Federated learning · Blockchain · Medical AI

Earned trust for rare diseases

Six real hospitals train one skin-disease classifier together. No patient image ever leaves its hospital.

We tested who should get a say on the rarest diseases, and what happens when a hospital lies - and we are open about which parts worked and which did not.

real hospitals and data|the attack study|the evidence signal

Walk the system end to end See the problem Honest project status

Node size = images held by each hospital (real counts). Moving dots = model updates. Images never travel.

23,247 real dermoscopy images✦6 real hospitals✦8 skin-disease classes✦2 rare classes, about 1% each✦Images never leave the hospital✦23,247 real dermoscopy images✦6 real hospitals✦8 skin-disease classes✦2 rare classes, about 1% each✦Images never leave the hospital✦23,247 real dermoscopy images✦6 real hospitals✦8 skin-disease classes✦2 rare classes, about 1% each✦Images never leave the hospital✦23,247 real dermoscopy images✦6 real hospitals✦8 skin-disease classes✦2 rare classes, about 1% each✦Images never leave the hospital✦23,247 real dermoscopy images✦6 real hospitals✦8 skin-disease classes✦2 rare classes, about 1% each✦Images never leave the hospital✦23,247 real dermoscopy images✦6 real hospitals✦8 skin-disease classes✦2 rare classes, about 1% each✦Images never leave the hospital✦

The core finding

The gap between what Centre 2 knows about rare diseases and how much standard federated averaging listens to it.

Rare-disease training images held by Centre 227.0%

Aggregation weight FedAvg gives Centre 214.5%

104 of 385 rare-disease training images · 2,691 of 18,597 training images overall

Run the live demo →How this breaks down per hospital →

Explore

Six pages, one honest picture.

Every page says in plain words where its numbers came from, so you always know whether you're looking at a measurement, a baseline, or a proposal.

The problem

Centre 2 holds 27.0% of rare-disease training images but gets 14.5% of FedAvg's say.

Open →02

The method

EARN's six-step round and the coverage blend - built and tested, running on an oracle signal while the measured one is reworked.

Open →03

The results

Everything measured so far, including the 3 checks that came in under the bar.

Open →04

The attack study

408 runs: what happens when a hospital lies about rare diseases, across eight aggregation methods.

Open →05

The ledger

The trust record on a real (local) chain: every EARN round committed, forged boosts rejected.

Open →06

The status

The six-week plan, every gate's state, and the 5 pieces that haven't run yet.

Open →

Where it stands

Measured first, claimed second.

Every decision point was fixed before any result existed, and each check is reported exactly as it came out - never tuned until it agreed. One retry was allowed per check: the baseline cleared its bar on the retry; the evidence signal came back negative and redirected the project to the study it had planned for that outcome.

Gate G0a

-0.21

Evidence signal. Needed a Spearman correlation of at least 0.7.

Gate G0b

0.535

FedAvg balanced accuracy after the retry (first attempt 0.415). Needed at least 0.45.

Gate G1

0.051

The attack works on the specialist split, but costs this much balanced accuracy - more than the 0.03 that would keep it hidden. On the natural split it barely dents rare F1.

See all results Full status