Steganography Encoding for Security Documents Using GANs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing steganography methods are not resistant to physical medium alterations or lossy physical transmission, leading to significant loss of hidden information when images are printed or displayed, and they require computational resources beyond the capacity of mobile devices, making them unsuitable for offline document integrity validation.
Innovation Solution
Utilizing adversarial generative neural networks to encode and decode images with minimal changes, incorporating a training subsystem that includes a generator and discriminative generative neural network to simulate noise and maintain image integrity during physical transmission, suitable for mobile devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If simple steganography methods (replacing least significant bits) are used to encode hidden information in images, then encoding and decoding is computationally simple, but the method is not resistant to noise or image manipulations such as compression and physical transmission
Solution Approach 1:
The patent applies preliminary action by pre-training generative adversarial networks with simulated physical transmission noise and compression artifacts during the encoding phase. This allows the steganography system to anticipate and compensate for distortions that will occur during physical transmission, enabling robust decoding even after JPEG compression or printing processes
Solution Approach 2:
The patent changes the parameter space by transitioning from simple bit-replacement steganography to deep learning-based generative models. The system learns optimal encoding parameters through adversarial training, adjusting how hidden information is embedded in the image to maintain robustness against various transformations while preserving image quality
2Reliability
If advanced steganography methods are used to improve security and resistance to manipulations, then reliability and security are improved, but computational resources required exceed the capacity of mobile devices
Solution Approach 1:
The patent performs computationally intensive training operations in advance during the encoding phase, where generative adversarial networks are pre-trained with simulated physical transmission noise. Once trained, the encoding and decoding operations on mobile devices require minimal computational resources, as the heavy lifting was done beforehand during system initialization
Solution Approach 2:
The patent replaces complex real-time computational mechanisms with pre-computed generative models. Instead of performing heavy adversarial training during decoding on mobile devices, the system uses pre-trained generators and discriminators that can operate efficiently on mobile hardware with minimal energy consumption
3Reliability
If hidden information is encoded in images for security verification, then document integrity verification is improved, but the hidden information is lost when images are transmitted through physical media such as printing or displaying
Solution Approach 1:
The patent applies preliminary action by pre-training the generative adversarial network with simulated physical transmission noise and compression artifacts during the encoding phase. This allows the steganography system to anticipate and compensate for distortions that will occur during printing, displaying, or JPEG compression, enabling successful decoding after physical transmission
Solution Approach 2:
The patent implements beforehand cushioning by incorporating noise simulation and adversarial training during the encoding phase to protect against future physical transmission losses. The generative model learns to embed hidden information in a way that is resilient to anticipated distortions, cushioning the hidden data against information loss during subsequent physical transmission
Data Source
Figure 1~2
Figure 3
AI summary
The invention comprises an encoding system for a security document with a printer-proof steganography-encoded image and a decoding and integrity validation system for a security document with a printer-proof steganography-encoded image, wherein the said systems operate based on a generator generative adversarial neural network and a discriminative generative adversarial neural network. The invention comprises an encoding method, a decoding method, security documents, computing devices, computer programs and reading means (scanner) by an associated computing device. The invention applies to image encoding in general, and is particularly useful for concealing a secret message in facial images, also called portraits, in the context of security documents with a printer-proof steganography-encoded image-such as civil identification documents and personal machine-readable documents.