Parth Rana

DeepScope

Deepfake detection from the involuntary signals of a real face

3rd Place Grand Award, Software Design · Regeneron ISEF 2026

camera

xi(t) ∈ ℝ3,  i = 1, …, 478

fs = 30 frames per second

procrustes

(s, R, b) = arg min Σi ‖s R xi + b − μi‖2

ri(t) = s R xi(t) + b − μi

signals

b(t) = ½(βL(t) + βR(t)),   blink ⇔ b(t) > 0.5

s(t) = SSIM(ct, ct+1),  ct = motion-compensated face crop

spectrum

E = ∫812 P(f) df,   P = Welch PSD of ri(t), averaged

fN = fs / 2 = 15 Hz

twelve metrics

cascade

p1 = g1(landmark metrics),   settle if |p1 − ½| ≥ m

otherwise p = g2(all fifteen features),  g1, g2 gradient boosted

The pipeline running live on a synthetic face. The motion, blinks and texture signal are simulated in your browser, and the alignment, spectrum and metrics are computed from them as you watch. The stages and equations are the ones DeepScope uses.

The question

As fake videos get convincing enough to use for scams and misinformation, most detectors learn the artifacts of the generators they were trained on, and they struggle when a new generator comes out.

I’m testing whether detectors built on involuntary biological signals, like micro-tremors and blinking, can catch deepfakes from generators they’ve never seen. The idea is simple: generators aren’t trained to fake biology.

What I built

DeepScope tracks the face with MediaPipe, removes head motion so only the small non-rigid movements are left, and turns each clip into twelve biological-plausibility measurements, from blink timing to skin-texture persistence. A two-stage cascade of gradient-boosted classifiers makes the call: the first stage sees only the landmark metrics and settles a clip early when it’s confident, and the rest go on to a second stage that sees all fifteen features.

How I’m testing it

The first version was my science fair project. Then I rebuilt it from scratch on open datasets and put the idea to a harder test, with forgeries I generated myself: a face swap and LivePortrait reenactments, which I made on GPUs at the University of Minnesota’s Minnesota Supercomputing Institute.

Every fake can be checked against the real clip it came from, and every number comes straight out of code.

Parth at his DeepScope booth during ISEF judging
At my DeepScope booth during judging, Regeneron ISEF 2026 in Phoenix.

Timeline

Started as my science fair project.

1st Place, Twin Cities Regional Science Fair, and the ProEquip Award for Safety in AI.

3rd Place Grand Award in Software Design at Regeneron ISEF, the world’s largest pre-college science competition.

Rebuilt from scratch on open datasets, with forgeries I generated myself.

More

ChemPrintAll projects