Welcome by
Prof. Simon S. Woo (WDC General Chair)
1:50-2:00 SGT
Keynote I
2:00-2:45 SGT
Title: FakeAmplifier: Using Adversarial Attacks to Enhance DeepFake’s Visual and Social Influence
Keynote speaker: Dr Qing Guo, A*STAR's Centre for Frontier AI Research (CFAR), Singapore
[Slides]
Speaker’s Bio
Dr. Qing Guo is currently a Senior Research Scientist and principal investigator (PI) at the Centre for Frontier AI Research (CFAR), Agency for Science, Technology, and Research (A*STAR) in Singapore, and an adjunct assistant professor at the National University of Singapore (NUS). In 2019, he joined Nanyang Technological University (NTU) as a Research Fellow and was subsequently appointed as a Wallenberg-NTU Presidential Postdoctoral Fellow in 2020. He received the Best Platinum Paper Award at ICME in 2018, the ACM Tianjin Outstanding Doctoral Dissertation Award in 2020, the third place in the AI SG Trusted Media Challenge 2022, won the Best Paper Award at the ECCV 2022 AROW workshop, awarded AISG Robust AI Grant Challenge in 2023 and the Digital Trust Centre Research Grant for research on AI model fairness in 2024. His research mainly focuses on AI safety and computer vision. He has published over 50 papers in top-tier conferences and journals. He serves as a Senior PC for AAAI 2023/2024 and vertical chair for Resilient and Safe AI at the IEEE Conference on Artificial Intelligence (CAI) in 2024.
Abstract
With the rapid development of deep generation technologies, it has become increasingly easy for people to create realistic yet fake images and videos. To prevent the misuse of DeepFake content, researchers have developed various methods to detect fake media and ensure safe generation. In this talk, we introduce our work on leveraging adversarial attacks to evade DeepFake detections, bypass safety filters, and embed sensitive social attributes, thereby revealing the potential to amplify DeepFake’s negative visual and social impact. Specifically, we propose distribution-aware adversarial methods to align fake content with the distribution of real content, making DeepFakes difficult to identify through visual cues. We also introduce JailBreak methods for black-box commercial generation models to bypass safety filters. Additionally, we propose emotion-aware backdoor attacks and attribute-aware attacks to compel generators to produce content sensitive to social groups, highlighting the vulnerability of deep generators from a social perspective. These findings can help in developing more advanced DeepFake detectors and safer deep generation techniques.
Full & Short Papers
Session Chair: Simon S. Woo
2:45-3:00 SGT
Honeyfile Camouflage: Hiding Fake Files in Plain Sight
Authors: Roelien C. Timmer, David Liebowitz, Surya Nepal and Salil Kanhere
3:00-3:15 SGT
Towards Generalized Detection of Face-Swap Deepfake Images
Authors: Faraz Ghasemzadeh, Tina Moghaddam, Jingming Dai, Joobeom Yun and Dan Dongseong Kim
3:15-3:30 SGT
On the Correlation Between Deepfake Detection Performance and Image Quality Metrics
Authors: Hyunjoon Kim, Jaehee Lee, Leo Hyun Park and Taekyoung Kwon
3:30-3:50 SGT
Tea Break
Keynote II
3:50-4:35 SGT
Title: User Literacy about Deepfake Technology
Keynote speaker: Assoc. Prof. Dr. Chei Sian Lee, Nanyang Technological University, Singapore
[Slides]
Speaker’s Bio
Associate Professor Chei Sian Lee is currently Associate Chair (Faculty) at the Wee Kim Wee School of Communication and Information at the Nanyang Technological University. She is actively involved in research on issues related to everyday user-information interaction within digital environments. Recently, her research has focused on how digital and emerging technologies can be designed to facilitate everyday information practices. Specifically, she is currently investigating deepfakes and generative artificial intelligence phenomena from an information-oriented perspective, focusing on credibility assessment mechanisms that are beneficial to student and adult learners. She is on the Editorial Advisory Board for Computers and Education, Online Information Review, The Electronic Library and Journal for STEM Education Research. Dr Lee received her B.Sc. and M.Sc. degrees in Computer and Information Sciences from the National University of Singapore and her PhD in Management Information Systems from the University of Illinois at Chicago, Liautaud Graduate School of Business.
Abstract
Deepfakes are artificially created media posing as actual video recordings and are a potential source of fake news or disinformation. Although research has been done in developing algorithms for the automatic detection of deepfakes, more work needs to be conducted on user literacy related to deepfake technology. This is a critical missing link because algorithms are currently not performing at a level where human judgment is unneeded. This presentation will examine the concept of user literacy on deepfake technology from the user studies we conducted to understand users’ perceptions and judgments towards deepfakes as well as the verification strategies adopted when engaging with content from deepfakes. Opportunities and challenges will also be discussed.
Poster and Discussion Papers
Session Chair: Tim Walita
4:35-4:50 SGT
Exploiting LLMs for Scam Automation: A Looming Threat
Authors: Gilad Gressel, Rahul Pankajakshan and Yisroel Mirsky
4:50-5:05 SGT
A Photo and a Few Spoken Words Is All It Needs?! On the Challenges of Targeted Deepfake Attacks and Their Detection
Authors: Raphael Antonius Frick and Martin Steinebach