Downloads · 30 days
0
Daemons-Q/qsh-medical-breastmnist-ckks-student
qsh-medical-breastmnist-ckks-student is a image classification model from Daemons-Q. Use it when you need a label for an image. It is set up for onnx. The card lists the license as other.
1 conv block (1-4ch, 3x3), avg-pool, dense head, 14x14 input
Downloads · 30 days
0
Access
Public
Updated Nov 13, 2026
Parameters
3.2K
163 KB on disk
Likes
0
Public
Click a slice to open those files.
.json513 KB · 92%
From the Hugging Face model README
1 conv block (1->4ch, 3x3), avg-pool, dense head, 14x14 input
Part of QSMPC-QKD-QHE-AI-Hybrid,
a quantum-safe orchestration demo.
This is the encrypted-tier model for the medical_fl use case: a
distilled student small enough to evaluate under CKKS. Whether the public demo
encrypts THIS model differs by use case; the notes below say what actually runs.
It is not the plaintext teacher, and its numbers should be read against that
teacher rather than against the published literature. Its agreement with the
teacher was measured offline, in plaintext.
| metric | value |
|---|---|
agreement | 0.826923 |
auc | 0.788221 |
disagreement_on_positives | 0.149123 |
majority_class_rate | 0.730769 |
metric_delta | 0.096154 |
metric_delta_pp | 9.6154 |
n_classes_predicted | 2 |
n_eval | 156 |
student_metric | 0.769231 |
student_params | 3226 |
teacher_metric | 0.865385 |
teacher_params | 11171266 |
This scores far below the 0.901 ResNet-18 baseline. The reason is representational rather than cryptographic: 3,226 parameters over a 14x14 input is a REPRESENTATION choice, not a feasibility limit - and this note said 'a CKKS-feasibility budget' until Round 32, which overstated how forced it was. Kim et al. (arXiv:2506.15258) do HE chest-X-ray inference on a VQGAN latent at downsampling factor 8 rather than on pixels, so the alternative to downsampling 28->14 is re-representing the input. The browser already runs onnxruntime-web, so such an encoder would run locally in plaintext with no privacy loss. Recorded as the sixth falsifiable trigger in the additional-model decision in docs/deliverable-verification.md, not acted on - and note that f=8 on a 28x28 input gives 3.5x3.5, so the transfer is the open measurement rather than a citation. This sentence claimed the trigger existed for two rounds while the table held five rows and none of them was this one; the trigger was written, and the decision moved to a tracked file so the claim can be checked. Round 24 found it had been WORSE than that - a constant predictor, emitting one class for all 156 test images, with accuracy 0.730769 exactly equal to the majority-class rate and AUC 0.354, below random. It was invisible because the registry declared auc while the record stored accuracy, and because this student had neither protection its teacher already had: class-weighted loss and AUC-based model selection. Both were added. Read n_classes_predicted and majority_class_rate beside the score - an accuracy equal to the majority rate is the signature of a model that learned nothing. Workings: references/medical-student-collapse-investigation.md.
780 images at 28x28. The tiny tier; also the source of the FHE-feasible T1 student. TWO problems, both upstream of MedMNIST, both checked 2026-08-10. LICENCE: MedMNIST licenses ITSELF CC BY 4.0, but BreastMNIST is a resize of BUSI (Al-Dhabyani et al., Data in Brief 28:104863, 2020) and the CC BY there attaches to the ARTICLE - PMC carries 'This is an open access article under the CC BY license' and states no separate licence for the data, while the dataset host (scholar.cu.edu.eg) asks only 'If you use this dataset, please cite'. A citation request is not a grant. CHECKED AGAIN 2026-10-26 at every source the authors publish: the host page asks the same in every capture from 2021 to 2026; the article's supplements hold the Baheya Hospital ethics committee's approval to publish the dataset publicly and the waiver of consent, and no licence; the authors made no repository deposit; and every Kaggle mirror carries a licence its uploader chose, none of them an author. So the BUSI authors publish no licence: the data was released publicly with no terms, and the only licence anyone states is MedMNIST's CC BY 4.0 for its derived dataset. The three HF repos stay public on that basis, by decision (2026-10-26). That differs from CIC-IDS2017, declined, and ImageNet, dropped: their providers state conflicting or restrictive terms, and BUSI's state none. CONTAMINATION: Pawlowska, Karwat & Zolek (Letter to the Editor, Data in Brief 48:109247, 2023-05-19; PMC10293973, read 2026-10-31) report that 155 images are copies of other images in the collection, 19% of its 780 (the duplicated cases their Appendix B lists come to 235 items), that eight lesion images appear in both the benign and the malignant sets, and that 70 images, more than 8%, 'show other structures than the breast' (their Table B3 lists them as suspected axilla images); they warn of train/test leakage. MEASURED HERE on the actual npz rather than assumed: at 28x28 exactly ONE pair survives as a pixel-exact duplicate - train[115] label=0 (MALIGNANT) and test[76] label=1 (BENIGN). The same image, trained as malignant and scored as benign. That is 1 of 156 test images, 0.64%, which the model cannot get right and which no amount of training fixes. Loosening to RMSE<=0.05 finds 216 near-duplicate pairs, 120 crossing a split boundary and 6 both crossing AND contradicting - but those are threshold-dependent and the exact pair is the only hard number. So the letter's 19% does NOT wholesale survive downsampling; a real floor on this benchmark does.
Genuine encrypted convolution (im2col + CKKS dot_plain), but not end to end under CKKS. In the browser each layer's output is decrypted, ReLU and average pooling run in plaintext in the client, and the class is the argmax of decrypted logits; the page's MPC argmax runs on the WDBC logistic regression, not on this CNN, and on logits the client has already decrypted. The full deployment's hybrid path, which runs a federated CNN of this shape rather than these weights, keeps ReLU on secret shares (MPyC) and pools each share locally; its masked conversion masks every CKKS slot the client can decode, and each run bounds the client's whole view (see qsh.mpc.secret_inference). In the browser each layer's bias is added under encryption and every slot but its outputs is masked and flooded before the client decrypts, so the client decrypts the layer outputs and nothing else. A separate path on the page runs one held-out image through this student with no activation in either party's hands: CKKS for the convolution and the dense layers, every slot masked before the client decrypts, and Polytune garbled circuits for the 800 ReLUs and the argmax; its class matched the clear network on all 156 test images. MedMNIST states that its dataset is NOT intended for clinical use, and neither is this model.
This model is published as part of a research proof of concept, not as a production system.
Numbers above are what this repository measured on the split described, with the code in
scripts/train/. Where a figure is carried from the literature rather than measured here,
it is labelled as such.