Checklist EUROAI (CEOAI) 2026 Contest Day 2 · Task 1
Stowaway
Hide and recover 256-bit messages in the weights of a small MNIST CNN: choose prunable weights, decode hidden LSB schemes, and encode messages that survive fine-tuning.
The task
StowawayCNN (conv1 16 filters, conv2 32 filters, a 32·7·7→10 linear layer; about 99% MNIST accuracy) has 20,432 weights in conv1.weight, conv2.weight and fc.weight. A bit can be hidden in a weight by setting the lowest bit of int(round(w × 1e6)).
Subtask 1 (Prunability Desk, 30 pts): nominate exactly 12,000 weights to prune; the reference is the 12,000 weights with lowest Fisher information. Subtask 2 (Decoder Desk, 35 pts): recover the 256-bit messages from eight provided networks encoded with four hidden schemes (tensor, 256-wide slice, bit position). Subtask 3 (Retraining Desk, 35 pts): write eight given messages into fc.weight as per-weight changes with |delta| ≤ 0.02 so that a fixed public reader (Gaussian jitter σ = 0.001, orthonormal DCT, sums of coefficients 64 + 61i to 124 + 61i) still recovers them.
Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.
At a glance
- You get
- Starter kit:
baseline_cnn.pt,reference_messages.json,decoder_grammar.md,encoded_models/model_00.pt…model_07.pt,mnist_val_sample.npz(1,000 images),mnist_finetune_batch.npz, 'Stowaway (RAW).ipynb' with local evaluators. - You submit
- One CSV (subtaskID, datapointID, answer): prune flags by global weight index (subtask 1), recovered bits with datapointID = model × 256 + bit (subtask 2), per-weight deltas with datapointID = model × 15680 + widx (subtask 3).
- Scoring
- points = round(clamp((metric − floor)/(ceil − floor), 0, 1) × max). Subtask 1: F1 with the Fisher reference (exactly 12,000 ones required), floor 0.583, ceiling 0.90. Subtask 2: per-network bit accuracy, floor 0.60, ceiling 1.00, averaged over 8 networks. Subtask 3: bit accuracy after the reader, floor 0.50, ceiling 0.90; any |delta| > 0.02 zeroes that network.
- Rules
- Biases may not be modified.
- Nothing uploaded is executed.
- Submission limit 50, two final submissions (platform settings).
- Format
- Contest Day 2, 17 July 2026 (Nitro AI Judge 07:10–12:50 UTC), on-site in Cluj-Napoca; three tasks, 100 points each.