Label Shield
MobileNetV3-Small classifier trained on synthesised hologram patches — authentic vs tampered. Weights transferred from SPIRIT's Lock-3 model (shared RNG seed); a regenerated training set is shown below for transparency.SIMULATED
Validation accuracy
100.0%
Architecture
mobilenet_v3_small
Resize(224) → Normalize(ImageNet)
Train / val per class
60 / 20
Inference device
cpu
224×224
Confusion matrix · validation
| predicted | ||
|---|---|---|
| Authentic | Tampered | |
| Authentic | 20 | 0 |
| Tampered | 0 | 20 |
Sample patches
The model learned that authentic holograms have centred radial gradients, unbroken concentric rings and faint diagonal stripes; tampered patches add pixelation, broken arcs, off-axis colour blotches and heavier noise.
Authentic
8 samples
Tampered
8 samples
Provenance
DRISHTI lifts SPIRIT's pre-trained Lock-3 weights. The synthetic dataset is regenerated locally for transparency; weights remain valid because the RNG seed (11) is shared. Phase-3 swaps the synthetic-tampered class for real seized-counterfeit photos collected in the 3-district pilot.
weights · transferred from SPIRIT lock3_label_shield.pt (RNG seed 11 shared)