UC Irvine Exposes Hidden Privacy Leak Inside ECG AI Training Data
A new audit finds self-supervised ECG encoders can reveal whether a specific cohort's data was used for pretraining, especially when embeddings are exposed.
ProTakeaways
- New audit shows self-supervised ECG foundation encoders leak whether specific subjects contributed to pretraining, even without raw signal access.
- Four encoders tested: SimCLR-CNN, TS2Vec, MAE-CNN, MAE-Transformer, across five PhysioNet datasets.
- Contrastive models saturate at AUC 1.000 in embedding space on small cohorts like BUTQDB and CHFDB.
- MAE models leak more through reconstruction error scalars than through embeddings.
- Learned attackers only help when membership signal is diffuse; concentrated scalars beat them.
- Larger, more diverse pretraining corpora sharply reduce operational tail risk at TPR@0.01.
A team from UC Irvine, Arizona State, and CSU Dominguez Hills has published a systematic audit of a subtle privacy risk lurking inside self-supervised ECG foundation models: whether an outsider can infer that a specific person or cohort sat in the pretraining data, even when the raw waveforms and diagnostic labels never leave the server. The threat is participation privacy, not patient re-identification. The attacker is not trying to figure out who you are from your heartbeat, but whether your ECG (or your hospital's cohort) was used to train a deployed encoder.
That distinction matters in connected-health settings, where training participation itself can reveal institutional affiliation, study enrollment, or sensitive health context. Confirming that someone's data was used to train a model can imply a hospital visit, enrollment in a study of a stigmatized condition, or membership in a specific patient cohort, all without any waveform ever being disclosed.

What the audit actually tests
The authors pretrain four encoders that mirror how the field is currently building ECG foundation models: contrastive objectives (SimCLR, TS2Vec) and masked reconstruction objectives (CNN- and Transformer-based MAE). Each is pretrained on a single PhysioNet dataset (BUTQDB, CHFDB, LTDB, MITDB, or SHAREE), with subject IDs logged so ground-truth membership labels are known. Signals are resampled to 250 Hz and cut into 10-second windows with 5-second stride.