CN-122024026-A - Lightweight non-repudiation model fingerprint efficient tracing method

CN122024026ACN 122024026 ACN122024026 ACN 122024026ACN-122024026-A

Abstract

The invention discloses a lightweight non-repudiation model fingerprint efficient tracing method, and belongs to the related technical fields of cryptography application, digital content tracing and the like. The method comprises the specific steps of inputting an original image data set into a computer system, training a learning lightweight encoder and a decoder, mapping binary fingerprints and images by the encoder to generate residual errors, restraining the residual errors to generate images containing fingerprints through a Sigmoid function, generating metadata of PIT parameter binding salt values and promise values, embedding the fingerprints into the images, extracting decoded fingerprints by a loading decoder during detection, binarizing the decoded fingerprints through a threshold function, and comparing the decoded fingerprints with the promise values through a fingerprint polynomial to realize double verification. The fingerprint verification method is used for converting the fingerprint verification mode into polynomial verification through the polynomial equivalent detection technology, reducing the communication complexity to sub-linear expenditure, realizing light-weight high-efficiency verification, and being suitable for efficient traceability and authenticity verification of deep counterfeit images.

Inventors

CHEN YULING
LUO YUN
DOU HUI
YANG YUXIANG
GUO MINYI

Assignees

贵州大学

Dates

Publication Date: 20260512
Application Date: 20260126

Claims (10)

1. A lightweight non-repudiation model fingerprint efficient tracing method is characterized by comprising the following steps: inputting an original image data set to a computer system, wherein the original image data set comprises an image channel number, an image height and an image width; Sampling binary fingerprints of each batch, training a learning encoder and a corresponding decoder based on the original image dataset, wherein the encoder maps the binary fingerprints with the original image dataset to obtain residual errors; Restricting the residual error in a range of intervals through a hash number Sigmoid function to generate a fingerprint-containing image, and calculating a binary cross entropy loss sum after the decoder extracts fingerprints from the fingerprint-containing image Loss; loading the encoder, the weights and the parameters, generating a fixed fingerprint if the batch images are uniformly marked, otherwise randomly sampling the fingerprint, generating PIT parameters, embedding the fingerprint into the images and generating metadata; The fingerprint is prevented from being tampered by binding the fingerprint with a salt value, and the metadata is stored; Loading the decoder, inputting an image to be detected and the metadata, extracting a decoded fingerprint, and obtaining a binary fingerprint through a threshold function; comparing whether the polynomials of the original fingerprint and the binary fingerprint are consistent, verifying the matching of the promised value, judging whether the embedded fingerprint is tampered, and outputting a matching result and a tampering detection result.
2. A lightweight non-repudiatable model fingerprinting efficient tracing method according to claim 1, wherein said binary fingerprinting specifically comprises generating based on each image in said original image dataset; the training learning encoder and the corresponding encoder specifically comprise that the data set is divided into small batches for training by training round by round, and the data set is disturbed at the beginning of each round to prevent the model from being fitted.
3. The efficient tracing method for lightweight non-repudiation model fingerprints of claim 1, wherein the hash Sigmoid function is: ; Wherein, the Representing a hash Sigmoid function, R representing a residual error, e representing a natural constant; the interval range is An inner part; The calculated binary cross entropy loss is obtained by a binary cross entropy loss function, the Losing constraint on similarity of the fingerprint-containing image to the original image by A loss function is obtained.
4. A lightweight non-repudiatable model fingerprint efficient tracing method according to claim 3, wherein said binary cross entropy loss function is: ; Wherein, the For binary cross entropy loss, j represents the sample index within the batch, B represents the batch size, Representing a fingerprint; the said The loss function is: ; Wherein, the Representation of A loss function, B, represents the batch size, j represents the sample index within the batch, Representing an image containing a fingerprint.
5. The efficient tracing method of lightweight non-repudiatable model fingerprints of claim 1, wherein said fingerprint embedded image comprises the steps of: embedding the fingerprint into the image through residual calculation, and activating and clipping through Sigmoid to ensure that the pixel value of the fingerprint-containing image accords with a range; The PIT parameters comprise prime numbers, random numbers and fingerprint polynomial verification values; The metadata comprises the PIT parameters, salt values and corresponding fingerprint promises, wherein the salt values are obtained by sampling each fingerprint; The fingerprint commitment is calculated based on the corresponding salt value.
6. The efficient tracing method for lightweight non-repudiation model fingerprints of claim 1, wherein the threshold function is: ; The calculation logic of the threshold function is specifically that if the binary fingerprint is The result of the threshold function is 1, if the binary fingerprint The result of the threshold function is 0.
7. The efficient tracing method of a lightweight non-repudiation model fingerprint according to claim 1, wherein said comparing whether the polynomials of the original fingerprint and the binary fingerprint are identical or not, verifying the promise value matching specifically includes a first re-verification and a second re-verification; The first re-verification specifically comprises the steps of comparing polynomials of the original fingerprint and the binary fingerprint, judging that the fingerprints are matched if the polynomials are equal, judging that the fingerprints are not matched if the polynomials are not equal, and activating the second re-verification.
8. The efficient tracing method of lightweight non-repudiation model fingerprints of claim 7, wherein said second re-verifying comprises: when the first re-verification judges that the fingerprints are not matched, opening a promise to verify, calculating a reconstruction promise, reconstructing promise values through a promise function based on the binary fingerprint and the salt value, and integrating to form a promise value set to be verified; And carrying out consistency check on the fingerprint promise in the metadata of the promise value set to be verified, if so, not falsifying the embedded fingerprint after being embedded, and if not, judging that the embedded fingerprint is falsified.
9. An electronic device, comprising: A memory communicatively coupled to the processor; The memory has stored therein a computer program which, when executed by the processor, enables a lightweight non-repudiation model fingerprint efficient tracing method as claimed in any one of claims 1-8.
10. A computer readable storage medium having stored thereon a computer program, which when executed by a processor implements a lightweight non-repudiation model fingerprint efficient tracing method as claimed in any one of claims 1-8.

