UTS: Unified Token-Space steganography for multimodal data via fast implicit neural representation

Abstract Steganography aims to conceal secret messages within carrier data without introducing perceptible distortion. Existing methods are typically designed for a single type of carrier (e.g., images) and cannot meet the requirements of the current trend toward multimodal media. Moreover, many deep-learning–based schemes require transmitting a message extractor alongside the carrier for accurate recovery, which significantly undermines steganographic stealthiness. To address these limitations, we propose Unified Token-Space Steganography (UTS), a multimodal steganographic framework based on fast implicit neural representations. Our core idea is to eliminate modality discrepancies through unified representation, enabling multimodal adaptation within a single framework. We employ a fast implicit neural representation method to encode multimodal data into compact spatial vector tokens and embed secret messages under the constraint of a fixed message extractor—whose parameters are initialized with a pre-agreed random seed and then frozen. During communication, the sender and receiver only need to exchange stego tokens to recover the secret message, avoiding the transmission of an additional extractor. Experiments on three modalities (images, weather data, and 3D scenes) demonstrate that UTS achieves 100% message accuracy, with PSNR exceeding 35 dB for stego images and weather data, and high-quality reconstruction for 3D scenes. These results demonstrate that UTS provides an efficient and practical steganographic framework for modern multimodal media.