Open Access

ARTICLE

Functa Hiding: Steganography via Modulated Implicit Representations

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

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

* Corresponding Author: Jia Liu. Email:

Computers, Materials & Continua 2026, 88(3), 63 https://doi.org/10.32604/cmc.2026.081589

Received 05 March 2026; Accepted 09 June 2026; Issue published 23 July 2026

Abstract

Implicit Neural Representation (INR) is a technique that models continuous signals using neural networks, replacing traditional discrete grid representations with a coordinate-to-value mapping function. As a data carrier, INR is gradually being adopted as the target for steganographic processing. However, existing INR-based steganographic schemes typically require modifying network structures (e.g., weights, nodes) and retraining to obtain stego INRs, leading to high time consumption and the need for re-training when replacing cover images. To address this issue, this paper proposes StegaMIR (Steganography via Modulated Implicit Representations), an image steganographic scheme based on modulated implicit representations. It adopts the Functa (Fast Implicit Neural Representation Framework) framework for fast data implicit representation, which represents data as a base network and a corresponding modulation vector. To embed messages, a fixed message extractor is introduced to constrain the image reconstructed by the modulation vector and the base network, yielding stego modulation, which is then combined with the base network to construct the stego INR. The sender samples a stego image from it for transmission, while the receiver recovers the secret message using the fixed message extractor, thereby completing covert communication. Experiments demonstrate that StegaMIR can efficiently construct an INR-based steganographic model for images, achieving a message extraction accuracy of over 97% with an average embedding time of 8 s per image, representing a 95.6% reduction compared to existing methods.

---

Keywords

---