Open Access

ARTICLE

Multimodal Implicit Representation Steganography Based on Point Cloud Representation

Yuwei Lu^1,2, Jia Liu^1,2,*, Qiya Wang^1,2, Yujie Liu^1,2, Peng Luo^1,2

1 College of Cryptography Engineering, Engineering University of PAP, Xi’an, China
2 Key Laboratory of Network and Information Security of PAP, Engineering University of PAP, Xi’an, China

* Corresponding Author: Jia Liu. Email:

Computers, Materials & Continua 2026, 89(1), 40 https://doi.org/10.32604/cmc.2026.084506

Received 23 April 2026; Accepted 22 June 2026; Issue published 13 August 2026

Abstract

Existing deep-learning-based steganography methods are typically designed for single-modality cover data and often rely on modality-specific network structures, which limits their cross-modal adaptability. To address this limitation, this paper proposes a multimodal implicit neural representation (INR) steganographic framework based on a point-cloud intermediate representation. The framework first fits the cover data as a carrier INR and samples the fitted carrier into a noisy point cloud. A pre-shared noise seed and secret key are then used to reproduce the carrier-derived point cloud and select a key-dependent point subset as the secret point cloud. Finally, a separate extractor, which is architecturally independent of the carrier INR, is trained to reconstruct the secret image from the secret point cloud. Instead of being treated as a decoder attached to the carrier network, the extractor can be encapsulated as a submodule within another neural network for delivery. On the receiver side, the secret image can be recovered only when the carrier INR, noise seed, secret key, and corresponding extractor are jointly available. Experimental results show that the reconstructed secret images achieve peak signal-to-noise ratio (PSNR) values above 40 dB on the CelebFaces Attributes-High Quality (CelebA-HQ), Common Objects in Context (COCO), and DIVerse 2K resolution (DIV2K) datasets.

---

Keywords

---