[Paper Review] A Brief Survey on Deep Learning Based Data Hiding

Chaoning Zhang, Chenguo Lin+4 more|arXiv (Cornell University)|Mar 2, 2021

Digital Media Forensic Detection6 citations

TL;DR

This survey provides a comprehensive overview of deep learning-based data hiding (deep hiding), classifying methods by capacity, security, and robustness while analyzing three core architectures. It reviews applications in image, audio, video, and text steganography and watermarking, and highlights the emerging link between deep hiding and adversarial attacks, revealing shared vulnerabilities in high-frequency signal sensitivity and mutual technique transferability.

ABSTRACT

Data hiding is the art of concealing messages with limited perceptual changes. Recently, deep learning has enriched it from various perspectives with significant progress. In this work, we conduct a brief yet comprehensive review of existing literature for deep learning based data hiding (deep hiding) by first classifying it according to three essential properties (i.e., capacity, security and robustness), and outline three commonly used architectures. Based on this, we summarize specific strategies for different applications of data hiding, including basic hiding, steganography, watermarking and light field messaging. Finally, further insight into deep hiding is provided by incorporating the perspective of adversarial attack.

Motivation & Objective

• To systematically categorize deep learning-based data hiding methods by their core properties: capacity, security, and robustness.
• To identify and analyze three fundamental neural network architectures used in end-to-end deep hiding pipelines.
• To review application-specific strategies in large-capacity data hiding, steganography, watermarking, and light field messaging.
• To examine data hiding in non-image media, including audio, video, and text, with a focus on deep learning integration.
• To explore the emerging intersection between deep hiding and adversarial machine learning, particularly shared vulnerabilities and cross-application techniques.

Proposed method

• Classifies deep hiding methods along three essential dimensions: capacity (payload size), security (resistance to steganalysis), and robustness (resistance to transmission distortions).
• Outlines three primary neural network architectures: (1) preparation network (P), (2) hiding network (H), and (3) reveal network (R), forming an end-to-end pipeline for embedding and extraction.
• Employs a joint optimization objective: minimizing perceptual distortion between cover and container images ($dist_c$) and reconstruction error between original and retrieved secrets ($dist_s$), typically using L2 or cross-entropy loss.
• Applies a weighted loss function $\mathcal{L} = \|C' - C\| + \beta\|S' - S\|$ to balance imperceptibility and accuracy, with $\beta$ controlling trade-offs.
• Reviews multi-stage networks for cross-modal hiding (e.g., image in audio), where residual errors are progressively embedded and reconstructed.
• Integrates insights from adversarial machine learning by analyzing shared sensitivity to high-frequency perturbations, using Fourier analysis to unify understanding of deep hiding and adversarial attacks.

Experimental results

Research questions

• RQ1 How do deep learning-based data hiding methods balance the trade-off between capacity, security, and robustness in image-based information hiding?
• RQ2 What are the three fundamental neural network architectures that enable end-to-end training in deep hiding, and how do they contribute to performance?
• RQ3 How do deep learning models enable effective data hiding in non-image media such as audio, video, and text, and what are the key challenges in these domains?
• RQ4 In what ways do adversarial attacks and deep hiding share underlying vulnerabilities, particularly in high-frequency signal sensitivity?
• RQ5 How can techniques from adversarial machine learning improve the robustness of deep hiding systems, and vice versa?

Key findings

• End-to-end deep learning models for data hiding significantly outperform traditional methods in capacity, security, and robustness, often achieving superior performance through joint optimization of embedding and extraction networks.
• The use of a weighted loss function $\mathcal{L} = \|C' - C\| + \beta\|S' - S\|$ enables effective trade-off management between perceptual quality and secret retrieval accuracy.
• Multi-stage networks for audio-based data hiding successfully embed image content into audio by progressively encoding residual errors, achieving high fidelity in reconstruction.
• Text-based data hiding using models like the Adversarial Watermarking Transformer (AWT) enables unobtrusive message embedding through joint encoder-decoder training and adversarial fine-tuning.
• A unified Fourier perspective reveals that deep neural networks are highly sensitive to high-frequency content, explaining both the success of adversarial attacks and the effectiveness of deep hiding techniques.
• Cross-application techniques between adversarial attacks and watermarking are effective: adversarial robustness methods improve watermark resilience, and watermarking defenses can mitigate model-extraction attacks.