The FREUID Challenge 2026 — Solution (Team EliasTSJ + Sekinal)

Fraud detection on identity documents for The FREUID Challenge 2026 (IJCAI-ECAI).

Results

Model	Public LB	Cross-country (LOCO)	Capture canary (AUC)
Public specialist (bm1024)	0.02514	0.033 Moz / 0.010 Benin / 0.007 Maur	0.726
Robust / private (slot2v3 3-seed ensemble)	0.11925	0.025 (Mozambique, epoch 0)	0.917

Both final submissions share the same backbone (EffNetV2-M) and training data (official FREUID train set only — no external data). They differ in exactly two training flags, which specialize them for the two announced test regimes:

	Public specialist	Robust model
Resolution	1024	1024
Capture-simulating augmentation	off (preserve fragile digital artifacts)	heavy (JPEG recompress, blur, moiré, perspective)
Epochs	3	1 (early stopping — more epochs memorize fragile artifacts)
Seeds	1	3, probability-averaged

Key findings (see the technical report for details)

1. Resolution is the universal lever. Fraud edits leave pixel-scale traces (resampling seams, JPEG-grid mismatches). Downscaling to 384–768 destroys them; at 1024 (near-native) both in-distribution AND cross-country scores improve by an order of magnitude (public 0.123→0.025; unseen-country 0.40→0.01–0.03). The traces generalize across countries — templates do not.
2. Early stopping controls the robustness trade-off. With capture-aug, epoch 0 generalizes best (0.025 vs 0.097 at epoch 2 on a held-out country).
3. External data poisoned everything it touched (measured, then removed): IDNet-scanned killed real-capture detection (canary AUC 0.50); GenAI synthetic fraud made from FREUID genuines taught "unknown country = fraud" (LOCO 1.0); pseudo-labels dragged the student toward a weaker teacher.
4. Leave-One-Country-Out (LOCO) validation — not the public LB — was the compass for every robustness decision, mirroring the announced private-test design (unseen countries + print-and-capture).

Repository layout

freuid/          core package: data pipeline, transforms, model, metrics
scripts/         numbered experiment scripts (EDA → training → submission)
infer.py         standalone inference entrypoint (used by the Docker image)
docker/          Dockerfile + build instructions (no-network sandbox contract)
docs/            notes and experiment logs

Reproducing the final models

Hardware used: 1× NVIDIA L40S 48 GB, ~3 h total training. Data: official competition data only, extracted to data/extracted/.

# 1) Public specialist (single seed, 3 epochs, no capture-aug, 1024 px)
python3 scripts/24_train_fusion.py --train-all --no-fusion --loss bce \
    --img-size 1024 --no-capture-aug --batch-size 12 --workers 10 --epochs 3 --seed 4

# 2) Robust model (3 seeds, 1 epoch, heavy capture-aug, 1024 px)
for S in a b c; do
python3 scripts/30_train2.py --train-all \
    --backbone tf_efficientnetv2_m.in21k_ft_in1k --img-size 1024 --batch-size 12 \
    --workers 10 --loss bce --aug-strength heavy --epochs 1 --save-name slot2v3_1024_$S
done

Inference

python3 infer.py --data-dir /path/to/images --out submission.csv \
    --checkpoints weights/slot2v3_1024_a.pt,weights/slot2v3_1024_b.pt,weights/slot2v3_1024_c.pt \
    --batch-size 32 --workers 8

Reads a flat directory of images (.jpeg/.jpg/.png/.webp/.bmp/.tif/.tiff), writes id,label CSV (id = filename without extension, label = fraud score). Checkpoints are probability-averaged with horizontal-flip TTA.

Final model weights are published as a GitHub Release (see Releases tab).

Docker (no-network sandbox)

See docker/README.md. The image embeds the weights and code; it reads /data/ (read-only) and writes /submissions/submission.csv, with --network none, per the competition reproducibility contract.

External resources credited

• timm — ConvNeXtV2 / EfficientNetV2 backbones (Apache-2.0)
• PyTorch / torchvision
• Explored but not used in the final models: IDNet-2025 (CC-BY-4.0), FantasyID, InsightFace inswapper, SDXL-inpainting (all documented in the report)

Code-freeze compliance

Timeline anchors (all organizer-verifiable):

• Final model weights frozen 2026-07-08 17:12 UTC — the final-models GitHub Release asset timestamps.
• Selected final submissions produced 2026-07-07 / 2026-07-08 (Kaggle submission timestamps) — five days before the private test release.
• Last change to model / training / inference code: commit 19630e2, pushed 2026-07-12 03:07 UTC — before the private test release (2026-07-13 07:02 UTC).

Commits after the private test release only add documentation, result artifacts, and archival copies of pre-freeze experiment launcher scripts (scripts/experiments/, runs corroborated by Kaggle submission timestamps). Verify with one command — no solution code changed after 19630e2:

git diff 19630e2 HEAD --stat -- freuid/ infer.py docker/Dockerfile \
  docker/entrypoint.sh 'scripts/*.py' 'scripts/*.sh'   # additions only, 0 modifications

License

MIT — see LICENSE.