Search

EP-4738124-A1 - SYNTHETIC DATA SIGNAL

EP4738124A1EP 4738124 A1EP4738124 A1EP 4738124A1EP-4738124-A1

Abstract

Proposed concepts thus aim to provide schemes, solutions, concepts, designs, methods and systems pertaining to generating a synthetic data signal. In particular, embodiments aim to provide a method for generating a synthetic data signal by first modifying a data signal (to be input into a trained generative network) by adding error-correction information to the data signal. The modified data signal (i.e., the data signal with the added error-correction information) is then input into the trained generative network to generate a synthetic data signal. The added error-correction information allows errors to be corrected during subsequent reconstruction of the data signal from the synthetic data signal.

Inventors

  • VAN DEN DUNGEN, WILHELMUS ANDREAS MARINUS ARNOLDUS MARIA
  • GARCIA MORCHON, OSCAR
  • VAN ACHT, VICTOR MARTINUS GERARDUS

Assignees

  • Koninklijke Philips N.V.

Dates

Publication Date
20260506
Application Date
20241105

Claims (15)

  1. A computer-implemented method (100) for generating a synthetic data signal, the method comprising: adding error-correction information (110) to a data signal to generate a modified data signal; and inputting the modified data signal into a forward generator of a trained bidirectional generative network (120) to generate a first synthetic data signal, wherein the forward generator is trained to generate a synthetic data signal based on an input data signal.
  2. The computer-implemented method of claim 1, further comprising: generating the error-correction information (202) by processing the data signal with at least one of: a forward error correction polar algorithm; a Reed-Solomon algorithm; a Golay algorithm; a BCH algorithm; a multidimensional parity algorithm; a Hamming algorithm; and a convolutional algorithm.
  3. The computer-implemented method of claim 1 or 2, wherein the method further comprises encrypting the data signal (204) prior to adding the error-correction information.
  4. The computer-implemented method of claim 3, wherein adding the error-correction information to the data signal comprises: structuring the encrypted data signal (206) in a first domain; and adding error-correction information to the structured encrypted data signal (210) to generate a modified data signal.
  5. The computer-implemented method of any of claims 1 to 4, wherein the method further comprises adding marking data to the modified data signal (215) such that the first synthetic data signal comprises an indicator corresponding to the marking data.
  6. The computer-implemented method of claim 5, wherein the method further comprises adding second error-correction information to the marking data to generate modified marking data, and wherein the modified marking data is added to the modified data signal such that the first synthetic data signal comprises an indicator corresponding to the marking data.
  7. The computer-implemented method of claim 5 or 6, wherein the method further comprises encrypting the marking data.
  8. A computer-implemented method (300) for reconstructing a data signal, the method comprising: inputting a first synthetic data signal into a backward generator of a trained bidirectional generative network (310) to generate a modified data signal, wherein the backward generator is trained to generate a data signal based on an input synthetic data signal; and processing the modified data signal with an error-correction mechanism (320) to generate a predicted data signal; wherein the first synthetic data signal comprises additional information based on error-correction information, and wherein the predicted data signal is a prediction of the data signal.
  9. The computer-implemented method of claim 8, wherein the method further comprises: decrypting the predicted data signal (430) to generate a decrypted predicted data signal.
  10. The computer-implemented method of claim 8 or 9, wherein the synthetic data signal comprises an indicator corresponding to marking data, and the method further comprises extracting the indicator from the synthetic data signal (440) and inputting the indicator into the backward generator (450) to generate the marking data.
  11. The computer-implemented method of claim 8 or 9, wherein the synthetic data signal comprises an indicator corresponding to marking data, and the method further comprises: extracting the indicator from the synthetic data signal; inputting the indicator into the backward generator to generate modified marking data; and processing the modified marking data with a second error-correction mechanism to generate predicted marking data, wherein the predicted marking data is a prediction of the marking data.
  12. A computer-implemented method (500) for training a bidirectional generative network, the method comprising: adding error-correction information to a training data signal (510) to generate a first modified training data signal; inputting the first modified training data signal into a forward generator of a bidirectional generative network (520) to generate a synthetic training data signal; inputting the synthetic training data signal into a backward generator of the bidirectional generative network (530) to generate a second modified training data signal; processing the second modified training data signal with an error-correction mechanism (540) to generate a predicted training data signal; determining an error-correction quality indicator (550) based on error detection and correction levels of the error-correction mechanism during processing of the second modified training data signal; and further training the bidirectional generative network (560) based on the determined error-correction quality indicator.
  13. The computer-implemented method of claim 12, wherein the error-correction quality indicator comprises a score between 0% and 100%.
  14. A computer program comprising code means for implementing the computer-implemented method of any preceding claim when said program is run on a processing system.
  15. A system (600) for generating a synthetic data signal, the system comprising: a processor (620) configured to: add error-correction information to a data signal to generate a modified data signal; and input the modified data signal into a forward generator of a trained bidirectional generative network to generate a first synthetic data signal, wherein the forward generator is trained to generate a synthetic data signal based on an input data signal.

