You're cordially invited to Jiameng Pu's Ph.D. preliminary proposal presentation on Thursday, 18 November 2021, at 1pm. Zoom link: https://virginiatech.zoom.us/j/81881081504 Title: Defending Against Misuse of Synthetic Media: Understanding Real-world Challenges and Building Robust Defenses Committee: Dr. Bimal Viswanath (chair) Dr. Daphne Yao Dr. Chandan Reddy Dr. Tijay Chung Dr. Gang Wang (UIUC) Abstract: Recent advances in deep generative models have enabled the generation of realistic synthetic content or deepfakes. This includes synthetic images, videos, and text. However, synthetic media can be misused for malicious purposes and damage users' trust in online content. This preliminary proposal aims to address several challenges associated with defending against misuse of synthetic media. Key contributions and ongoing work include the following: (1) Understanding challenges with the real-world applicability of existing defenses that aim to detect synthetic content. We curate synthetic videos and text from the wild (i.e., the Internet community), and assess the effectiveness of state-of-the-art defenses on content in the wild. In addition, we measure the adversarial robustness of existing defenses, and also propose novel low-cost adversarial attacks against existing defenses. Our findings reveal that most defenses show significant degradation in performance under real-world detection scenarios, which leads to the second thread of my research work: (2) Building synthetic content detection schemes with improved generalization and robustness. Most existing synthetic image detection schemes are highly content-specific (e.g., only designed for faces), thus limiting their application. I propose an unsupervised content-agnostic detection scheme called NoiseScope, which can be applied to a wide variety of synthetic images, and is more resilient against an adaptive attacker. My ongoing work further aims to extend this approach to synthetic videos. For the text modality, my preliminary work reveals that state-of-the-art defenses that mine sequential patterns in the text using Transformer models are vulnerable to simple evasion schemes. My ongoing work aims to address this challenge by developing defenses that leverage the semantic structure of text. --Bimal