基於零知識證明之區域感知影像保證系統

Abstract

隨著多媒體技術與社群軟體的快速發展，數位影像在我們的生活中變得無所不在。然而，影像編輯軟體也越來越強大，甚至連AI都可以直接生成出影像。這些發展使得有心人士可以輕易的竄改或製造出影像內容，進而散播錯誤或惡意的假資訊，最終導致社會成本增加。 影像認證是一項可以保護影像完整性與真實性的技術。傳統的主動影像認證方法可以有效地保護原始影像的完整性，但它們無法使編輯後的影像被認證，然而，對原始影像進行合理的影像編輯往往是影像發布前重要過程。零知識證明的出現為主動影像認證開闢了新的可能性，它使影像可以在不洩露其原始影像的情況下證明編輯後影像的完整性與真實性。然而，目前基於零知識證明的方法在滿足實際用途方面存在不可接受的性能。 在本論文中，我們提出了一個基於零知識證明之區域感知影像保證系統，以克服現有方法在效能上所面臨的挑戰。進一步來說，本系統利用馬賽克、拼貼等局部編輯功能保護影像的敏感區域，針對編輯區域進行零知識證明，非編輯區域進行數位簽章來，進而使得本系統仍然保證了整個影像的認證保證。我們全面分析了系統性能，並透過一個真實世界隱私影像資料集（Redactions Dataset）證明了本系統的實用性。實驗結果顯示，本系統相較目前最先進零知識證明影像認證系統，分別在 KeyGen 與 Proof 運算上快 15 倍與 60 倍的運算效率，在 Proof 的大小上降低了 25 倍的記憶體用量；此外，在所選的 Redactions Dataset 中影像中，有 180 倍到 2700 倍的效能提升，顯著地減少了編輯影像所需的證明時間。 綜合上述，我們透過廣泛的實驗證明了系統的實用性，期望我們的努力能使進階影像認證技術朝著現實世界應用的實際解決方案向前邁進一步，從而為構建可信賴的互聯網這難題做出小小的貢獻。

With the rapid advancement of multimedia technology and social software, digital images have become ubiquitous in our lives. However, image editing software is growing increasingly sophisticated, and even artificial intelligence can directly generate images. Such development has made it effortless for those with nefarious intentions to manipulate or fabricate image content, thereby spreading misinformation and disinformation, ultimately leading to an escalation in social costs. Image authentication is a technology designed to safeguard the integrity and authenticity of digital images. Traditional active image authentication methods effectively protect the integrity of a source image, but they cannot make edited images authenticated. Nevertheless, reasonable editing of images is often essential before their release. The advent of Zero-knowledge Proof (ZKP) opens up new possibilities for active image authentication that can prove the integrity of an image after editing without revealing its source. However, ZKP-based methods suffer from unacceptable performance in fulfilling real-world usages. In this paper, we propose a Zero-Knowledge Proof Based Region-aware Photo Assurance System to overcome the performance challenges of existing works. The system utilizes local editing functions such as mosaic and collage to protect sensitive areas of an image, performs ZKP for the edited region, and digital signature for the non-edited region to ensure complete image authentication. We comprehensively analyze the system's performance and demonstrate its practicality with a real-world privacy image dataset (Redactions Dataset). Experimental results show that, compared to the most advanced ZKP-based image authentication system, our system is 15 times faster in KeyGen and 60 times faster in Proof operations, respectively, and reduces memory usage by 25 times in the size of Proof. Moreover, the system's performance is optimized by 180 times to 2700 times on selected images in the Redactions Dataset, leading to a significant reduction in the proof time required for image editing.