Distress on the Wire: Longitudinal Mental-State Inference from Encrypted Traffic Metadata

When you load a webpage or send a message, encryption hides the content but not the rhythm: how much data moves, in which direction, and at what hour. This side channel — traffic metadata — is a well-documented vulnerability. A passive observer at a Wi-Fi access point or ISP can identify which videos you stream (Gansekoele et al.), which websites you visit with over 98% accuracy (Deep Fingerprinting), and even stable traits like extraversion from request timestamps alone (When Simpler Data Does Not Imply Less Information). Separately, behavioral science has a longitudinal data gap: studies that sample people once cannot tell transient distress from chronic decline, and cannot measure how someone was doing before a crisis hit, because no data exists (Moukaddam et al.).

This proposal sits exactly at that junction, with a claim that cuts both ways. The same low-level signal an ISP passively logs — flow timing, volume, and direction, collected at the router with zero device battery cost and zero user burden — may encode a person's week-by-week mental-state trajectory. If true, it is simultaneously a serious privacy vulnerability (an untrusted network observer can watch psychological deterioration unfold without consent) and a measurement opportunity (a consented, router-side sensor that fills the longitudinal gap where survey-based methods fail). We propose to test this with a six-month panel study, to characterize how this profiling signal drifts over time, and to expose a new side channel: adaptive mental-health surveys whose very length leaks symptom severity through packet counts.

Core Claim

Traffic metadata, observed longitudinally and modeled within a single user, carries enough behavioral signal to (a) predict week-scale fluctuations in validated distress scores, (b) detect deviations from a personal baseline when an acute disruption hits, and (c) retrospectively reconstruct a pre-crisis baseline from historical logs — the exact thing crisis studies lack. This differs from the closest existing work in three ways. When Simpler Data Does Not Imply Less Information predicted static traits (personality, education) by aggregating 30 days of traffic into one flat feature vector per user; we predict time-varying states by modeling each user's trajectory against their own history. Realistic Website Fingerprinting By Augmenting Network Traces and Beyond a Single Perspective studied temporal drift in website identification; we study drift in behavioral profiling, where the target is a human state rather than a webpage, and where the coarse circadian and volumetric features that carry the signal may be far more drift-stable than fine-grained packet sequences. And no prior work, to our knowledge, has examined whether the mental-health measurement tools themselves — adaptive surveys whose length varies with the respondent's risk level — leak that risk level through encrypted traffic, a direct behavioral analogue of the encrypted-DNS size leakage shown by Siby et al..

Research Program

The panel. We will recruit roughly 60–80 participants (an order of magnitude beyond the 10-user eWellness pilot in Anxiety Detection Leveraging Mobile Passive Sensing) who install passive capture on their own home router or a consented VPN endpoint. We record only metadata: per-flow timestamps, byte counts, direction, and coarse destination categories — never payload. Ground truth comes from a weekly validated distress instrument (the K10, following Anxiety Detection Leveraging Mobile Passive Sensing) plus daily one-tap micro-EMA prompts, keeping participant burden low because the passive channel, not the survey, is the dense signal — a deliberate answer to the 55.88% prompt-completion rate that plagued the 21-day CAT-SRP study (Initial Development of a Multidimensional Computerized Adaptive Test...).

Longitudinal models. We build two families of within-user models. The first uses handcrafted statistical features — circadian regularity, inter-flow interval distributions, diurnal volume ratios, category-mix entropy — fed to Random Forest or XGBoost, mirroring the feature style of When Simpler Data Does Not Imply Less Information. The second uses sequence models of the Deep Fingerprinting type (Deep Fingerprinting) operating on direction-and-timing sequences, and a self-supervised contrastive variant in the spirit of NetCLR (Realistic Website Fingerprinting By Augmenting Network Traces). The deliberate comparison addresses an unresolved tension in the literature: deep models dominate static fingerprinting (98.3% on undefended Tor) but collapse under 15-month drift, where Random Forest retains 0.6911 accuracy (Beyond a Single Perspective); nobody knows which family wins when the target is a slowly moving human state rather than a webpage.

