Research Questions
What makes DeepFake detection break under unstable face sequences, reposting, re-encoding, or audio-only circulation?
Preview Page · Not in Top Menu
DeepFake & Media Forensics
This academic-facing page is not just a short publication list. It connects ACVLab’s research line to real misinformation incidents, source-backed teaching materials, and public-interest verification workflows. The goal is to show how research questions emerge from actual cases rather than only from benchmark design.
How This Page Is Structured
Real Incidents
Every highlighted case below is tied to a concrete public event, not a generic stock example.
Reusable Teaching
The AI literacy course turns research concepts into evidence-backed modules for students, collaborators, and outreach.
Deployment Bridge
This line sits between lab research, media verification, fact-checking, and public communication.
Real Incidents That Motivate the Research
Case *1
Conflict Footage That Looks Cinematic
A viral clip claimed that Tel Aviv was under a meteor-like airstrike. TFC identified it as highly likely AI-generated. This is a good example of synthetic conflict footage that spreads because it looks visually dramatic and urgent.
Case *2
Edited Speech, Reframed Narrative
The so-called “20 years ago Lai Ching-te speech” case is not only about generation. It is also about editing, reframing, and fabricated context. This is exactly why real-world media forensics cannot stop at a single detector score.
Case *3
Audio Misinformation Travels Differently
The fake audio of Taiwan’s president criticizing his predecessor shows why verification cannot stay video-only. Audio claims, reposted clips, and platform-level distortion create different failure modes from clean benchmark settings.
Case *4
Emotionally Warm, Socially Shareable, and False
The “elderly duet” case shows how AI-generated clips can look harmless or uplifting while still training audiences to lower their guard. These cases matter because social sharing often rewards emotion before verification.
Case Archive *5
Deepfake Events Need Longitudinal Analysis
READr’s deepfake event archive is useful because it moves beyond one-off panic. It shows the range of cases, recurring manipulation patterns, and why some false media matters more than others.
Guideline *6
Verification Practice for Newsrooms
FactLink’s AI verification guideline is worth showing alongside the cases because it turns the lab’s broader argument into newsroom-facing practice: suspicious imagery should be checked with workflow, not intuition.
Representative Research Outputs
Reusable Teaching and Verification Stack
Forensics Module
Explains synthetic media, forensic cues, and why detectors fail in realistic settings.
Verification Tools
Shows practical search, source-checking, and evidence comparison workflows.
Case Studies
Uses source-backed examples instead of only abstract warnings.
Detection Demo
Public-facing system entry for trying detector-assisted verification.
Source Notes
- TFC, “網傳『特拉維夫被流星雨式空襲』影像，極可能為AI生成”
- TFC, “20年前賴清德演講是虛構變造的影片”
- RFA / Asia Fact Check Lab, leaked audio case
- TFC, “九旬老夫婦二重奏是虛構的AI生成影片”
- READr deepfake event archive
- FactLink, AI for Trust newsroom guideline
All external images on this page are tied to the corresponding original report or article above. They are used as cited event references, not as detached stock visuals.