LamTNguyen/ridgelora-cross-sensor-sd302d-f-to-m-20260825

RidgeLoRA-FP: SD302A-F to SD302D-M cross-sensor experiment

This public archive contains the leakage-controlled direct cross-sensor experiment used to evaluate whether Stage-2 synthetic target-sensor images help recognition on a physically different real sensor.

Locked protocol

• Source/condition sensor: NIST SD302A device F.
• Target sensor: NIST SD302D device M.
• Identity: subject:finger-position; the same fingers exist across both collections.
• Subject split: 160 train / 20 validation / 20 test subjects, fixed before optimization.
• Stage-2 LoRA training uses only target-M images belonging to training subjects: 1,580 paired fingers from 158 subjects. Device M is missing all ten fingers for two of the 160 training subjects.
• Generation conditions: 1,600 SD302A-F training fingers, three deterministic synthetic impressions per finger (4,800 images total).
• Validation: 200 held-out F-to-M finger pairs. Final test: 199 held-out F-to-M finger pairs. Target test pixels are not loaded until model selection ends.
• No validation/test subject is used to train the generative model, LoRA, style/pose pool, or recognizer.

The recognizer comparison has four arms: real only, matched real oversampling, matched standard augmentation, and real plus RidgeLoRA-FP Stage-2 synthetic images. Matched arms have exactly 6,400 slots per epoch and the same four-slot-per-class histogram. The recognizer is an ImageNet-1K V1 ResNet-18 trained for 60 epochs with seeds 123, 456, and 789.

Checkpoints and generation settings

• Stage-2 base checkpoint: clean SD302A train-160 split, step 50,000.
• Target-M adapter: rank 8, alpha 8, middle self-attention qkv and proj_out, 12,288 trainable parameters, 5,000 steps.
• Sampling: DDIM-50, eta 0, control scale 1.0, BF16, 512 x 512.
• Target-pose alignment: bounded similarity transform with scale in [0.75, 1.25], followed by target-mask clipping.

Important SHA-256 values:

• IMPOSE VQ/base checkpoint: 3048e170b8fd4687a18b54ca823dcb01ca1a0f5fe2f5affc297a1f00072763ab
• Clean Stage-2 checkpoint: 67d62399ea8f027ed1de6a81cd736312a43993d63860ef922d9f63ecee3090dd
• Target-M LoRA checkpoint: a8f164f63892bfa0fdaa6d1167e1327850ac86d6d13565a509703e259d20330c
• Full synthetic manifest: cde96bcedb91564ca8a16d5a0faa3ed2aa10fd77533f24de2d23230382929eef

Restore

Download the archive into a fresh working directory:

hf download LamTNguyen/ridgelora-cross-sensor-sd302d-f-to-m-20260825 \
  --repo-type dataset --local-dir cross_sensor_sd302d_f_to_m

Obtain SD302A and SD302D through their respective public dataset pages, then place them at the paths recorded in the manifests or rewrite only the path prefixes. Do not alter subject IDs, finger positions, row order, generation seeds, or synthetic-impression indices.

The source/ directory contains the exact preparation, LoRA-training, generation, QC, and recognizer scripts. Reports in protocol/, lora/, qc/, synthetic/, and recognizer/ record the arguments and hashes needed to resume or audit the experiment.

Results

Final held-out SD302D-M test, mean +/- sample SD over seeds 123/456/789:

Training arm	EER (%)	TAR@FAR 1% (%)	TAR@FAR 0.1% (%)	Rank-1 (%)	Rank-5 (%)
Real only	45.72 +/- 1.84	1.01 +/- 0.50	0.17 +/- 0.29	1.68 +/- 0.29	4.69 +/- 1.05
Real oversampling	43.72 +/- 0.50	0.84 +/- 0.29	0.00 +/- 0.00	0.84 +/- 0.29	4.69 +/- 0.77
Standard augmentation	42.51 +/- 0.27	2.68 +/- 0.58	0.34 +/- 0.58	2.35 +/- 0.58	5.53 +/- 0.87
Stage-2 target-M synthetic	40.91 +/- 1.52	2.68 +/- 0.58	0.17 +/- 0.29	2.35 +/- 0.29	6.87 +/- 1.16

Stage-2 augmentation lowers mean EER by 1.61 percentage points relative to matched standard augmentation and by 2.81 points relative to oversampling. TAR@1% and Rank-1 are tied with standard augmentation. These three-seed differences are descriptive; no statistical-significance claim is made. The absolute TAR and identification rates remain weak under the F-to-M shift.

recognizer_pretrained_full60_3seed/cross_sensor_results.json contains all per-seed test metrics, validation histories, selected checkpoints, and paired seed differences.

Scope

This experiment directly tests real SD302A-F conditions -> synthetic SD302D-M training augmentation -> held-out real SD302D-M downstream recognition. It is separate from the paper's full Stage-1 virtual-identity pipeline and does not by itself establish end-to-end Stage-1 utility.