FADG：频域感知的自适应中间域生成方法用于载体源失配的隐写分析

FADG: Frequency-Aware Adaptive Domain Generation for Cover Source Mismatch in Steganalysis

摘要: 随着数字媒体的广泛应用，信息隐藏技术在保密通信中扮演着重要角色，但同时也被恶意利用进行非法信息传播，对网络安全构成严重威胁。隐写分析作为检测和识别隐蔽信息的关键技术，在维护信息安全方面具有重要价值。近年来，深度学习的引入显著提升了隐写分析的检测性能，众多基于卷积神经网络的隐写分析模型取得了优异成果。然而，这些隐写分析模型面临一个严峻的挑战：当训练数据(源域)与实际检测数据(目标域)来源不一致时，模型性能急剧下降，这一现象被称为载体源失配(Cover Source Mismatch, CSM)问题。CSM问题严重制约了深度隐写分析模型的实用化进程。本文围绕CSM问题展开研究，提出了一种有效的跨域隐写分析方法，频域感知自适应中间域生成方法(Frequency-Aware Adaptive Domain Generation, FADG)。FADG通过光谱残差理论识别图像中频域能量波动大的区域，进而计算源域和目标域图像的频域隐写嵌入概率图，接着利用源域和目标域的嵌入概率图得到混合权重图用于生成中间域，从而实现从源域到目标域的平滑过渡。通过这种自适应生成中间域的方法，从而有效缓解了跨域隐写检测性能下降的问题。

Abstract: With the widespread application of digital media, information hiding technology plays an important role in secure communication, but it is also maliciously exploited for illegal information dissemination, posing serious threats to network security. Steganalysis, as a key technology for detecting and identifying covert information, holds significant value in maintaining information security. In recent years, the introduction of deep learning has significantly improved the detection performance of steganalysis, with numerous steganalysis models based on convolutional neural networks achieving excellent results. However, these steganalysis models face a severe challenge: when the training data (source domain) and the actual detection data (target domain) originate from different sources, model performance drops sharply, a phenomenon known as the Cover Source Mismatch (CSM) problem. The CSM problem severely restricts the practical application of deep steganalysis models. This paper focuses on the CSM problem and proposes an effective cross-domain steganalysis method, Frequency-Aware Adaptive Domain Generation (FADG). FADG identifies regions with large frequency-domain energy fluctuations in images through spectral residual theory, then calculates frequency-domain steganographic embedding probability maps for both source and target domain images. Subsequently, it utilizes the embedding probability maps of the source and target domains to obtain blending weight maps for generating intermediate domains, thereby achieving a smooth transition from the source domain to the target domain. Through this adaptive intermediate domain generation method, the problem of performance degradation in cross-domain steganalysis detection is effectively alleviated.