Description

Lightweight non-repudiation model fingerprint efficient tracing method Technical Field The invention belongs to the related technical fields of cryptography application, digital content tracing and the like, and particularly relates to a lightweight non-repudiation model fingerprint efficient tracing method. Background With the rapid evolution of artificial intelligence technology, the depth generation model provides a core technical support for the maturation and wide application of the generated image technology, especially for the continuous iteration of the generation of a countermeasure network (GAN). The technology can generate vivid image content and has great application value in the fields of film and television production, virtual interaction and the like. However, the application of the technology also promotes the problem of deep forgery detection, false image content permeates into all corners of real scenes such as social media, news propagation and the like, so that strong worry is caused on the controllability of the generated data, and the research on the deep forgery detection and tracing is stimulated. Currently, in the detection and tracing methods for deep counterfeiting, an active sustainable detection method becomes a mainstream scheme because of the advantages of advanced prevention and control and whole course tracking. However, the existing active sustainable detection method still has the defects that on one hand, the communication complexity is higher, on the other hand, the model fingerprint is not counterfeitable, the design is not perfect enough, malicious tampering and counterfeiting attacks are difficult to effectively resist, the reliability and authority of a tracing result are difficult to guarantee, and the strict requirements on tracing of generated data in practical application cannot be fully met. In view of the foregoing, there is a need for an efficient, reliable and practical deep counterfeiting tracing method, which can effectively reduce the complexity of communication while strengthening the non-counterfeitability of model fingerprints on the basis of inheriting the core advantages of the active sustainable detection method, and provides a new scheme for deep counterfeiting tracing. Disclosure of Invention The invention mainly aims to provide a lightweight non-repudiation model fingerprint efficient tracing method, which aims to convert a fingerprint verification mode into polynomial verification through a polynomial equivalent detection technology for a deep counterfeit image tracing scene, reduce communication complexity to sub-linear expense, realize lightweight high-efficiency verification, and reduce communication complexity, strengthen fingerprint non-counterfeitability, and improve the efficiency and reliability of deep counterfeit book tracing by means of technologies such as encoder and decoder collaborative training, PIT parameter and promised value dual verification and the like. Based on the first main aspect of the invention, a lightweight non-repudiation model fingerprint efficient tracing method is provided, which comprises the following steps: inputting an original image data set to a computer system, wherein the original image data set comprises an image channel number, an image height and an image width; Sampling binary fingerprints of each batch, training a learning encoder and a corresponding decoder based on the original image dataset, wherein the encoder maps the binary fingerprints with the original image dataset to obtain residual errors; Restricting the residual error in a range of intervals through a hash number Sigmoid function to generate a fingerprint-containing image, and calculating a binary cross entropy loss sum after the decoder extracts fingerprints from the fingerprint-containing image Loss; loading the encoder, the weights and the parameters, generating a fixed fingerprint if the batch images are uniformly marked, otherwise randomly sampling the fingerprint, generating PIT parameters, embedding the fingerprint into the images and generating metadata; The fingerprint is prevented from being tampered by binding the fingerprint with a salt value, and the metadata is stored; Loading the decoder, inputting an image to be detected and the metadata, extracting a decoded fingerprint, and obtaining a binary fingerprint through a threshold function; comparing whether the polynomials of the original fingerprint and the binary fingerprint are consistent, verifying the matching of the promised value, judging whether the embedded fingerprint is tampered, and outputting a matching result and a tampering detection result. Through the technical scheme, a complete technical process closed loop from training, embedding, detecting and finally verifying is constructed, communication complexity is reduced based on a lightweight model design, fingerprint non-counterfeitability and tamper resistance are enhanced by means of a double verification mec