SaTC: CORE: Medium: Analytic versus Data Driven Models in Steganography: Empowering Each Other

隐写术是一种秘密通信方法，它通过将实际信息隐藏在其他无害的物体中来实现隐私。数字媒体，如图像，是理想的，因为它们可以被稍微修改以编码秘密，而不会使这些修改对人类可见或被计算机检测到。因此，隐写术为那些审查通信渠道或禁止使用加密的国家的公民提供了隐私。现代隐写术及其使用的检测严重依赖于机器学习和人工智能。虽然这些“黑匣子”系统提供了上级的性能，现有的分析方法，相对知之甚少，这些工具的性能限制，甚至更少关于如何以及为什么这些系统的工作。该项目旨在通过将传统的分析建模工具与现代机器学习相结合来解决这些缺陷，以便深入了解现代人工智能系统如何在检测隐写术中工作，建立其限制，并发现有关数字媒体中隐写术的新基础知识。该项目的核心思想是将传统分析建模工具的解释能力与机器学习的灵活性联合收割机结合起来，利用它们的互补优势来解决该领域中一些最基本的开放问题。人工但真实的图像数据集将被用来计算限制和评估现代数据驱动的隐写算法和隐写分析检测器的最优性。机器学习模型在学习任务、训练示例的形式和可学习架构方面受到适当限制，将用于知识发现和构建与从业者相关的新型分析模型。总之，这些将导致更深入地了解实际的秘密系统的局限性和他们的检测。这将有助于开发更好的反欺骗方法，并在恐怖分子、犯罪分子和间谍用于计划和协调破坏性活动时对隐写术进行防御。这反过来又将推动隐写术的研究前沿和实际影响，而基于研究的方法可能对相关领域有用，包括数字取证和数据出处。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Steganography is a secret communication method that achieves privacy by hiding the actual message in some other innocuous looking object. Digital media, such as images, are ideal for this purpose because they can be slightly modified to encode a secret without making these modifications visible to a human or detectable by a computer. Steganography thus offers privacy to citizens in countries that censor communication channels or prohibit the usage of encryption. Modern steganography and the detection of its use heavily rely on machine learning and artificial intelligence. While these "black box" systems offer superior performance to existing analytical methods, comparatively little is known about the performance limits of such tools and even less about how and why these systems work. This project aims at addressing these deficiencies by combining conventional analytical modeling tools with modern machine learning in order to obtain insight into how modern artificial intelligence systems work in detecting steganography, establish their limits, and discover new fundamental knowledge about steganography in digital media. The key idea of the project is to combine the explanatory power of conventional analytic modeling tools with the flexibility of machine learning, leveraging their complementary strengths to address some of the most fundamental open problems in this field. Artificial but realistic image datasets will be used to compute limits and assess optimality of modern data-driven steganographic algorithms and steganalysis detectors. Machine learning models suitably restricted in terms of the task being learned, the form of the training examples, and the learnable architecture will be used for knowledge discovery and for building novel analytic models relevant to practitioners. Together, these will lead to deeper understanding of the limits of practical covert systems and their detection. This will facilitate development of better counter-deception methods and defenses against steganography when used by terrorists, criminals, and spies for planning and coordination of disruptive activities. This, in turn, will advance both the research frontier and the practical impact of steganography, while methods based on the research may be useful for related fields including digital forensics and data provenance.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.