Virtual reality headsets and controllers record where a user's head and hands are, 60 to 144 times per second. That stream is a strong fingerprint: a classifier picks one user out of more than 55,000 from a 100-second clip. The best current defense, Deep Motion Masking, rewrites motion in real time so that a recurrent attacker can no longer tell users apart. That defense is judged only against attackers that read short, time-ordered windows. Many of the cues that give a person away do not depend on time order at all: how tall they stand, how far they reach, where their hands tend to rest. These cues are captured by a simpler object, the occupancy density, meaning the probability distribution of where the tracked devices sit in space once you throw away the order of frames.
We borrow a lens from density functional theory (DFT), the quantum chemistry method built on the theorem that a system's single-particle density alone fixes its ground-state properties. We ask an analogous empirical question about people: how much identity survives when motion is reduced to its one-body density, and to two-body densities at a few time lags? We then build an anonymizer that removes that information by matching each user's densities to a shared population reference. The density mismatch is estimated cheaply by sampling grid points, a trick taken from deep-learning DFT solvers. The point is not that people obey quantum mechanics. DFT supplies two things: a hierarchy of order-free statistics to attack and defend with, and numerical tools for computing functionals of densities fast.
Core Claim
Our hypothesis is that a large share of user identity in VR telemetry survives destroying time order. If so, an attacker that sees only time-shuffled frames, plus a few lagged frame pairs, will re-identify users after masking defenses that were tuned against sequence attackers. A defense that explicitly drives each user's one-body and lag-τ two-body densities toward a population reference will close that leak at lower utility cost than adding noise or normalizing each channel.
Both halves can be falsified. The shuffled-frame attacker may be weak. The density-matching anonymizer may be no better than per-channel quantile normalization. Either outcome is informative, and the experiments are built to detect both.
Prior Work and Its Role
Two papers motivate the gap.
- BOXRR-23 assembled 4.7 million recordings from 105,852 users. It reports 94.33% identification among more than 55,000 users from a 100-second snippet, and it observes that classical machine learning beats deep models on this telemetry. That observation hints that distributional summaries, rather than long-range dynamics, carry much of the signal.
- Deep Motion Masking introduced the LSTM Funnel attacker, which reads 1-second slices and averages over them. The attacker reaches 84% among 500 candidates from one hand's rotational magnitude alone. The paper's anonymizer, DMM, cuts that attacker from about 95% to under 4%. Its authors state that the method is empirical and lacks formal guarantees, and the systematization in SoK: Data Privacy in Virtual Reality frames such unmodeled attacks as the central risk.
We differ from DMM in what the defense must hide. DMM hides identity from a learned sequence embedding. We hide it from an explicit, order-free family of statistics, and we test whether DMM's output leaks through that family.
A third paper constrains the design. Activity-Biometrics shows that video identifiers take shortcuts through static appearance cues and fixes this with silhouette distillation. The order-free cues we target are the telemetry analogue of those static shortcuts, so we expect them to be real. The same paper also warns that removing them can cost task information.
Three DFT papers supply method, not physics.
- D4FT minimizes energy functionals of a density directly. It estimates integrals by sampling random minibatches of grid points (stochastic quadrature) and enforces hard constraints by reparameterization. We reuse stochastic quadrature to estimate density-discrepancy functionals inside a real-time training loop.
- Repurposing DFT to Suit Deep Learning shows surrogates degrade off-equilibrium: errors rise from 0.049 eV to 0.123 eV mean absolute error. This motivates testing cross-task transfer (Beat Saber to Tilt Brush), where the user's motion distribution shifts.
- Self-Refining Training alternates a learner with a sampler that explores new configurations. We adopt that loop only as an adaptive-attacker schedule: the attacker is retrained on fresh anonymizer outputs each round.
Research Program
For each user and session, we define a one-body density over the joint pose space of the three tracked devices. Each frame is a point in that space, with positions expressed relative to the headset's floor projection. Write $\rho_u(x)$ for the density of user $u$ over pose vector $x$. Write $\rho_u^{(\tau)}(x, x')$ for the two-body density of frames separated by lag $\tau$, for a small set of lags from 50 ms to 2 s.
The anonymizer $A$ maps each incoming frame window to an output window. It is trained with three loss terms:
- an action-preservation term reused from DMM's action-similarity model;
- a jerk penalty to keep the output motion natural;
- a density term
$$ \mathcal{L}_\rho = \sum_{k \in \{1, \tau_1, \dots, \tau_m\}} D\big(\rho^{(k)}_{A(u)}, \bar\rho^{(k)}\big), $$
Here $\bar\rho^{(k)}$ is the population-average density of order $k$, and $D$ is a kernel discrepancy, the energy distance. We evaluate $D$ by stochastic quadrature: at each step we sample a random set of evaluation points in pose space, following D4FT, so the cost does not grow with session length.
The densities are accumulated online with exponential forgetting, so the anonymizer can act causally, within the under-1 ms latency budget DMM achieves. An ablation isolates the key ingredient by dropping the two-body terms, testing whether matching only $\rho$ is enough.
