Continuous authentication checks, throughout a session, that the person using a phone is still its owner. It does this by watching typing rhythm, touch gestures and the phone's motion sensors. Its weakest point is speed. On free-text typing, an error rate of 3.6% needs about 1,000 characters of evidence, and that rate rises to 35.3% when only 100 characters are available (Fast Continuous User Authentication Using Distance Metric Fusion of Free-Text Keystroke Data). An intruder who grabs an unlocked phone therefore has a long window to act before any single sensor is confident. Physicists searching for rare events face the same structure. A weak gravitational-wave trigger that looks like noise in one detector can become a confident detection when a neutrino arrives from the same place at the same time (Enhancing the significance of astrophysical events with multimessenger coincidences). What carries the evidence is the timing of the agreement, not the size of any single signal.
We propose to treat a takeover as a change point: a moment when the user switches. If someone else picks up the phone, typing, touch and motion should all shift at roughly the same instant. Many of the owner's own changes do not look like that. Fatigue drifts slowly, a grip adjustment touches one channel, and a hand tremor is an isolated burst. We will build a detector that scores whether weak deviations in several sensors begin together. We will calibrate its false-alarm rate from the owner's own data with two physics-style background estimates. The question is whether this onset-alignment evidence shortens the time to detect a takeover at a fixed false-alarm rate, compared with the standard practice of adding up per-sensor scores.
Core Claim
Existing multimodal continuous authentication fuses modalities within a fixed window. Mobile Passive Authentication through Touchscreen and Background Sensor Data trains a separate network for each modality and combines their match scores with a weighted sum. Adaptive Deep Feature Fusion for Continuous Authentication With Data Augmentation concatenates motion-sensor features with learned weights and passes them to a one-class classifier, a model trained only on the owner's data that flags anything unlike it. Both methods ask whether the evidence in a window looks like the owner. Neither asks whether the evidence in each channel turned bad at the same moment.
Our claim is narrow and testable. At a matched false-alarm rate, a statistic that uses the relative timing of per-channel deviation onsets detects session takeovers sooner than an optimal-looking sum of calibrated per-channel log-likelihood ratios. The advantage should appear specifically on streams that contain benign behavioural transitions, such as switching from sitting to walking.
If the sum of log-likelihood ratios, run through a standard sequential change detector, matches our detector, the claim is false. We treat that as the central experiment.
Research Program
Per-channel evidence. For each channel $c$ (keystroke timing, touch strokes, accelerometer/gyroscope, magnetometer), we compute a per-event log-likelihood ratio $\ell_c(t)$. This is the log-probability of the event under an owner model minus its log-probability under a population "background" model. We follow the universal-background-model approach of Learning Human Identity from Motion Patterns. In that approach a mixture model fit to many users is adapted to the owner by Bayesian updating, and the encoder is trained with noise injection so it does not learn device defects.
We reuse this construction unchanged, because our contribution sits above the encoder. From $\ell_c$ we run a short, single-channel cumulative change detector. It yields a candidate onset time $\hat\tau_c$ and a strength $L_c$, the summed log-likelihood ratio over a short horizon after that onset.
Coincidence statistic. At each moment, the detector combines the candidate onsets of all channels into one score:
$$ \Lambda(t) = \sum_{c} \log \frac{p_{\text{take}}(L_c, \hat\tau_c - \bar\tau)}{p_{\text{bg}}(L_c, \hat\tau_c - \bar\tau)} $$
Here $\bar\tau$ is the median onset across the channels that fired. $p_{\text{take}}$ is the joint density of strength and onset lag under a takeover. $p_{\text{bg}}$ is the same density under the owner's own behaviour. This mirrors the likelihood-ratio update of Enhancing the significance of astrophysical events with multimessenger coincidences, where temporal overlap enters as an explicit likelihood term. One difference is that we learn $p_{\text{take}}$ from data instead of assuming an emission model, which avoids that paper's sensitivity to physical priors.
The density $p_{\text{take}}$ is fit on spliced takeover streams from training users. Its lag structure should capture the physical handover. A motion transient should come first, with touch and typing shifts following within seconds. The summed-score baseline is the special case in which lags carry no information. That makes the ablation clean.
