Noise-NeRF: Hide Information in Neural Radiance Fields using Trainable Noise

TL;DR

Noise-NeRF introduces trainable input noise to NeRF for lossless steganography without altering model weights, addressing confidentiality and data security in neural radiance fields. The method leverages an Adaptive Pixel Selection strategy and a Pixel Perturbation strategy to improve both steganography accuracy and convergence efficiency, while maintaining high rendering fidelity. It enables not only standard steganography within NeRF scenes but also effective super-resolution image hiding by distributing hidden content across viewpoints. Experimental results on diverse NeRF scenes and datasets demonstrate state-of-the-art steganography quality and robust rendering performance, with high recovery accuracy and SSIM metrics.

Abstract

Neural Radiance Field (NeRF) has been proposed as an innovative advancement in 3D reconstruction techniques. However, little research has been conducted on the issues of information confidentiality and security to NeRF, such as steganography. Existing NeRF steganography solutions have shortcomings in low steganography quality, model weight damage, and limited amount of steganographic information. This paper proposes Noise-NeRF, a novel NeRF steganography method employing Adaptive Pixel Selection strategy and Pixel Perturbation strategy to improve the quality and efficiency of steganography via trainable noise. Extensive experiments validate the state-of-the-art performances of Noise-NeRF on both steganography quality and rendering quality, as well as effectiveness in super-resolution image steganography.

Figures (6)

• Figure 1: Overview of Noise-NeRF.
• Figure 2: Framework of Noise-NeRF. We first add random initial trainable noise to a specific view, and use pre-trained NeRF for prediction. Then, we perform supervised training on input noise using secret images. We employ Adaptive Pixel Selection strategy and Pixel Perturbation strategy during the training process to improve the quality and efficiency of steganography.
• Figure 3: Noise-NeRF performances on multiple scenes. Each column displays the initial rendering, rendering after 100 loops, rendering after 300 loops, and the residual image. We also show the SSIM between the steganography image rendered by Noise-NeRF and the real hidden image.
• Figure 4: Noise-NeRF performance on super-resolution images. Each column displays the initial rendering, rendering after 100 loops, rendering after 300 loops, and the residual image. We also show the SSIM between the steganography image rendered by Noise-NeRF and the real hidden image.
• Figure 5: More qualitative results of Noise-NeRF on multiple super-resolution results.