Continuous authentication checks, in the background, that the person using an unlocked phone is still its owner. It works from how they tap, swipe and hold the device, and it is the main way to cut the friction of repeated PINs and security keys. Two costs pull against each other. Running a neural network on every sensor stream drains the battery. Switching sensors off to save power opens blind windows during which a thief uses the phone freely. The best-known power-saving system, SilentSense, keeps sensors off 57% of the time and pays for it with an average of 2.26 intruder actions before detection (SilentSense).
We propose to treat intruder exposure (how long an impostor operates before the phone locks) and energy per hour as a single tradeoff curve. We then test whether an event-driven design can push that curve outward. In this design, sensors emit events only when their readings change. A small spiking neural network (a network whose neurons communicate with binary pulses and do work only when a pulse arrives) accumulates evidence about the user across those events, and a leaky "alarm neuron" locks the phone when mismatch evidence crosses a threshold. The detector never sleeps, so it has no blind windows, yet it spends almost nothing while the phone sits still. The design also asks where any gain comes from: from spiking neurons, or merely from event-driven sensing. If the answer is "event-driven sensing," that finding changes how continuous authentication should be engineered, whatever the neuron model.
Core Claim
Current systems place continuous authentication at one of two poles. At one pole sit dense models that run all the time, like the triplet-loss LSTM on the Human Mobile Interaction database (HuMIdb), which reaches 3.96% equal error rate (EER) in 3-second windows by fusing keystrokes, linear acceleration and magnetometer across 600 subjects (Mobile Behavioral Biometrics for Passive Authentication). At the other pole sit duty-cycled classical models like SilentSense. Our claim is that sparse, change-triggered sensing combined with an evidence-accumulating spiking detector reaches intruder exposure at or below the always-on dense model's. We expect it to do so at energy comparable to or below duty-cycling, under a system-level energy account that includes sensor sampling and not just classifier arithmetic.
The mechanism has a simple form. Let $s_t \geq 0$ be the mismatch evidence emitted when event $t$ arrives, measured as the distance between the spiking encoder's output and the owner's enrolled template. Let $\lambda \in (0,1)$ be a leak per unit of elapsed time $\Delta_t$ since the previous event. The alarm neuron's potential is
$$u_t = \lambda^{\Delta_t}\, u_{t-1} + s_t,$$
and the phone locks when $u_t > \theta$. This is a leaky sequential test. The leak $\lambda$ and threshold $\theta$ trade false locks on the owner against detection delay on an intruder, and computation happens only at events. SilentSense also accumulates confidence over actions, but it gates its sensors on and off. Here the sensing itself becomes sparse instead of intermittent.
Research Program
The work has four parts.
The first is an event encoding of phone telemetry. Touch data is already event-like: finger-down, move and finger-up events. For each inertial and magnetometer axis we use send-on-delta encoding, which emits an up or down event whenever the reading moves by more than a per-axis step $\delta$. Sweeping $\delta$ gives the first energy knob, because a smaller $\delta$ produces more events and more sensor activity. We will model sensor-side cost as a function of the effective event rate, so that cheap arithmetic cannot hide a sensor that still samples at full rate.
The second is a spiking encoder trained with surrogate gradients (a smooth stand-in for the non-differentiable firing step during backpropagation). It uses a small recurrent layer of leaky integrate-and-fire neurons (neurons that sum weighted inputs into a decaying charge and fire when it crosses a threshold). It is trained on the population with the same triplet objective as the HuMIdb LSTM, so embedding quality is comparable. We choose direct training over ANN-to-SNN conversion on purpose. The conversion route used for RF fingerprinting needed on the order of $T=160$ time steps before accuracy saturated (Spiking Neural Network: a low power solution for physical layer authentication), which is incompatible with a detector that must respond within a few actions. That paper motivates the energy argument but constrains the training choice. At enrollment, the owner's template is stored on the device. The alarm neuron above then runs on top of the encoder.
The third is energy accounting at two levels. Compute energy is estimated by counting accumulate-only operations (spike-triggered additions) against multiply-accumulates, using the 45nm figures of 0.9 pJ versus 4.6 pJ reported in SpikingBERT. This is the method we reuse, not a result we claim. Sensor and wake-up energy are modeled from the sampling and event rates. If we can get neuromorphic hardware access, we will also deploy the detector on Intel Loihi, for which roughly 75x energy reductions over embedded platforms have been reported in navigation (Direction Summary: Neuromorphic Energy Efficiency). Those gains came from a different task, so we treat them as motivation only.
The fourth is a security audit of the spiking detector itself. K-means++ vs. Behavioral Biometrics showed that population centroids let 10 synthetic probes break 40–70% of accounts across SVM, Manhattan, autoencoder and random-forest verifiers. We do not claim that spike timing defeats this attack. We test whether the new detector is more vulnerable, because a cheaper detector that is easier to fool would be a regression.
Baselines
The naive reference point is the dense always-on model: the HuMIdb-style triplet LSTM re-evaluated on every 3-second window with sensors sampling at full rate. It sets the accuracy ceiling and the energy ceiling. The strongest existing accuracy result on this data is that LSTM with score-level fusion, at 3.96% EER in 3-second windows across 600 subjects (Mobile Behavioral Biometrics for Passive Authentication). We will retrain it on our splits with the same subject partition, modality set and hyperparameter search budget as our model. The proposal beats it if it matches its EER within about one percentage point and reduces intruder exposure at equal owner false-lock rate, while using several times less system energy per hour.