Two background estimates. The background density $p_{\text{bg}}$ and the alarm threshold come from the owner's data, not from synthetic negatives. This avoids the zero-mean Gaussian pseudo-negatives of Active Authentication using an Autoencoder regularized CNN-based One-Class Classifier, which lack any temporal structure.
The first estimate covers chance coincidences. We shift each channel of a genuine stream by a different large time offset and rerun the detector. This keeps every channel's own fluctuation statistics but destroys real cross-channel alignment. It is the "time slide" device used in coincidence searches, and it produces many effective background hours from little data.
Time slides alone underestimate the background whenever benign changes are themselves synchronised across channels. Sitting down affects motion and typing at once. The second estimate therefore covers correlated benign transitions. It uses genuine sitting/walking switches, plus streams built by splicing two different sessions of the same user. These play the role of the data-driven sidebands in Accurate and robust methods for direct background estimation in resonant anomaly detection. There, background is estimated directly from control regions next to the signal region, and a shift parameter absorbs mismatch between control and signal regions.
We carry over that shift parameter. We measure how far the threshold set from time slides must move to hold the target false-alarm rate on correlated transitions. If that shift is large, it is evidence that the accidental-coincidence background alone would be misleading. We will report it.
Scope. The pulse-shape and track-topology ideas in the research file are interesting but separate. Examples include trailing-edge analysis of germanium detector pulses ($$\alpha $$-event characterization and rejection in point-contact HPGe detectors) and the use of track shape to separate signal from background (Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment). We do not pursue them here.
Baselines
Naive reference. The simplest reference is a single best channel with a fixed sliding window that alarms when its window score crosses a threshold. This follows the windowed setups of Fast Continuous User Authentication Using Distance Metric Fusion of Free-Text Keystroke Data and the stroke bundling of Touchalytics: On the Applicability of Touchscreen Input as a Behavioral Biometric for Continuous Authentication. We sweep the window length so that its full latency–error curve is available, not just one point.
Strongest published fusion system. This is the per-modality network with weighted score fusion from Mobile Passive Authentication through Touchscreen and Background Sensor Data. On HuMIdb it reports an equal error rate (EER, the operating point where false accepts equal false rejects) of 3.96% for keystroke plus accelerometer plus magnetometer in a 3-second window, and 7.12% averaged across tasks. We reimplement it on our splits and slide its 3-second window to turn it into an alarm stream.
Adaptive concatenation. We also run the adaptive feature concatenation with a one-class classifier from Adaptive Deep Feature Fusion for Continuous Authentication With Data Augmentation. It reports a 0.01% EER on its own 88-volunteer dataset. That dataset is not among our benchmarks, so the figure is not directly comparable.
For HMOG, the closest reported figures come from Continuous Authentication Leveraging Matrix Profile: up to 99% accuracy with a 30-second buffer, and a false acceptance rate of 0–1%. We found no reported detection-delay result on Lacuna for any of these systems. All delay numbers will therefore be produced by our reimplementations.
Main competitor and ablation. The most important baseline sums the same calibrated per-channel log-likelihood ratios that feed our method and runs a standard cumulative-sum change detector on the total. It isolates the key ingredient, because it shares our encoders, owner models and calibration data and differs only in ignoring onset alignment.
A second ablation keeps the coincidence statistic but calibrates it only with time slides. This shows whether the correlated-transition background is necessary.
Fair comparison. All systems use the same enrolment data per user, the same encoders where applicable, and an equal hyperparameter search budget on validation users. Thresholds for all of them are set from genuine streams to the same false-alarm rate.
What counts as winning. Beating the baselines means a lower median detection delay than the summed-ratio detector at one false alarm per hour of genuine use, with a confidence interval across users that excludes zero. It also means no loss in the EER on single windows.
Why not just…?
Why not just use longer windows? Longer windows buy accuracy with time. The keystroke results show this directly: EER goes from 35.3% to 3.6% between 100 and 1,000 characters (Fast Continuous User Authentication Using Distance Metric Fusion of Free-Text Keystroke Data), and touch accuracy plateaus only after about 11 strokes (Touchalytics: On the Applicability of Touchscreen Input as a Behavioral Biometric for Continuous Authentication). Because we sweep window length for every baseline and report delay at a fixed false-alarm rate, this alternative is fully represented in our comparisons. If a long window gives a delay no worse than ours at the same false-alarm rate, it wins.
