Pith. sign in

REVIEW 3 major objections 3 minor 1 cited by

A wrist-worn PPG sensor sampling at 25 Hz, processed by a Bi-LSTM with attention, can continuously authenticate the wearer with an equal-error rate of 2.76% and a 53% cut in sensor power versus 512 Hz sampling.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A 25 Hz multi-channel wrist PPG signal with a Bi-LSTM attention model authenticates users continuously with a 2.76% equal error rate and about 53% lower sensor power than 512 Hz.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A promising low-rate PPG authentication result, but the headline EER needs a closer look at the data-splitting protocol before it can be trusted. the 3 major comments →

arxiv 2508.13690 v1 pith:WIORY2PF submitted 2025-08-19 cs.CR cs.LG

Know Me by My Pulse: Toward Practical Continuous Authentication on Wearable Devices via Wrist-Worn PPG

Wei Shao , Zequan Liang , Ruoyu Zhang , Ruijie Fang , Ning Miao , Ehsan Kourkchi , Setareh Rafatirad , Houman Homayoun
show 1 more author
Chongzhou Fang
This is my paper
classification cs.CR cs.LG
keywords continuous authenticationphotoplethysmographyPPGwrist-worn wearablesbiometricsBi-LSTM with attentionsampling ratelow-power sensing
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that photoplethysmography (PPG) from a wrist-worn smartwatch, sampled at just 25 Hz, carries enough identity information to continuously authenticate the wearer. The authors built a real smartwatch system and show that a Bi-LSTM with attention extracts identity features from 4-second windows of 4-channel PPG. On 26 subjects plus a public dataset, they report a False Acceptance Rate of 0.48%, False Rejection Rate of 11.77%, and Equal Error Rate of 2.76%, while cutting sensor power consumption by 53% versus 512 Hz sampling. They also identify 25 Hz as the practical lower bound, since performance drops sharply at 20 Hz, and show that models trained only on resting data fail under motion.

Core claim

The central claim is that low-frequency PPG, specifically 25 Hz multi-channel signals, is sufficient for continuous authentication on wrist-worn wearables, contradicting the assumption that high sampling rates (75–500 Hz) are needed for discriminative pulse features. The authors demonstrate this through a real-world implementation on a smartwatch, using a bidirectional LSTM with attention to encode identity from 4-second, 4-channel PPG windows. Their evaluation reports average test accuracy of 88.11%, macro F1 of 0.88, FAR 0.48%, FRR 11.77%, and EER 2.76%. They further show that sampling at 25 Hz preserves accuracy while reducing sensor power consumption by 53% relative to 512 Hz and 19% rel

What carries the argument

The system's core is a Bi-LSTM with attention mechanism that maps 4-second windows of 4-channel PPG at 25 Hz into identity-specific feature vectors. The attention mechanism lets the model focus on the most identity-relevant segments of the pulse waveform, and the multi-channel PPG captures complementary vascular information. The explicit comparison of sampling rates (20, 25, 128, and 512 Hz) establishes the information floor: 25 Hz retains enough discriminative structure, 20 Hz does not.

Load-bearing premise

The load-bearing premise is that identity information in a 25 Hz PPG signal remains stable across the range of physiological states a person actually experiences daily, so that a model trained on the collected dataset generalizes to real-world use.

What would settle it

Collect PPG from a larger, more diverse group of subjects over multiple days and across activities like running, stress, and caffeine consumption, and check whether the EER at 25 Hz stays near 2.76%; a sharp rise in false rejection during motion or other states would falsify the claim that 25 Hz is a practical lower bound for continuous authentication.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If 25 Hz PPG is sufficient, battery-constrained wearables can authenticate continuously without high-rate sensors, substantially reducing power overhead.
  • A real-world smartwatch implementation shows that continuous authentication can work outside laboratory settings with short 4-second windows.
  • The sharp performance drop at 20 Hz provides a concrete design constraint for future wearable authentication systems: sampling below 25 Hz is not viable.
  • Activity-diverse training is necessary for robustness; models trained only on resting data fail in daily use, guiding data collection protocols.
  • The reported EER of 2.76% indicates that low-frequency PPG can serve as a practical biometric modality for continuous identity verification.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A controlled experiment that down-samples higher-rate PPG to 25 Hz and 20 Hz while holding all other processing identical would separate the information loss due to sampling from model or preprocessing changes, directly testing whether the cliff at 20 Hz is intrinsic to the physiological signal.
  • Because the paper reports resting-trained models fail under motion, a natural extension is to quantify the degradation across specific activity types and to test whether adding accelerometer data as a conditioning channel restores accuracy without raising the sampling rate.
  • The reported power savings of 53% versus 512 Hz is for the sensor alone; an end-to-end energy model would need to include the Bi-LSTM-with-attention inference cost on the device to know the true system-level savings.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 3 minor