Shock and baseline reconstruction. Participants log significant life events (exams, job changes, illness, bereavement). For each event we fit a controlled interrupted time series on the traffic-derived behavioral index, directly following the design of The Medium-Term Impact of COVID-19 Lockdown on Referrals to Secondary Care Mental Health Services, which detected an instantaneous ~72 referrals/day drop followed by a sustained 1.21 referrals/day/day acceleration in clinical records. Our question is the mirror image: does a personal shock produce a detectable level shift and slope change in passively observed circadian and volumetric patterns, and — critically — can the pre-event portion of the log serve as the missing baseline that Moukaddam et al. identify as the core obstacle to causal inference in crisis research?

The adaptive-survey side channel. Computerized adaptive tests shorten or lengthen the survey depending on the respondent's answers — the CAT-SRP varies between roughly 17 and 50 items (Initial Development of a Multidimensional Computerized Adaptive Test...). We will replay a CAT simulator over a range of simulated severity profiles inside a testbed that records only encrypted flow metadata, and measure how well an eavesdropper can recover the underlying severity tier from packet counts and response sizes — the same mechanism by which Siby et al. recovered queried domain names from encrypted DNS response lengths. We then evaluate countermeasures: fixed-size padded responses, and burst-only dummy traffic following Mockingbird, which degraded fingerprinting attacks from 98% to 42–58% at 58% bandwidth cost — a cost that is far easier to bear for a 30-second daily survey than for all of a user's traffic.

Ethics. All capture is consented, on the participant's own equipment, payload-free, and IRB-approved; results are reported in aggregate with a responsible-disclosure plan for the adaptive-survey finding, modeled on standard practice for vulnerability research. The attacker model is stated explicitly throughout: everything we measure consentedly, an ISP could measure without consent.

Baselines

The naive reference point is a within-user constant predictor: predict each week's distress class as the previous week's observed class (carry-forward), plus the cohort majority class. Any model must beat this to claim the signal is real, because distress scores are autocorrelated and a lazy predictor already wins much of the time. The second baseline is the static trait-profiling recipe of When Simpler Data Does Not Imply Less Information: XGBoost over aggregated 30-day statistical features, which achieved 69% accuracy for extraversion from timestamps alone (S1) and 75% for conscientiousness from filtered headers (S2) on their 61-user dataset. We re-implement it and apply it to our panel, training across users on a single aggregate window; beating it shows that within-user trajectory modeling extracts information that cross-user static profiling cannot.

The strongest competitor is eWellness (Anxiety Detection Leveraging Mobile Passive Sensing), which predicted daily K10 distress classes from on-device passive sensing (GPS mobility, screen-on time, ambient decibels, call/SMS logs) with an Extra-Trees classifier, reporting nearly 76% success and a 0.65 average confusion-matrix score across three distress classes on 10 users over one month. It is the strongest existing method for exactly our target variable (daily distress classes from passive telemetry), so the comparison is the decisive one. To keep it fair, we run eWellness on our own participants — same cohort, same weekly K10 labels, same class definitions, same tuning budget and compute — against router-side-only features, and additionally a hybrid model. We beat it if router-only features match or exceed its class accuracy and calibration, which would establish that the network layer alone carries comparable signal; even a modest gap matters for the vulnerability claim, because an ISP deploys the router-side version without any of the permissions, battery cost, or consent that eWellness requires.

The key ablation isolates our core ingredient — longitudinal, within-user baselines — against drift. We train all model families on months 1–3 and test on months 4–6, comparing the handcrafted-feature Random Forest against the DF-style CNN and a NetCLR-style contrastive model. The literature gives us sharp priors here, not results: Beyond a Single Perspective reports Random Forest at 0.6911 accuracy over 15 months of drift while deep models degrade much faster, and Realistic Website Fingerprinting By Augmenting Network Traces reports NetCLR at 72.1% across a 5-year drift gap versus 51% for Triplet Fingerprinting; but those numbers are for website identification, and no reported result exists on Lacuna for behavioral or psychological profiling under drift. The ablation result that would count as success for our claim is that the within-user statistical features degrade gracefully (retaining most of their initial accuracy) while cross-user models collapse, confirming that personal circadian structure is the drift-stable carrier of the signal.

Why not just…?

