the newly added samples reaches a preset number threshold or the performance of the verification set is reduced to exceed a preset performance threshold; sampling challenge samples from a dynamic challenge sample library; mixing the countermeasure sample with the original training data to construct a fine tuning training set; and fine-tuning the detection model based on the fine-tuning training set.
8. The method of claim 7, wherein the fine training set is expressed as: ; Wherein, the Represents the fine-tuning training set and, Representing the number of samples sampled from the dynamic challenge sample library; Representing the number of samples of the raw training data; the fine adjustment of the detection model is realized through incremental learning: ; Wherein, the Representing the parameters of the model after fine tuning, Representing the parameters of the original model and, The fine-tuning learning rate is indicated, Representing the gradient of the model parameters, Representing a loss function; The elastic weight consolidation method is adopted to prevent forgetting: ; Wherein, the Representing the loss of consolidation of the elastic weights, Indicating the loss of a new sample, The regularization coefficient is represented as a function of the regularization coefficient, Representation of Information matrix of the first kind The number of diagonal elements is one, Representing the parameters of the current model, Representing the original model parameters.
9. The method for sustained evolution of a AIGC detection model based on anti-reflux according to claim 1, wherein during the detection of AIGC detection model: If it is Directly outputting a detection result; If it is Fusing the multi-module result to output a final detection result; Wherein, the Representing the detection model versus the sample Is used to determine the degree of confidence in the prediction, A first preset threshold value is indicated, Represents a second preset threshold value, and ; The fusion of the multi-module results employs a weighted voting mechanism: ; Wherein, the The result of the final test is indicated, 、、、 Respectively representing weight coefficients of AIGC detection models, a depth consistency analysis module, a semantic discrimination module and a metadata evidence obtaining module, 、、、 Respectively representing AIGC the judging results of the detection model, the depth consistency analysis module, the semantic judging module and the metadata evidence obtaining module, The classification decision function is represented, and when the input value is greater than the threshold value, the image is decided to be generated, otherwise, the image is decided to be a real image.
10. A AIGC detection model continuous evolution system based on anti-reflux, which is characterized by comprising a AIGC detection model and a reflux channel; AIGC a detection model, which is used for inputting a sample to be detected and outputting a detection result, wherein the detection result comprises a real image or a prediction label of a generated image and a corresponding prediction confidence; The return passage includes: The value sample mining module is used for screening out high-confidence erroneous judgment samples and low-confidence fuzzy samples from the detection results according to the prediction confidence, and taking the high-confidence erroneous judgment samples and the low-confidence fuzzy samples as value samples to flow back to the sample library, and respectively classifying the value samples into a real image sample set or generating an image sample set according to the prediction labels; The semantic segmentation module is used for selecting a real image sample from the sample library and generating an image sample, and extracting an object region through semantic segmentation; The cross-sample semantic recombination module is used for carrying out cross-sample recombination on the extracted object region to generate a mixed enhancement sample with local semantic inconsistency; The disturbance countermeasure generation module is used for adding disturbance countermeasures to the mixed enhancement sample to generate a disturbance countermeasure sample; a dynamic challenge sample library for injecting challenge samples; and the continuous learning training module is used for continuously learning based on the dynamic countermeasure sample library, and performing iterative optimization on the AIGC detection model so as to continuously evolve the AIGC detection model in the detection process.

Description

AIGC detection model continuous evolution method and system based on anti-reflux Technical Field The invention relates to a AIGC detection model continuous evolution method and system based on anti-backflow, and belongs to the technical field of artificial intelligence safety. Background With the rapid development of the generation type artificial intelligence technology, the image generation method based on the deep learning technology such as the generation countermeasure network, the diffusion model and the like can synthesize the highly realistic image. However, these techniques are also used maliciously for illicit activities such as false information dissemination, identity falsification, etc., which pose a threat to social security. Therefore, AIGC detection technology becomes an important means for guaranteeing the authenticity of digital content. The existing AIGC detection technology mainly has the following defects: First, the existing detection system adopts a static model architecture, the detection model is fixed once training is completed, and the detection model cannot adapt to the rapid iteration of the generation technology. When the novel generation model appears, the detection performance of the existing detector is drastically reduced, the training data and the training model need to be collected again, the response period is long, and the maintenance cost is high. Second, existing detection systems lack a robust guarantee mechanism against attacks. An attacker can bypass detection by adding an anti-disturbance, but the existing system cannot learn and evolve from an attack sample, and is difficult to cope with a continuously-upgraded anti-attack means. Third, existing detection systems lack active mining mechanisms for value samples. The difficult sample that runs into in the actual testing process can not be effectively used for model improvement, causes the waste of data resource. Fourth, the sample enhancement means of the existing detection system is single, mainly adopts traditional data enhancement methods such as geometric transformation and color dithering, lacks semantic enhancement strategies aiming at AIGC detection scenes, and is difficult to generate challenging training samples. The root of the technical problems is that the existing detection system lacks a closed-loop evolution mechanism driven by a detection result, and the continuous improvement of the detection capability cannot be realized. Disclosure of Invention The invention provides a AIGC detection model continuous evolution method and system based on anti-reflux, which solve the problems disclosed in the background technology. In order to solve the technical problems, the invention adopts the following technical scheme: The method for continuously evolving based on AIGC detection models against reflux: AIGC the detection model inputs a sample to be detected, and outputs a detection result, wherein the detection result comprises a real image or a prediction label of a generated image and a corresponding prediction confidence; The pre-constructed backflow channel screens out high-confidence erroneous judgment samples and low-confidence fuzzy samples from detection results according to the prediction confidence coefficient, and the high-confidence erroneous judgment samples and the low-confidence fuzzy samples are used as valuable samples to be backflow to a sample library, and are respectively classified into a real image sample set or an image sample set is generated according to a prediction label; Selecting a real image sample from a sample library, generating an image sample, extracting an object region through semantic segmentation, and performing cross-sample recombination to generate a mixed enhancement sample with local semantic inconsistency; Adding opposite disturbance to the mixed enhanced sample to generate an opposite sample, and injecting the opposite sample into a dynamic opposite sample library; And (3) based on continuous learning of the dynamic challenge sample library, carrying out iterative optimization on the AIGC detection model, so that the AIGC detection model continuously evolves in the detection process. Further, the value samples are:; ; Wherein, the Representing a high confidence misjudgment sample set,Representing a sample to be tested and,Representing the detection model versus the sampleIs used to determine the degree of confidence in the prediction,Representing a preset high confidence thresholdRepresenting the predicted tag(s),Representing the real label confirmed by external feedback verification; ; Wherein, the Represents a low confidence fuzzy sample set, delta represents a preset low confidence threshold,AndRespectively represent the firstAnd (b)Sample by each detection moduleIs determined by the determination result of (2). Further, the method for generating the mixed enhancement sample with local semantic inconsistency by extracting the object region through semantic segm