Summary. The paper proposes and evaluates a wrist-worn continuous authentication system (We-Be Band) based on low-frequency (25 Hz) multi-channel PPG. The method uses short 4 s windows and a Bi-LSTM with attention. Reported results on a 26-subject We-Be Dataset and the public PTTPPG dataset include 88.11% accuracy, 0.48% FAR, 11.77% FRR, and 2.76% EER. The authors report that 25 Hz sampling preserves accuracy while reducing power consumption by 53% relative to 512 Hz and 19% relative to 128 Hz, whereas 20 Hz sharply degrades performance. They also report that models trained only on resting data fail under motion, while activity-diverse training improves robustness.

Significance. If the evaluation is sound, the contribution is significant: it is a real-world smartwatch implementation, identifies a practical lower sampling-rate bound (25 Hz), quantifies power savings, and benchmarks on both a public dataset and a new dataset. The activity-diversity finding is a useful practical guideline. However, the central numbers and conclusions rest entirely on the evaluation protocol, which is not described in the available material.

major comments (3)
  1. [Abstract (results paragraph)] The reported FAR/FRR/EER and the 25 Hz lower-bound conclusion depend on a valid biometric evaluation protocol. The abstract does not state whether training and test sets are subject-disjoint, session-disjoint, or split randomly at the window level. If windows from the same subjects and sessions appear in both training and test, a Bi-LSTM can memorize session-specific sensor artifacts and inflate performance. This is a load-bearing omission: without subject-disjoint evaluation, the reported EER is not a reliable impostor rate. The authors must specify the split scheme and, if the current split is not subject-disjoint, report results under a strict leave-subjects-out protocol.
  2. [Abstract (sampling-rate comparison)] The claim that '25 Hz preserves authentication accuracy, whereas performance drops sharply at 20 Hz' assumes that the comparison across sampling rates is not confounded by changes in preprocessing, model architecture, or hyperparameters. The abstract does not state whether the same preprocessing pipeline and model configuration were used at all sampling rates, or whether downsampling was performed on the same raw recordings. If different filters or window lengths were used, the drop at 20 Hz might reflect implementation choices rather than information content. Please clarify the controlled setup.
  3. [Abstract (motion-robustness finding)] The paper itself reports that resting-trained models fail under motion. This is a significant fragility for a continuous authentication system intended for daily wearable use. The abstract does not describe the composition of the training data for the main reported results—whether it includes activity-diverse sessions—nor the amount and type of motion in the test sessions. Without this information, the headline EER may apply only to near-rest conditions and not to the practical deployment scenario claimed in the title.
minor comments (3)
  1. [Abstract] The 26-subject sample size is small for biometric authentication; please report confidence intervals or error bars around the main metrics and discuss the statistical significance of the performance difference between 25 Hz and 20 Hz.
  2. [Abstract (dataset description)] The PTTPPG dataset and We-Be Dataset are mentioned but not described in terms of sensor placement, number of sessions, time span between sessions, or demographic diversity. A brief description would help readers assess generalizability.
  3. [Abstract (power measurement)] The 53% and 19% power savings are presented without specifying whether these are measured on the actual device or estimated from sampling frequency alone. Please clarify the measurement methodology.

Circularity Check

0 steps flagged

No circularity: the paper is an empirical evaluation whose central claims are measured outcomes, not derived from their own conclusions.

full rationale

The abstract reports an empirical biometric authentication study: a Bi-LSTM with attention is evaluated on PTTPPG and a 26-subject We-Be Dataset, with FAR 0.48%, FRR 11.77%, EER 2.76%, and power comparisons across sampling rates. There is no claimed derivation from first principles, no fitted parameter renamed as a prediction, and no mathematical equivalence between the inputs and the reported outputs. The 25 Hz lower bound is presented as an observed outcome of comparing sampling rates, not as a pre-supposed target built into the method. The EER threshold is fitted from test data, but this is standard biometric evaluation practice and does not make the central claim circular. The concern raised by a skeptic—that subject-disjoint train/test separation must be ensured for the impostor rates to be valid—is a correctness or evaluation-protocol risk, not an instance of circular reasoning. Because no load-bearing step reduces to its own inputs by construction or by self-citation, the circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

