blinktrace · research program

Which signals survive generation shift?

BlinkTrace studies vision-based deepfake detection under generation shift, with a long-term goal of supporting reliable real-time detection workflows.

01 · research thesis

BlinkTrace treats deepfake detection as a representation problem under generation shift, not a closed-set classification problem. The core question is which forensic signals remain useful when manipulation methods change, compression changes, and shortcut cues stop being reliable.

vision-onlyno audio trainingno audio detectiongeneration shiftshortcut-risk aware

02 · why real-time detection is hard

No shortcuts survive contact with a live stream.

Real-time deepfake detection cannot depend on convenient shortcuts or familiar generator fingerprints. A detector that only works on known distributions becomes fragile as soon as the generation family, encoding path, or source conditions shift.

held-out generation methods

source-matched evaluation

shortcut-risk awareness

deployment-relevant testing

03 · methodology

Branches compared under one protocol.

Appearance-based detection

An EfficientNet-B0 branch analyzes tracked, normalized face crops to learn transferable signals — blending, shading, texture and resampling artifacts — under held-out generation-family evaluation across families including FaceFusion, LivePortrait, HelloMeme, Diff2Lip, LatentSync and Memo.

x → EffNet-B0 → GAP → MLP → p_cnn

Temporal facial dynamics

A GRU branch models normalized landmark sequences — position, velocity and acceleration at every timestep — to test whether facial motion geometry provides useful signal beyond still-frame appearance cues when the same manipulation families are held out.

L_t = [x̂_t, v_t, a_t] · h_t = GRU(L_t, h_(t−1))

Fusion and evaluation

Shared held-out validation enables honest comparison between branches, and tests whether fusion adds new ranking power or mainly improves operating-point behavior. Exploratory motion, frequency and pulse-consistency layers are evaluated under the same protocol.

p_final = fusion(p_cnn, p_gru)

04 · curated findings

What the current evidence says.

0.82 AUC

Appearance branch ceiling in a held-out run

A representative CNN run climbs from near-chance initialization to the low 0.8 AUC range — real signal, without trivial initialization leakage.

0.62–0.67 AUC

Temporal branch on shared held-out tests

Landmark-only modeling carries some signal, but currently transfers less reliably than appearance-based detection.

0.95+ acc

Fusion can improve hard decisions

Late fusion improves operating-point accuracy on shared held-out splits, even when ranking metrics remain dominated by the appearance branch.

1.0 AUC cases

A warning sign, not a victory lap

Some generation families may still be too easy — likely generator fingerprints or codec shortcuts, which is exactly why evaluation rigor is part of the research problem.

05 · toward real-time deployment

A model is not a system.

Deployment direction

Real-time detection requires more than a promising model. BlinkTrace builds the surrounding research operations too: delta-only processing, training triggers, validation reports, storage checks, and orchestration logic that supports repeatable inference-oriented experimentation over time.

Research operations

New visual inputs are discovered and routed through frame extraction, face cropping, landmark generation, normalization and sequence building, while storage limits, crop yield, landmark miss rates and training reports are monitored along the way. Preprocessing reliability and evaluation discipline are treated as part of the detection system — not separate cleanup work.

06 · limitations and next steps