Read before reproducing
I first converted the source paper into an auditable specification: what the paper states, what it leaves underspecified, and what would require an explicit implementation assumption.
Why I built it
Identity-document systems often focus on recognition, while presentation attacks can preserve readable content and alter the physical path between document and camera. I wanted to test whether explicit texture and edge representations add useful evidence beyond a strong semantic vision backbone.
What “paper-inspired” means
I do not claim authorship of the original architecture or equivalence to the paper's full experiment. Repository documentation separates paper-described components, my implementation choices, and the results observed on Protocol A-Reduced.
The structured approach I followed
Make data validity a stop gate
Before training, I audited dataset licenses, provenance, class/source relationships, document lineage, duplicates, and split leakage. The mixed-source composite proposal was rejected rather than hidden as a limitation.
Design for constrained hardware
On an 8 GB laptop, acquisition streamed selected frames, decoding stayed serial, preprocessing used CPU batch 1 and zero workers where needed, and the full dataset was never cached in RAM.
Freeze the experimental protocol
Protocol A-Reduced fixed 300 balanced samples, group-aware splits, portrait crops, normalization, seed, no-augmentation policy, and validation/test discipline before architecture comparisons.
Add one idea at a time
M3 established semantics. M4 isolated texture without edge. M5 isolated fixed-Sobel edge evidence without texture. The untested three-branch model was deliberately not reported.
Analyze failures, not only scores
I compared confusion matrices, class F1, correction overlap, new failures, padding behavior, learned fusion weights, and persistent errors—while avoiding causal claims the data cannot support.
What made the 24 hours possible
Scope discipline
Each milestone had explicit outputs and a stop gate. No API, frontend, or later model work entered an earlier research milestone.
Compute discipline
Local CPU smoke tests validated shapes, gradients, optimizers, and checkpoints. Real training used a temporary free Tesla T4 with mixed precision.
Evidence discipline
Validation selected checkpoints. Test was evaluated once. Negative results and shortcut risks remained in the final story.
Observed engineering envelope
~573 MiBportrait-extraction peak RAM
~0.303 smedian CPU preprocessing / image
~0.398 sp95 CPU preprocessing / image
8 × 4GPU batch × gradient accumulation
Limitations
300-sample reduced protocol
Only 44 test samples
One primary random seed
Full DLC corpus not evaluated
Unequal base-document coverage
Possible padding-related shortcut
No stochastic augmentation
M3 vs M4/M5 is not head-matched
No statistical significance claim
No production deployment evaluation
No liveness guarantee or security certification
No state-of-the-art claim
Open research, without exposing private data
Safe source code, aggregate metrics, reproducibility metadata, confusion matrices, training curves, and methodology are public. Raw dataset media, prediction-level identity-document crops, private bundles, and large checkpoints remain excluded.
View the implementation on GitHub ↗