BlinkTrace
Building toward reliable real-time deepfake detection across evolving generation methods.
BlinkTrace is an independent research effort focused on a practical question: which forensic signals still hold up when manipulation methods change? The work combines appearance-based detection, temporal facial dynamics, and repeatable evaluation to support a future real-time detection pipeline. The current research is vision-focused and does not train on or detect audio.
Detection quality should be measured under generation shift, not only on familiar training distributions.
BlinkTrace compares visual appearance cues, temporal facial geometry, and fusion under held-out validation.
Real-time inference is the destination, with evaluation rigor and automation built first.
Current Research Focus
BlinkTrace is organized around a practical research question: which detection signals remain reliable when the generation method changes, and how does that translate into a future real-time system? The current scope is strictly computer vision: no audio training and no audio detection are part of the pipeline.
Real-Time Direction
The long-term goal is a real-time deepfake detection pipeline that remains useful as manipulation methods evolve, not just a benchmark classifier.
Generation-Shift Evaluation
Experiments emphasize held-out generation methods and source-matched validation to reduce misleadingly easy results and surface real transfer behavior.
Visual Signal Comparison
BlinkTrace compares appearance-based detection with temporal facial dynamics to understand which visual signal families remain useful under distribution shift.
Automated Research Pipeline
Processing, training, validation, and reporting are being shaped into a repeatable loop that can support future deployment-minded experimentation.
What The Project Is Testing Now
Appearance-Based Detection
Face-crop models learn texture, blending, shading, and resampling artifacts that may transfer across held-out generation methods such as FaceFusion, LivePortrait, HelloMeme, Diff2Lip, LatentSync, and Memo.
Temporal Facial Dynamics
Landmark-sequence models test whether motion consistency and facial geometry add signal beyond still-frame artifacts.
Evaluation Discipline
Results are framed around held-out generation methods, source-matched validation, and skepticism toward inflated random-split performance.
Deployment Path
Training, validation, monitoring, and delta-only pipeline automation are being shaped into a repeatable real-time detection workflow.
Recognition
BlinkTrace was selected for UBC Innovation's 51st cohort.
That selection reflects the project's direction in deepfake detection research, especially its focus on robust evaluation, vision-focused experimentation, and real-time deployment potential.
Public Repository
BlinkTrace will publish its public repository once the first shareable version of the project is ready. For now, this site captures the research direction, technical story, and current progress.