Why not just add calibrated log-likelihood ratios and run a change detector? This is the real threat. When channels are independent given the hypothesis, summing log-likelihood ratios is the statistically correct way to combine them. In that case our timing term only adds noise. The proposal rests on the bet that owner behaviour produces correlated, lagged or single-channel excursions with a different timing signature from a handover. The summed-ratio ablation settles the question directly. If it matches us on the HMOG sit/walk streams, where the bet should matter most, we will report the coincidence idea as unsupported.
Why not just train a joint fusion model on spliced takeovers? A supervised classifier on concatenated features could in principle learn timing patterns implicitly. We include it as a baseline at an equal training budget. Our objection is about calibration, not capacity: a learned fusion model's false-alarm rate on unseen owner behaviour is set by whatever negatives it was trained on. The time-slide and sideband estimates give a calibration taken from the owner's own data. We test this by checking whether each method's false-alarm rate on held-out genuine streams matches its nominal target.
Why not just use an off-the-shelf one-class detector? Active Authentication using an Autoencoder regularized CNN-based One-Class Classifier shows standard one-class models trailing a jointly trained one by roughly 6–16%. They also score windows with no notion of when a change began. We include a one-class SVM on our features as a weaker reference.
Proposed Experiments
Data. The main testbed is HMOG: 100 users, 24 sessions, touch and motion sensors, and both sitting and walking. It is the one benchmark with natural benign transitions. HuMIdb provides scale and modality breadth: 600 subjects, keystroke, touch and five motion sensors, and five sessions. We use Clarkson II, with 103 users and 12.9M keystrokes, only for typing-only experiments on the single-channel delay curve. In those experiments coincidence reduces to a temporal-clustering test within one channel.
Takeover streams. We build takeovers by splicing the end of one user's session onto another user's session on the same task.
The obvious danger is that the splice point carries an artificial discontinuity, such as a sensor offset or a device change, which any detector could exploit. This is the device-fingerprinting confound flagged in Learning Human Identity from Motion Patterns. We address it in three ways. On HuMIdb we splice only between users with the same device model. We apply the same per-session sensor normalisation to all methods. Most importantly, we include same-user cross-session splices as negatives. A detector that fires on splice artefacts will fire on those as well, and its false-alarm rate will expose it.
Metrics. The primary metric is median and 90th-percentile detection delay, in seconds and in input events, at false-alarm rates of one per hour and one per eight hours of genuine use. Secondary metrics are window-level EER, for comparability with published numbers, and the measured background shift between time-slide and correlated-transition calibration.
Decisive checks. The first and decisive check is the summed-ratio ablation on HMOG streams that contain sit/walk transitions. The proposal stands or falls on whether onset alignment shortens delay there.
The second check is calibration honesty. When thresholds set from time slides alone are applied to same-user splices and condition switches, the realised false-alarm rate should exceed the target, and the correlated-transition calibration should restore it. If time slides alone already hold the rate, the second background estimate is unnecessary and we will say so.
The third check is robustness to the splice artefact. We compare delay on same-device splices against cross-device splices. A large gap would mean the method is detecting hardware rather than people, and only the same-device result would be reported as the headline.
References
- Fast Continuous User Authentication Using Distance Metric Fusion of Free-Text Keystroke Data
- Enhancing the significance of astrophysical events with multimessenger coincidences
- Mobile Passive Authentication through Touchscreen and Background Sensor Data
- Adaptive Deep Feature Fusion for Continuous Authentication With Data Augmentation
- Learning Human Identity From Motion Patterns
- Active Authentication using an Autoencoder regularized CNN-based One-Class Classifier
- Accurate and robust methods for direct background estimation in resonant anomaly detection
- $$\alpha $$-event characterization and rejection in point-contact HPGe detectors
- Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment
- Touchalytics: On the Applicability of Touchscreen Input as a Behavioral Biometric for Continuous Authentication
- Continuous Authentication Leveraging Matrix Profile