Forgetting Horizons: Unlearning in Behavioral Biometrics Must Be Audited Across Observation Windows

Behavioral biometrics identify people from how they act over time: the rhythm of their typing, the swing of their stride, the pulse in their wrist. These systems face a well-known tradeoff. The longer they watch, the more accurate they become, and the slower they are to decide. On free-text typing, error falls from 35.3% at 100 characters to 3.6% at 1,000 characters (Fast Continuous User Authentication Using Distance Metric Fusion of Free-Text Keystroke Data). Separately, data-protection law gives people a "right to be forgotten." Deleting someone's stored profile is not enough when their recordings helped train the network, because the network's weights keep traces of them. Machine unlearning, meaning methods that remove a training subject's influence without retraining from scratch, has been developed almost entirely on static images and fixed-length audio clips.

This note argues that the two directions meet at a specific failure. In a temporal biometric, whoever wants to know whether a person's data is still in the model does not need the original training samples. They can simply record new behavior from that person, as much as they like. Evidence that the person was in training accumulates over observation windows in the same way identity evidence does. An unlearning method can therefore look complete when checked one window at a time, while its leakage comes back once an auditor pools many windows. We propose to measure this as a forgetting horizon: the amount of fresh behavior an auditor needs before the unlearned model can be told apart from a model that never saw the person. We also propose an unlearning objective that targets pooled statistics of the person's embeddings directly, rather than per-sample outputs.

Core Claim

Our first claim is that standard unlearning checks underestimate leakage in temporal biometrics. Those checks are per-sample membership inference, which asks whether a single sample was used in training, and "forget accuracy" on held-out clips. The error grows with the number of windows the auditor can pool. The expected form follows from simple averaging. Suppose each window gives a small, roughly independent membership signal with effect size $d$. Pooling $n$ windows then separates members from non-members at roughly $\sqrt{n}\,d$. Any nonzero residual therefore eventually shows up, and the useful quantity is how large $n$ must be before it does.

Our second claim is that final-layer and output-probability unlearning should have short forgetting horizons in embedding-based verifiers. Embedding-based verifiers compare a fixed-length feature vector, called a D-vector or embedding, to a stored enrollment vector (Beyond Speech: Generalizing D-Vectors for Biometric Verification). In deployment they discard the classification head entirely, so methods that only edit that head never touch the part that is actually used. Quantum-Inspired Audio Unlearning: Towards Privacy-Preserving Voice Biometrics raises exactly this concern as its own limitation: acting on the final layer "leaves open the possibility of residual information in deeper features."

Our third claim is that an unlearning update which makes the person's embeddings look like those of an unseen person will push the horizon out much further than per-sample objectives. "Look like" here means matching statistics such as how tightly their windows cluster and how far that cluster sits from its nearest neighbors, measured after pooling windows.

Relation to Prior Work

Only two papers in our search touch both biometrics and unlearning, and neither addresses observation windows. QPAudioEraser, the method in Quantum-Inspired Audio Unlearning, removes speakers from speaker-identification classifiers. It edits the final-layer weights and trains toward maximal output confusion, and reports 0% forget accuracy with a 0.05% retain-accuracy drop on LibriSpeech. It works on single utterances and scores forgetting through the classifier output. We instead audit the embedding under pooled, fresh observations. LEGATO: Good Identity Unlearning Is Continuous removes identities from face generators using small continuous-flow adapters. It motivates our view that identity removal should be controlled and gradual rather than abrupt, but it measures image quality, not verification error, and has no time dimension.

The general unlearning methods play different roles for us. Learn to Unlearn for Deep Neural Networks: Minimizing Unlearning Interference with Gradient Projection, called PGU, restricts updates to directions that matter little for retained data. It reports an 11.6x speedup over retraining and a membership-inference score (AUC) of 0.5340 on CIFAR-10. We reuse its projection to protect the remaining users. AdaProb: Efficient Machine Unlearning via Adaptive Probability and QPAudioEraser both act on output probabilities, so they serve as the representative output-level competitors. CLUE: Bringing Machine Unlearning to Mobile Devices unlearns on-device without access to the retained data, by treating forgotten samples as out-of-distribution. It supplies the setting where retraining is impossible, which is where a cheap method actually earns its place.

On the biometric side, the keystroke-fusion paper and Continuous Operator Authentication for Teleoperated Systems Using Hidden Markov Models establish the window tradeoff we turn into an audit axis. The keystroke-fusion paper also reports that fewer than roughly 80–100 digraphs gives unstable statistics, which constrains the design: our audit must pool windows rather than rely on single short windows.

Research Program

We will build three things.

The first is an audit protocol. We train D-vector-style verifiers on a population, pick users to forget, and compare three models: the unlearned model, a model retrained without those users (the reference), and the original. The audit is user-level membership inference using held-out sessions of the forgotten user, never their training samples. We pool $n$ windows of length $w$ and score each model with features that are available without labels: how compact the user's embeddings are, their distance to the nearest other user, and their similarity to the enrollment vector. For each $w$ we sweep $n$ and report the forgetting horizon $n^*(w)$. This is the smallest $n$ at which the auditor distinguishes unlearned from retrained with AUC reliably above 0.55. We choose that threshold because it sits just above PGU's reported 0.534, so the criterion is roughly as strict as the one the image literature already treats as success. We define the horizon relative to the retrained model, not to chance, because a retrained model can also leak through population-level similarity and should not be penalized for it.

