Unified Steganography via Implicit Neural Representation

TL;DR

The paper tackles the lack of generalizability in steganography by introducing Unified Steganography via Implicit Neural Representation (U-INR), which embeds secret data directly into the parameters of an INR. It replaces format-specific encoders/extractors with a private-key–driven implicit consensus that designates which INR weights carry the secret data, enabling cross-modal hiding across images, audio, video, SDFs, and NeRFs. The method uses a stega mask $\mathrm{M}_e$ derived from a private key $k_e$ to select weight positions, and optimizes secret and cover data within the masked parameters, with decryption achievable only by those who possess the key. Experiments across multiple data modalities demonstrate competitive secret/stea representations and robustness to pruning attacks, highlighting potential for secure, data-type-agnostic covert communication without external extraction components.

Abstract

Digital steganography is the practice of concealing for encrypted data transmission. Typically, steganography methods embed secret data into cover data to create stega data that incorporates hidden secret data. However, steganography techniques often require designing specific frameworks for each data type, which restricts their generalizability. In this paper, we present U-INR, a novel method for steganography via Implicit Neural Representation (INR). Rather than using the specific framework for each data format, we directly use the neurons of the INR network to represent the secret data and cover data across different data types. To achieve this idea, a private key is shared between the data sender and receivers. Such a private key can be used to determine the position of secret data in INR networks. To effectively leverage this key, we further introduce a key-based selection strategy that can be used to determine the position within the INRs for data storage. Comprehensive experiments across multiple data types, including images, videos, audio, and SDF and NeRF, demonstrate the generalizability and effectiveness of U-INR, emphasizing its potential for improving data security and privacy in various applications.

Figures (12)

• Figure 1: Illustration of our scenario. (a) Previous steganography approaches require designing specific frameworks for different data formats. Besides, malicious users could exploit the steganography encoder/decoder to expose or corrupt the secret data. (b) Our U-INR can work on various data formats like image, video and others. Besides, U-INR bypasses the need for external encoder/decoder architectures that could expose attack surfaces, ensuring exclusive access to key holders.
• Figure 2: Framework of U-INR. Our architecture establishes secure synchronization between multimedia sender and receiver through an implicit consensus using a shared private key $k_e$. This key precisely maps the weight positions distinguishing secret payloads from cover data in the neural representation. Enforcing consensus-based parameter coordination through the implicit neural network's weight-sharing mechanism eliminates the need for external auxiliary modules.
• Figure 3: Implicit Consensus. The initialized weight values of the INR are used to identify and select the weights for secret data based on the magnitudes.
• Figure 4: Examples of the stega and recovered images generated using different schemes. The left is the original image, and the right represents $\times 5$ magnified residuals. The cover/stega and secret/recovered images are given in the first and last rows. For our U-INR, we use stega ratio $\mathcal{S} = 0.3$ as it balances the quality of cover and secret representation.
• Figure 5: Quantitative and qualitative results of our method when applying to video data. The bike video and cat video are adopted as secret and stega representations. Compared to the normal neural representation, the quality of the stega and secret representations only experiences a slight decrease.