MWSF 2025 Program
Excel to HTML
MONDAY 3 FEBRUARY 2025
Watermarking and Steganalysis
Session Chair: Jessica Fridrich, Binghamton University
09:30 - 10:30
Grand Peninsula B
09:30MWSF-298
Natural language watermarking with ChatGPT, Martin Steinebach, Fraunhofer SIT (Germany) [view abstract]
Digital watermarks for texts come in numerous forms. The text itself, but also its appearance, i.e. font, letter spacing or line spacing, can be modified. Here we present an approach that marks the text itself, i.e. changes the written words. Here, too, numerous methods are known, such as change and active to passive, modulation of sentence lengths or synonyms. We use ChatGPT to supplement existing texts with suggestions for synonymous formulations. We also look at evaluating the transparency of the marked texts with the help of ChatGPT.
09:50MWSF-299
Bag size-agnostic transformer for pooled steganalysis, Edgar Kaziakhmedov, Binghamton University (US); Jessica Fridrich, Binghamton University (US) [view abstract]
Batch steganography offers secure covert communication by distributing the secret payload across multiple images. Recent studies have shown that, at a fixed communication rate, this approach is less detectable selecting the bag size optimally. Traditional detection methods analyze images individually using a Single Image Detector (SID) whose soft outputs are pooled to evaluate the entire bag. While machine learning-based poolers have demonstrated superior performance, they require retraining for each bag size and communication rate. This work proposes a flexible solution using Vision Transformer (ViT) for batch steganography detection. The key advantages of our approach include input size agnosticism, allowing construction of poolers that work with arbitrary bag sizes and communication rates, interpretable decision-making by design, and off-the-shelf usability without the need for retraining. Our main contribution is the development of a transformer-based pooler that offers adaptability and interpretability in batch steganography detection, thereby addressing these limitations of existing methods.
10:10MWSF-349
JIST-first Accepted (JIST1981): A consideration of JPEG resistance verification of correlation-based steganography, Mariko Aikawa, Shibaura Institute of Technology (Japan) [view abstract]
With large amount of personal information being managed on the Internet, privacy protection in communication channels has become essential. Therefore, steganography, a secret communication technique that conceals information in an image, is attracting attention. By using this method to communicate with embedded ciphertext, the detection rate of encrypted communication can be reduced and privacy protection can be strengthened by making it appears to a third party that only image communication is taking place. Currently, many steganography methods with improved capacity and robustness have been proposed. In addition, JPEG compression resistance verification is required to extend its versatility. Correlation-based BMP steganography in previous research has improved confidentiality by embedding and extracting secret information using correlation. However, the compression resistance of the generated images has not been verified. In this paper, we evaluate compression resistance based on image quality and information recovery rate. Moreover, we show the minimum compression quality that satisfies the required acceptable retention conditions for JPEG compression of cover images.
10:30 – 11:00 and 15:00 – 15:30 Coffee Breaks
AI Generated Content Forensics
Session Chair: Adnan Alattar, Digimarc
15:30 - 17:30
Grand Peninsula B
15:30MWSF-300
Face swap forensics, Martin Steinebach, Fraunhofer SIT (Germany); Marco Fruehwein, Fraunhofer SIT (Germany) [view abstract]
Multimedia forensics is an important field addressing the increasing misuse of digital content, such as deepfakes and face-swapping technologies. This paper focuses on detecting face swapping. Our goal is not to decide whether face swapping has occurred. We assume that we execute a forensic investigation in which it needs to learned which photo of a person's face has been used for the face swap. We take a number of potential source face photographs and compare their behavior when reproducing the face swap. We show that the photo used for the face swap can be identified even after lossy compression and scaling.
15:50MWSF-301
EU AI-Act: Tagging genAI content, Julian Heeger, Fraunhofer SIT (Germany); Waldemar Berchtold, Fraunhofer SIT (Germany); Simon Bugert, Fraunhofer SIT (Germany); Martin Steinebach, Fraunhofer SIT (Germany) [view abstract]
The necessity to mark generative AI (genAI) content stems from the increasing challenges posed by AI systems capable of producing highly realistic synthetic material that is difficult for humans to distin- guish from authentic, human-generated content. The rapid advancements in AI technologies, coupled with the wide availability of such systems, have raised significant concerns regarding the integrity and trust within the information ecosystem. These concerns include the potential for misinformation, large-scale manipulation, fraud, impersonation, and consumer deception. Without a reliable way to identify AI-generated content, users are left vulnerable to these risks, which can severely undermine societal trust in digital communications and information. To mitigate these risks, it is essential to implement mechanisms that ensure AI-generated or AI- manipulated content is clearly identifiable. Embedding technical solutions, such as digital water- marks, metadata, cryptographic techniques, or content logging methods, can help trace the origin of content and ensure transparency. These methods should be reliable, interoperable, and robust, capable of withstanding technological advancements and misuse while effectively alerting consumers and authorities to the artificial nature of the content. Tagging genAI content serves several critical purposes: it helps maintain the credibility of the in- formation ecosystem, supports efforts to combat misinformation, and reduces the potential for fraud and impersonation. Moreover, as AI models grow increasingly sophisticated, the need for mechanisms to distinguish AI-generated content becomes more pressing. By embedding such solutions at the AI model or system level, both developers and downstream providers can comply with regulatory frame- works, such as the EU AI Act, while ensuring that AI-generated content is transparent and traceable, protecting consumers from deception and maintaining public trust in digital environments.
16:10MWSF-302
Cooking spiders: Efficient OSINT with chefs and recipes, York Yannikos, Fraunhofer SIT / ATHENE (Germany); Julian Heeger, Fraunhofer SIT / ATHENE (Germany); Simon Bugert, Fraunhofer SIT / ATHENE (Germany) [view abstract]
Social media, online forums, darknet marketplaces, and various other digital platforms are increasingly used or targeted by cybercrime. Therefore, open source intelligence (OSINT) has become an important aspect in digital forensics and cybercrime investigations: leveraging publicly available data on the Internet provides new information and offers insights into criminal behavior, patterns, and relationships. Many different tools and services exist to collect and extract data from websites for digital forensic investigations. These are often expensive and prone to errors when target websites change their structure or content.In this paper we present a media acquisition and multi-processing framework (�MAMPF�) for OSINT tasks. The framework is able to collect and extract data from various websites with easy extensibilty and maintenance in mind. We show that our framework makes a self-hosted approach to efficient OSINT possible where a centralized core component is utilized in such a way that nodes performing crawling / scraping tasks no not require any maintenance at all. To describe our approach we use the analogy of a restaurant with chefs that prepare dishes following specific recipes.
16:30MWSF-303
Robustness of reverse image search engines in the era of AI-generated media, Raphael Frick, Fraunhofer SIT | ATHENE Center (Germany); Felix Stein, Technische Universit�t Darmstadt (Germany); Katharina Wallrabenstein, Technische Universitaet Darmstadt (Germany); Sascha Zmudzinski, Fraunhofer SIT | ATHENE Center (Germany) [view abstract]
Reverse Image Search engines play a critical role in Open Source Intelligence (OSINT) and media forensics. They help uncover the origins of images, restoring original contexts and verifying authenticity. This capability is essential for distinguishing benign images from manipulated content. However, the robustness of these search engines is increasingly challenged by advancements in generative AI, which can produce highly realistic synthetic images. This paper analyzes the robustness of current reverse image search engines and highlights several shortcomings, emphasizing the need for enhanced resilience against AI-generated content to maintain their effectiveness in OSINT and media forensics.