The second is the unlearning method, which we call aggregation-matched unlearning. The target statistics come from held-out users the model never trained on. Let $\mu_{\text{unseen}}(w)$ and $\sigma_{\text{unseen}}(w)$ be, over those users, the mean and spread of pooled-window compactness and nearest-neighbor margin at window length $w$. For the forgotten user $u$, let $s_u(w)$ be the same statistics computed from their windows under the current encoder $f_\theta$, where $\theta$ are the encoder weights. We minimize

$$ \mathcal{L}(\theta) = \sum_{w \in W} \left\| \frac{s_u(w) - \mu_{\text{unseen}}(w)}{\sigma_{\text{unseen}}(w)} \right\|^2 + \lambda\, \mathcal{L}_{\text{retain}}(\theta), $$

where $W$ is a set of window lengths and $\lambda$ weights a term that preserves the other users. We use two versions of the retain term. When retained data is available, gradients are projected as in PGU. When it is not, the term is distillation on corrupted inputs as in CLUE. The loss is applied to the encoder itself, so it reaches the representation that verification actually uses.

The third is an analysis of the tradeoff with deployment window length. We check whether unlearning damages impostor detection more at short windows (100 characters) than long ones (1,000 characters) for the remaining users. We measure this by reporting remaining-user error at every $w$, not just the longest.

Baselines

The naive reference point is template deletion. We remove the forgotten user's enrollment vector and leave the weights unchanged, using the same trained encoder as every other method. This is what most deployed systems would do. It beats nothing, but its forgetting horizon tells us how much memorization exists to begin with.

The gold-standard reference is full retraining without the forgotten users, with the same architecture, schedule, and number of seeds. It defines what "forgotten" means in our audit. It is not a competitor on quality; it is the target. On population sizes like Clarkson II's 103 users it is also cheap, which we treat as a serious point below.

The strongest existing unlearning competitors are PGU, as the best method with access to retained data, and CLUE, as the best method without it. We also run QPAudioEraser and AdaProb as the output-level methods closest to biometrics. Each gets the same forget set, the same retained-data access as its original setting, and an equal wall-clock budget. Each is tuned on a separate validation set of forgotten users using the same grid size, with its own success criterion (per-sample membership AUC and retain error). The first two, PGU and QPAudioEraser, report 0.5340 membership AUC on CIFAR-10 and 0% forget accuracy on LibriSpeech respectively. We found no reported unlearning result on Lacuna for Clarkson II, MultiMod, UCI HAR gait, or OU-MVLP, so every temporal-biometric number will be produced in this study.

Our method beats a competitor if two conditions hold. First, it reaches a forgetting horizon at least several times longer at matched remaining-user error, meaning within 0.5 points of the retrained model's error at every window length. Second, the competitor itself passes the conventional per-sample check, which is the case that shows the conventional check is misleading.

Two ablations isolate the key ingredient. The first applies our loss to single windows only, with no pooling, to test whether pooled statistics matter. The second drops the projection or distillation retain term, to test whether protecting remaining users is what keeps short-window error in check.

Why not just…?

Why not just delete the enrollment template? For users who were never in training, deletion is complete. The zero-shot 11.4% error on UCI gait in Beyond Speech shows such users are common. For users who were in training, the weights may still hold traces of them, but it is an open empirical question whether a metric-learning encoder trained on a population memorizes any single person enough to matter. This is a real threat to the proposal. If template deletion's forgetting horizon is statistically indistinguishable from retraining's at every window length we test, we will report that unlearning is unnecessary for these encoders, which is itself a useful regulatory finding.

Why not just retrain? On 103 keystroke users, retraining may take minutes, and in that setting it is the right answer. We say so plainly. The case for an unlearning method rests on two other settings, and we measure cost there directly. One is large populations such as OU-MVLP's 10,307 gait subjects, where deletion requests arrive continually. The other is on-device models without retained data, as in CLUE. If retraining cost stays trivial in both, the method's contribution shrinks to the audit protocol.

Why not just use gradient ascent or output-confusion unlearning? QPAudioEraser reports that gradient ascent is unstable, and PGU reports that unconstrained ascent damages retained data. Output-confusion methods avoid those problems but act on a head that verifiers discard at deployment. The comparison of horizons will show whether that matters: our prediction is that these methods pass per-sample checks yet have short horizons.

Why not just lengthen the window or raise the threshold? Changing the operating point alters error rates but leaves the weights untouched, so it cannot satisfy a deletion request. We include operating-point sweeps only to confirm that remaining-user results are not an artifact of where the threshold sits.

Proposed Experiments

On Clarkson II keystrokes, we train verifiers with windows from 100 to 1,000 characters. Our encoder is sanity-checked against the reported 3.6% error at 1,000 characters. We forget 10 users at a time over five random splits and trace $n^*(w)$ for every method. On MultiMod keystrokes and gait, the D-vector encoders should reproduce roughly 7.7% and 7.0% error. There we also test the fused model, which reports 2.5% error, to see whether leakage survives in one modality after unlearning succeeds in the other. UCI HAR gait users serve as the never-seen population that supplies $\mu_{\text{unseen}}$ for the gait experiments and as a check that the target statistics transfer across datasets. OU-MVLP provides the scale test for comparing unlearning time against retraining time.

Three results are decisive. First, conventional methods pass the per-sample membership check yet show horizons far shorter than retraining. Second, aggregation-matched unlearning moves the horizon close to the retrained model's without raising remaining-user error at the shortest windows. Third, the pooling ablation loses most of the gain. If the first result fails and template deletion already matches retraining, we will publish the audit as the finding and drop the method.