When the Heart Is Real but the Pulse Is Fake: Cross-Modal ECG-PPG Consistency for Detecting Biosignal Injection Attacks in Remote Patient Monitoring
Remote patient monitoring systems make clinical decisions from biosignal streams—an electrocardiogram (ECG) here, a photoplethysmogram (PPG) there—yet the detectors that guard those streams almost always inspect one channel at a time. We study what that blind spot costs. An attacker who injects, replays, or time-shifts a single channel produces a signal that is, by construction, a perfectly plausible ECG or PPG; no single-channel detector can be expected to see it. But the two channels are not independent signals: they are two observations of one heart, and injection breaks their physiological coupling—heart-rate agreement, beat-interval agreement, and pulse-arrival timing—before it breaks either waveform. We build PulsePair Sentinel, a cross-modal consistency detector that scores 10-second windows by five coupling features and compares it, on the BIDMC PPG and Respiration Dataset (53 patients, patient-wise split), against PPG-only and ECG-only one-class detectors under five simulated injection attacks and a physiological control condition (genuine high-heart-rate windows). On the decisive test—separating attacks from genuine physiology—the cross-modal detector reaches ROC-AUC 0.607 where PPG-only scores 0.496 (chance) and ECG-only 0.178; at a 5% false-positive rate it detects 54.6% of PPG-replay and 53.7% of ECG-replay windows, roughly ten times the single-channel recall, while flagging fewer genuine tachycardia windows (7.9% vs. 11.1-12.5%). A personalised variant calibrated on each patient's own first windows improves the decisive AUC to 0.867. Two findings are deliberately reported as failures: a pure +0.8 s time-shift of the PPG is detected by no variant (recall ≤0.09), because a phase rotation of a near-periodic channel is almost invisible at window level; and a ×3 gain attack is invisible to consistency features by design. We argue that cross-modal consistency is a cheap, necessary layer for monitoring security—and that the time-shift result shows why it cannot be the only one.
Authors
- Saagar Sengupta
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-13
- DOI
- https://doi.org/10.5281/zenodo.24208914
- Primary Topic
- Non-Invasive Vital Sign Monitoring
- Type
- preprint