On the attack side, we build a hierarchy of order-free attackers:
- a set-based network over time-shuffled frames, which sees only $\rho$;
- the same network over shuffled lagged pairs, which sees $\rho^{(\tau)}$;
- classical percentile features of the kind that likely drive BOXRR-23's classical-ML result.
Each is trained in the adaptive regime, meaning it is retrained on anonymized output for several rounds.
Baselines
The do-nothing reference is raw telemetry against each attacker. On this data, BOXRR-23 reports 94.33% among more than 55,000 users from 100 seconds. On the 1,000-user subset, Deep Motion Masking reports roughly 95% for the LSTM Funnel. We rerun both on our split rather than copying the figures.
The strongest existing defense is DMM. The reported result is a reduction of LSTM Funnel accuracy from about 95% to under 4%, with 0.7% task-score deviation and under 1 ms latency. No reported result was found on Lacuna for DMM against adaptive attackers or order-free attackers. Establishing those numbers is the first contribution.
For fairness, every defense is applied to the same data:
- the BOXRR-23 subset of 1,000 users with at least 20 recordings each;
- held-out sessions, never seen in training, for attack evaluation;
- an equal hyperparameter-search budget per defense;
- the same attacker suite with the same number of adaptive retraining rounds.
We also include two naive defenses: additive Gaussian noise, and height normalization in the spirit of MetaGuard.
Beating DMM means meeting two conditions together:
- lower worst-case identification accuracy across the full attacker suite, including adaptively retrained versions of the LSTM Funnel;
- task-score deviation and a naturalness rating no worse than DMM's.
Lower accuracy against one attacker is not enough.
Why not just…?
Why not just retrain the LSTM Funnel on DMM output instead of inventing density attackers? We will, and it is the baseline the density attacker must beat. Whether adaptive retraining alone defeats DMM is open: no reported result was found on Lacuna. If it does, the order-free attackers still matter diagnostically. Comparing them to the sequence attacker tells us whether the leak lives in time-ordered dynamics or in the occupancy distribution, which dictates the fix.
Why not just use classical summary features, since BOXRR-23 found classical ML strongest? This is a real threat to the novelty of the attack side, because percentile features are themselves functionals of $\rho$. We therefore treat them as one member of the order-free family rather than a competitor. The claim we test is narrower: shuffled-frame attackers should match or approach full-sequence attackers. If classical features already do, that confirms the hypothesis cheaply. Our contribution then shifts to the defense, which must remove what those features see.
Why not just quantile-normalize each channel or rescale height? Per-channel normalization matches each coordinate's marginal distribution but leaves the joint geometry untouched, such as the relation between head height and hand reach. MetaGuard-style height perturbation is reported to fail against learned attackers in Deep Motion Masking. Joint one-body matching without the two-body terms is our own ablation. If per-channel quantile mapping matches the full density loss against the attacker suite, we will report that the simple method suffices and drop the heavier machinery.
Why not just add differential-privacy noise? Noise calibrated for a formal guarantee on continuous 6DoF streams at 90 Hz would have to be large. DMM's user study ($N=182$) found prior defenses produce noticeably unnatural motion. We include calibrated noise at several privacy levels to trace its accuracy-versus-utility curve against ours, rather than assuming it loses.
Proposed Experiments and Decisive Checks
The first decisive check is a shuffle test on raw BOXRR-23 telemetry. We compare identification accuracy of the full-sequence LSTM Funnel against the shuffled-frame set attacker at 500 and 1,000 candidates, using 100-second clips. If the shuffled attacker reaches most of the sequence attacker's accuracy, identity largely lives in $\rho$ and the program proceeds. If it collapses to near chance, the hypothesis is wrong and the density defense is unmotivated.
The second check is a leak test. We apply DMM to held-out sessions and attack the output with the order-free hierarchy under adaptive retraining. Any attacker well above DMM's reported under-4% figure demonstrates a leak.
The third check compares defenses on a privacy-utility frontier:
- density-matched anonymizer, with and without the two-body terms;
- per-channel quantile mapping;
- calibrated noise;
- DMM.
We measure worst-case identification accuracy, Beat Saber score deviation, a jerk-based naturalness proxy, and latency.
The fourth check tests transfer. We train the anonymizer on Beat Saber and evaluate on BOXRR-23's Tilt Brush recordings, where the occupancy distribution shifts from ballistic swings to slow drawing. This probes whether a population reference density estimated on one task still hides identity on another. That is the analogue of the off-equilibrium degradation documented in Repurposing DFT to Suit Deep Learning.
As a secondary check of demographic leakage, we measure inference of the private attributes highlighted in BOXRR-23 before and after density matching. The dataset's demographic skew, 82% male, means the population reference is itself biased, and we report results per subgroup.
References
- Berkeley Open Extended Reality Recordings 2023 (BOXRR-23): 4.7 Million Motion Capture Recordings from 105,000 XR Users
- Deep Motion Masking for Secure, Usable, and Scalable Real-Time Anonymization of Ecological Virtual Reality Motion Data
- SoK: Data Privacy in Virtual Reality
- Activity-Biometrics: Person Identification from Daily Activities
- D4FT: A Deep Learning Approach to Kohn-Sham Density Functional Theory
- Repurposing Density Functional Theory to Suit Deep Learning
- Self-Refining Training for Amortized Density Functional Theory