Description

FIELD OF THE INVENTION This invention relates to the field of generating synthetic data. BACKGROUND OF THE INVENTION Bidirectional generative networks can be used for a variety of purposes, for example, to generate pictures from a command prompt, alter images, alter audio, etc. For example, a bidirectional generative network can create a synthetic output image based on latent space input pixels (which are noise-like) combined with a command prompt. If the latent input data is different, then for the same command prompt, the network will generate a different synthetic output image. Bidirectional generative networks can also be trained to recover the latent input data from the synthetic output image, i.e., bidirectional generative networks can be reversible. Large amounts of effort and resources are being spent to improve bidirectional generative networks. Currently, state-of-the-art of bidirectional generative networks provide around 90% accuracy, and therefore lack full traceability and reversibility. In other words, when trying to recover the latent input data, 10% of the reconstructed latent input data is wrong. This is sufficient for some purposes, for example, deep fake technology, but not sufficient for more demanding applications such as those requiring fully traceable reconstruction. For example, fully traceable reconstruction is needed for medical image processing, denoising, super resolution, steganography, or encrypted steganography. SUMMARY OF THE INVENTION The invention is defined by the claims. According to examples in accordance with an aspect of the invention, there is provided a computer-implemented method for generating a synthetic data signal. The method comprises: adding error-correction information to a data signal to generate a modified data signal; and inputting the modified data signal into a forward generator of a trained bidirectional generative network to generate a first synthetic data signal, wherein the forward generator is trained to generate a synthetic data signal based on an input data signal. Proposed concepts thus aim to provide schemes, solutions, concepts, designs, methods and systems pertaining to generating a synthetic data signal. In particular, embodiments aim to provide a method for generating a synthetic data signal by first modifying a data signal (to be input into a trained generative network) by adding error-correction information to the data signal. The modified data signal (i.e., the data signal with the added error-correction information) is then input into the trained generative network to generate a synthetic data signal. In other words, it is proposed that by adding error-correction information to a data signal to be input into a trained bidirectional generative network (e.g., adding error-correction information to latent space data), a synthetic data signal is generated with improved reversibility, such that a higher percentage (even 100%) of the original data signal can be reconstructed/recovered (e.g., using a backward generator of the trained bidirectional generative network). The added error-correction information allows errors to be corrected during subsequent reconstruction of the data signal from the synthetic data signal, for example, using an error-correction mechanism. For example, error-correction code can be added to latent space data before it enters a generator of a bidirectional generative network (such as a BiGAN). This thus produces a synthetic data signal which, when input into a backwards path of the same bidirectional generative network, produces an output data signal whose errors are able to be corrected such that a higher percentage (up to 100%) of the original data signal can be reconstructed. Also provided is a corresponding method for training a bidirectional generative network to facilitate this enhanced reversibility. For example, the training of the bidirectional generative network is modified (compared to traditional methods) in order to facilitate the bidirectional generative network being trained to enhance the correctability of any errors made during reconstruction of a data signal from a synthetic data signal. Ultimately, an improved method for generating a synthetic data signal. In some embodiments, the method may further comprise generating the error-correction information by processing the data signal with at least one of: a forward error correction polar algorithm; a Reed-Solomon algorithm; a Golay algorithm; a BCH algorithm; a multidimensional parity algorithm; a Hamming algorithm; and a convolutional algorithm. These may all be suitable algorithms for generating effective error-correction information to add to the data signal. In some embodiments, the method may further comprise encrypting the data signal prior to adding the error-correction information. This may enhance the security of the reconstructable data within the synthetic data signal. In some embodiments, adding the error-correction information to the data