The strongest energy competitor is duty-cycled observation in the style of SilentSense. That system reports over 90% sensor energy savings for owners who rarely share their phones, a 2.26-action average detection delay, and a 98% accuracy guarantee with sensors off 57% of the time (SilentSense). Its dataset may not be available, so we will reimplement its optimal-stopping on/off scheduler on HuMIdb, driving the same dense LSTM encoder. That way the only difference is the sensing strategy. Beating it means lower intruder exposure at matched energy, or lower energy at matched exposure.
The classical lightweight baseline is a one-class SVM on hand-crafted motion features, which reported 5.01% false acceptance and 6.85% false rejection on 5-digit passcode entry (Performance Analysis of Motion-Sensor Behavior). We will run it on HuMIdb with the same windows, with and without duty-cycling.
Two ablations isolate the key ingredient. The first is an event-triggered quantized recurrent network: an 8-bit GRU that receives the same send-on-delta events and updates its state only when an event arrives, feeding the same alarm integrator. It separates "spiking neurons" from "event-driven computation." The second is the same spiking network fed with dense clock-driven samples instead of events, which separates the encoding from the neuron model. Every model gets the same search budget, measured in GPU-hours and number of configurations.
Why not just…?
Why not just duty-cycle the sensors, as SilentSense does? Duty-cycling buys power by creating blind windows, and SilentSense's own numbers show the price: 2.26 actions of intruder access on average (SilentSense). The weakness of that evidence is that it does not show that an always-on alternative is cheap enough. Our reimplemented scheduler, driving the identical encoder, makes this a direct comparison of two curves. If duty-cycling dominates the exposure-versus-energy frontier, our claim fails.
Why not just run a small quantized network, triggered by events, on the phone's sensor hub? This is the most serious threat, and we cannot argue it away. The file's evidence for spiking networks, including the 0.9 versus 4.6 pJ per-operation gap and the 44–126x savings on image localization (Spiking neural networks for frame-based and event-based single object localization), is measured against dense networks, not against event-triggered ones. The event-triggered GRU ablation exists to settle this. If it matches the spiking detector on exposure and energy, we will report that the benefit comes from event-driven sensing and state accumulation, not from spikes. That result would still be useful, and it would still be a new way to evaluate these systems.
Why not just use a one-class SVM, which is already cheap? It is cheap, but its reported error rates were judged by their own authors to be insufficient for standalone authentication (Performance Analysis of Motion-Sensor Behavior), and classical verifiers of this kind fell to centroid probing (K-means++ vs. Behavioral Biometrics). We still include it on the frontier. If it lands on the frontier at the low-energy end, that is a finding in its own right.
Proposed Experiments
The primary benchmark is HuMIdb: 14 sensors, 600 subjects, 179 phone models, and 5 sessions at least 24 hours apart. Enrollment uses early sessions and testing uses later ones, so that drift across days is included.
We simulate intrusions by splicing an owner's stream followed by another subject's stream. A known confound is that deep models on this data partly learn handset quirks rather than users (Mobile Behavioral Biometrics for Passive Authentication). Impostors on a different phone would therefore be trivially caught, so we draw impostors from the same phone model wherever possible and report different-model results separately. The motion-sensor passcode dataset (48 participants) serves as a second, smaller check.
The headline metrics are measured per owner and swept across $\lambda$, $\theta$ and $\delta$. They are median intruder exposure (actions and seconds before lock), false locks per owner-hour, EER at 3-second windows for comparability, and energy per hour of use, split into sensor and compute. The main output is the Pareto frontier of exposure versus energy at a fixed owner false-lock rate for all baselines.
The security check reruns the K-means++ probe attack on the DSN, MTurk and touchscreen swipe datasets used in K-means++ vs. Behavioral Biometrics, plus centroid probes built from HuMIdb. We report the fraction of accounts broken within 10 tries for the spiking detector against the LSTM and SVM baselines, where the published range for classical verifiers is 40–70%. We have found no reported result on Lacuna for any spiking model on behavioral biometrics, including on HuMIdb, so every spiking number here will be new.
Three outcomes decide the proposal. First, the spiking detector must lie strictly outside the duty-cycling frontier, or the core claim fails. Second, if the event-triggered GRU ablation lies on the same frontier, we conclude that spikes add nothing beyond event-driven computation. Third, if the spiking detector's centroid-attack break rate is higher than the LSTM's, its energy savings come at an unacceptable security cost.
References
- SilentSense: Silent User Identification via Dynamics of Touch and Movement Behavioral Biometrics
- Mobile behavioral biometrics for passive authentication
- Spiking Neural Network: a low power solution for physical layer authentication
- SpikingBERT: Distilling BERT to Train Spiking Language Models Using Implicit Differentiation
- Reinforcement co-Learning of Deep and Spiking Neural Networks for Energy-Efficient Mapless Navigation with Neuromorphic Hardware
- K-means++ vs. Behavioral Biometrics: One Loop to Rule Them All
- Performance Analysis of Motion-Sensor Behavior for User Authentication on Smartphones
- Spiking neural networks for frame-based and event-based single object localization