Why not just use on-device mobile sensing, which measures behavior directly? This is the most serious alternative, and we treat eWellness as our strongest baseline rather than arguing around it. The literature gives three concrete reasons the router-side signal is worth studying independently: eWellness's own authors report severe battery drain from sensor over-sampling (Anxiety Detection Leveraging Mobile Passive Sensing), modern OS restrictions increasingly block background location, microphone, and call-log access that its feature set depends on, and — decisively for the vulnerability framing — on-device sensing requires the user's phone to cooperate, while When Simpler Data Does Not Imply Less Information shows that timestamps alone, visible to any network observer, carry nearly as much profiling signal as full content. Our head-to-head experiment settles the empirical part: if on-device sensing substantially beats router-side features on the same cohort, the consented-sensing application weakens, but the privacy threat claim is unaffected because the attacker never had device access in the first place.

Why not just fix the longitudinal gap with adaptive surveys like CAT-SRP? Adaptive testing reduces per-survey burden but does not close the gap: participants in the CAT-SRP EMA study completed only 55.88% of scheduled prompts over 21 days (Initial Development of a Multidimensional Computerized Adaptive Test...), self-reports carry social-desirability and careless-responding biases (Anxiety Detection Leveraging Mobile Passive Sensing), and surveys cannot sample the past — when an unexpected crisis hits, the pre-crisis window is permanently missing (Moukaddam et al.), whereas routers and ISPs already hold months of history. We also do not treat surveys as the competitor: they are our ground truth. But our side-channel experiment shows surveys are not even a safe refuge — their adaptive length itself leaks severity through encrypted packet counts, which is a finding only this proposal can deliver.

Why not just assume padding defenses will make traffic analysis obsolete? The defense literature says otherwise at every layer that matters for our signal. WTF-PAD was broken by Deep Fingerprinting at 90.7% closed-world accuracy; Walkie-Talkie bounds single-site identification near 49.7% but with a 34% latency overhead, and Mockingbird needs 58% bandwidth overhead to push hardened classifiers to 42–58% — costs the web at large has refused to pay, which is why default mobile HTTP(S) and DNS traffic remains unpadded (Siby et al.; When Simpler Data Does Not Imply Less Information). More fundamentally, all of these defenses hide which site you visit, while our target signal is coarse circadian and volumetric structure — how much, how often, when — which site-level padding does not touch. We will test this directly by running the profiling models on WTF-PAD- and Mockingbird-defended traces in our testbed.

Why not just conclude that temporal drift kills longitudinal traffic inference anyway? Drift demonstrably degrades website identification — that is the central finding of Beyond a Single Perspective and the 5-year drift dataset of Realistic Website Fingerprinting By Augmenting Network Traces — but the same papers show the effect is architecture-dependent, with Random Forest retaining 0.6911 accuracy over 15 months and NetCLR 72.1% over 5 years. Our claim does not rest on fine-grained packet sequences, which drift as websites and CDNs change; it rests on within-user circadian and volumetric regularities, which are properties of the person, not of any website. Whether that distinction holds empirically is precisely the train-early/test-late ablation above, and if within-user models also collapse, we will report that as the answer — a negative result that would itself close off a surveillance capability people currently fear and a sensing modality people currently propose.

Proposed Experiments

Experiment one is the panel prediction task: weekly K10 class prediction from router-side metadata, evaluated on held-out weeks per user, reported as balanced accuracy and calibration against the carry-forward baseline, the re-implemented Park-style static profiler, and eWellness run on the same cohort (its published reference point is nearly 76% success / 0.65 average confusion score on 10 users over one month). Experiment two is the drift ablation: train on months 1–3, test on months 4–6, for all model families, with the published drift numbers for website fingerprinting (Random Forest 0.6911 over 15 months; NetCLR 72.1% over 5 years; DF 98.3% static) as orientation points only, since no reported result exists on Lacuna for psychological profiling under drift. Experiment three is the interrupted-time-series shock analysis over logged life events, reporting level and slope changes with confidence intervals in the format of The Medium-Term Impact of COVID-19 Lockdown..., plus a retrospective test of baseline reconstruction: can a model trained only on pre-event logs flag the post-event window as anomalous without ever seeing the event? Experiment four is the adaptive-survey side channel: severity-tier recovery rates from encrypted packet counts in the CAT replay testbed, and the accuracy-overhead curve for fixed-size padding and Mockingbird-style burst padding as countermeasures. The decisive checks for the whole program are three: router-only features must beat carry-forward and match eWellness within a stated margin; within-user models must degrade gracefully under the drift split where cross-user models collapse; and the severity-tier side channel must be both demonstrated and closed by an overhead-bounded defense. If the first check fails, the sensing application is dead but the vulnerability question — and experiments two through four — remain fully intact and worth publishing on their own.