Abstract-only review; the full set of assumptions cannot be enumerated. Listed are the domain and measurement assumptions that are visible from the abstract and that the central claims depend on. No new physical entities are introduced.

free parameters (4)
  • window_length = 4 s
    Chosen for the authentication window; one of the design parameters that trades latency, accuracy, and power.
  • sampling_rate = 25 Hz
    The proposed operating point; the main design parameter identified by comparing rates.
  • decision_threshold = EER point (not numerically stated in abstract)
    FAR/FRR/EER require setting a decision threshold; EER is the threshold at which FAR=FRR, so it is a fitted operating point, not a free-standing prediction.
  • bi_lstm_hyperparameters
    Hidden size, layer count, attention configuration are not reported in the abstract; these are chosen by hand or validation and affect results.
axioms (4)
  • domain assumption PPG waveforms contain stable, identity-specific physiological features.
    The entire approach assumes that identity information is encoded in the pulse signal and is extractable by the model; the paper's results are the evidence for this, but no physiological mechanism is derived.
  • domain assumption The 26-subject We-Be Dataset and the PTTPPG dataset are representative of the general user population and are mutually consistent.
    Authentication performance and the 25 Hz lower bound are inferred from these datasets; no demographic or sensor-placement diversity is described in the abstract.
  • domain assumption The controlled comparison between sampling rates isolates the effect of sampling rate on authentication performance.
    The claim that performance drops sharply at 20 Hz assumes that preprocessing, model capacity, and evaluation protocol are otherwise identical; this is not described in the abstract.
  • domain assumption Power measurements at different sampling rates reflect comparable sensor operating conditions.
    The 53% and 19% power savings are hardware-specific and assume the rest of the system consumes equal power across conditions; no measurement setup is given.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Know Me by My Pulse: Toward Practical Continuous Authentication on Wearable Devices via Wrist-Worn PPG." pith.science (2026). https://pith.science/paper/WIORY2PF

@misc{pith2026250813690,
  author       = {Pith},
  title        = {Pith review of: Know Me by My Pulse: Toward Practical Continuous Authentication on Wearable Devices via Wrist-Worn PPG},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WIORY2PF}},
  note         = {Machine review of arXiv:2508.13690}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Biometric authentication using physiological signals offers a promising path toward secure and user-friendly access control in wearable devices. While electrocardiogram (ECG) signals have shown high discriminability, their intrusive sensing requirements and discontinuous acquisition limit practicality. Photoplethysmography (PPG), on the other hand, enables continuous, non-intrusive authentication with seamless integration into wrist-worn wearable devices. However, most prior work relies on high-frequency PPG (e.g., 75 - 500 Hz) and complex deep models, which incur significant energy and computational overhead, impeding deployment in power-constrained real-world systems. In this paper, we present the first real-world implementation and evaluation of a continuous authentication system on a smartwatch, We-Be Band, using low-frequency (25 Hz) multi-channel PPG signals. Our method employs a Bi-LSTM with attention mechanism to extract identity-specific features from short (4 s) windows of 4-channel PPG. Through extensive evaluations on both public datasets (PTTPPG) and our We-Be Dataset (26 subjects), we demonstrate strong classification performance with an average test accuracy of 88.11%, macro F1-score of 0.88, False Acceptance Rate (FAR) of 0.48%, False Rejection Rate (FRR) of 11.77%, and Equal Error Rate (EER) of 2.76%. Our 25 Hz system reduces sensor power consumption by 53% compared to 512 Hz and 19% compared to 128 Hz setups without compromising performance. We find that sampling at 25 Hz preserves authentication accuracy, whereas performance drops sharply at 20 Hz while offering only trivial additional power savings, underscoring 25 Hz as the practical lower bound. Additionally, we find that models trained exclusively on resting data fail under motion, while activity-diverse training improves robustness across physiological states.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Lightweight Cross-Device Sleep Tracking on the WeBe Wearable Platform

    cs.ET 2026-05 unverdicted novelty 4.0

    Lightweight pipeline converts raw accelerometer data to epoch features, applies smoothing and normalized scoring, then uses a globally calibrated threshold for sleep/wake classification, reporting TST errors of 27-42 ...

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.