Abstract
In deep steganography, the model size is usually related to the underlying mesh resolution, and a separate neural network needs to be trained as a message extractor. In this paper, we propose a generative image steganography based on point cloud representation, which represents image data as a point cloud, learns the distribution of the point cloud data, and represents it in the form of a continuous function. This method breaks through the limitation of the image resolution, and can generate images with arbitrary resolution according to the actual need, and omits the need for explicit data for image steganography. At the same time, using a fixed point cloud extractor transfers the training of the network to the point cloud data, which saves the training time and avoids the risk of exposing the steganography behavior caused by the transmission of the message extractor. Experiments prove that the steganographic images generated by the scheme have very high image quality and the accuracy of message extraction reaches more than 99%.
Introduction
Digital steganography is the art of concealing secret information within a non-secret carrier, such as an image. While traditional deep learning-based methods have made significant strides, they often rely on fixed pixel grids. This reliance introduces several constraints regarding resolution flexibility and security.
This research introduces a novel paradigm: Generative Point Cloud Steganography. By shifting from a discrete pixel-grid representation to a continuous point cloud representation, the proposed method achieves resolution independence and enhanced security by eliminating the need to transmit specialized extraction models.
The Challenges of Traditional Steganography
Current deep learning-based steganography typically encounters three primary obstacles:
- Resolution Rigidity: The capacity and robustness of the hidden message are often strictly tied to the image resolution. Changing resolutions can break the embedded message.
- The Extractor Dilemma: Senders usually must transmit a specific neural network (the extractor) to the receiver. This transmission is not only bandwidth-intensive but also serves as a potential indicator of steganographic activity to adversaries.
- Quality Degradation: Embedding data into pixels often introduces detectable noise, reducing the visual quality and making the carrier susceptible to steganalysis tools.
Proposed Methodology: Breaking the Grid
The core of this approach lies in treating an image as a continuous function $f(x, y) = (r, g, b)$ rather than a static grid. This is achieved through Implicit Neural Representation (INR) and Generative Adversarial Stochastic Processes (GASP).
1. Image as a Function
By modeling the image as a function, the system can sample the data at any desired resolution. This allows for "super-resolution" steganography where the secret message remains intact regardless of whether the final output is 64x64 or 256x256.
2. GASP Generator
A noise vector $z$ is processed by a "Function Generator" to produce a unique point cloud. This generator learns the underlying distribution of the data, ensuring the generated carriers look natural and authentic.
3. Fixed Extractor and Shared Seeds
To solve the "extractor delivery" problem, the authors utilize a fixed point cloud extractor. Both the sender and receiver share a random seed to initialize identical network architectures. Instead of training the extractor to find the message, the sender uses the L-BFGS optimization algorithm to slightly adjust the point cloud data so that the fixed extractor correctly decodes the secret message. This shifts the computational burden to the data generation phase rather than the model transmission phase.
Experimental Evaluation
The framework was evaluated using two distinct datasets to prove its versatility:
- CelebA: A large-scale face attributes dataset (28,000 images) for standard multimedia steganography.
- ERA5: A global climate reanalysis dataset. Using weather data demonstrates that the point cloud method can be extended to non-standard and scientific data formats.
Key Results
The point cloud-based method demonstrated superior performance compared to existing state-of-the-art models like StegaGAN and FNNS:
- Exceptional Image Quality: Achieved a Peak Signal-to-Noise Ratio (PSNR) of 66.99 dB, significantly higher than the 25-39 dB range typical of previous methods.
- Structural Integrity: Maintained a Structural Similarity Index (SSIM) of 0.99, ensuring the steganographic carrier is virtually indistinguishable from the original.
- Robust Extraction: Message extraction accuracy exceeded 99%.
- Efficiency: Point cloud generation remained constant at approximately 1.4 seconds, proving the generator is not bottlenecked by the target image resolution.
Conclusion and Future Work
By moving steganography into the domain of continuous functions and point clouds, this research overcomes traditional trade-offs between resolution, security, and image quality. While the method achieves high accuracy, future work will focus on reaching a 0% Bit Error Rate (BER) and extending these principles to 3D models and other complex data structures.
Summary updated