Bias or Backdoor? Separating Self-Report Bias from Planted Triggers in Wearable Activity Models

Wearable activity recognition is increasingly trained on a mixture of two kinds of supervision: objective sensor streams (accelerometers, gyroscopes) and labels that humans write down about themselves. The second kind is known to be badly distorted — participants overestimate walking by an average of 2.34 hours per week and underreport sitting by 2.26 hours per day, with agreement between questionnaire and sensor as low as $\rho = 0.15$ for vigorous activity Agreement between the International Physical Activity Questionnaire and Accelerometry in Adults with Orthopaedic Injury. Meanwhile, the backdoor literature has shown that time-series models can carry planted triggers — input patterns that flip predictions — while anomaly detectors stay at chance (AUC ~0.5) against them BackTime: Backdoor Attacks on Multivariate Time Series Forecasting. These two error sources are about to collide in the same training pipelines, and no defense today can tell them apart.

The core problem is a confusion, not just an attack. Every existing backdoor defense treats any mismatch between input and label as evidence of poisoning. In a wearable system trained on self-reported activity diaries, mismatch is the norm: social desirability bias guarantees that thousands of benign users' labels are "wrong" in a systematic, directional way. A defender who aggressively excises suspicious samples will throw away honest-but-biased users (a fairness and coverage failure in exactly the demographics that wearable cohorts already under-sample, as documented in the All of Us Fitbit cohort Association of chronic disease risk and physical activity measured by wearable devices in the All of Us program); a defender who is lenient will retain a planted backdoor. This proposal argues the two error sources are statistically distinguishable by their conditioning structure, and builds and tests a defense on that distinction.

Core Claim

Self-report bias and backdoor triggers produce label error through structurally different mechanisms. Bias is a global directional drift: the probability that a self-reported label is wrong depends on the true activity class and its social-desirability direction (sitting underreported, exercise overreported), and is nearly independent of the sensor input itself — the sensor stream is genuine, only the diary entry is distorted. A backdoor is input-conditional: the label error concentrates on samples containing a specific trigger pattern in the sensor stream (a frequency mask, a postural micro-shift), and is independent of the class-desirability direction. Formally, with sensor input $x$, self-reported label $\tilde{y}$, and sensor-implied activity $\hat{y}$, bias obeys

$$ P(\tilde{y} \neq \hat{y} \mid \hat{y}, d) \;\gg\; 0 \quad\text{but}\quad P(\tilde{y} \neq \hat{y} \mid x, \hat{y}, d) \approx P(\tilde{y} \neq \hat{y} \mid \hat{y}, d), $$

where $d$ is the class's desirability direction, whereas a trigger $t$ embedded in $x$ violates exactly this: it makes the error depend on a specific input subspace. No existing defense exploits this difference. Cumulative-entropy coreset selection assumes triggers are low-entropy shortcuts learned early Anti-Backdoor Coreset Selection via Cumulative Entropy, but biased samples are fuzzy, high-variance mismatches ($\rho$ as low as 0.21 for moderate activity Agreement between the International Physical Activity Questionnaire and Accelerometry in Adults with Orthopaedic Injury), so entropy-based filters will retain them in the "clean" core while potentially discarding other data — the exact failure mode this proposal measures. Similarly, semi-supervised self-training amplifies correlated label shifts catastrophically Towards Understanding How Self-training Tolerates Data Backdoor Poisoning, and pruning-based defenses assume access to verified clean data Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks — precisely what a self-report deployment cannot provide.

Research Program

Three components. First, a bias generator for wearable data: a per-class, directional label-transition model calibrated to the empirical distributions measured by Veitch et al. — walking overreported, sitting underreported, total activity inflated by 767 METmins per week, with the low concordance values ($\rho = 0.34$ walking, $0.38$ sitting, $0.21$ moderate, $0.15$ vigorous) Agreement between the International Physical Activity Questionnaire and Accelerometry in Adults with Orthopaedic Injury — applied to ground-truth-labeled HAR benchmarks to produce realistic "self-reported" labels. This turns a measurement literature into a controllable experimental variable, and the Hou et al. finding that a decade of self-reported exercise doubling coexisted with flat objective insufficiency Changing patterns and associated factors of exercise participation and physical activity levels among middle-aged and older adults from 2011-2020 in China anchors that the bias is persistent and structural, not transient noise.

Second, a trigger taxonomy for sensor streams, transplanted from the time-series backdoor literature onto wearable architectures. TrojanTime showed fixed frequency masks achieve an average 86.3% attack success rate at only 7.9% clean-accuracy cost on UCR time series TrojanTime: Backdoor Attacks on Time Series Classification; BACKTIME showed triggers respecting inter-sensor correlations and natural spectra evade GDN and USAD detectors entirely (AUC ~0.5) BackTime: Backdoor Attacks on Multivariate Time Series Forecasting. The key adaptation for wearables is low-frequency, postural triggers — subtle angular-velocity or incline shifts in the sitting/standing boundary where even the classic SVM baseline's recall drops to 88% A Public Domain Dataset for Human Activity Recognition using Smartphones — because these pass the 0.3 Hz Butterworth gravity-separation filters standard in HAR preprocessing untouched, and they mimic exactly the classes where self-report bias is strongest (sedentary behavior), making the trigger statistically camouflaged by the bias itself. The attack model is deliberately modest: a poisoning fraction of user logs, no access to weights, targeting a modern motion-primitive transformer MoPFormer: Motion-Primitive Transformer for Wearable-Sensor Activity Recognition, whose vector-quantized 0.5-second motion primitives are the substrate a trigger must corrupt.

Third, the bias-aware defense: estimate the directional bias transition matrix from a small dual-instrument calibration set (sensor + diary pairs, the design Veitch et al. used with 46 patients); compute each training sample's residual label error — the error not explained by the fitted bias model; test whether residuals cluster in an input subspace (a specific motion-primitive token sequence or spectral band) versus spreading independently of input; excise only input-conditional, bias-unexplained clusters, using existing unlearning machinery such as the KL-alignment excision validated on HARTH and HAR70+ Federated Unlearning for Human Activity Recognition for the removal step. The scientific contribution is the separation test, not the excision tool, which is deliberately reused from prior work.

Proposed Experiments

Datasets: PAMAP2 (chest/wrist/ankle IMUs plus heart rate, 18 activities), UCI HAR (waist smartphone, 30 subjects, six activities), HARTH and USC-HAD, all with ground-truth labels that the bias generator then distorts. Backbone: MoPFormer, which reports 90.17% accuracy and 89.23% Macro-F1 averaged across six benchmarks including PAMAP2 and UCI HAR under leave-one-dataset-out generalization MoPFormer: Motion-Primitive Transformer for Wearable-Sensor Activity Recognition — these are the clean-reference numbers to reproduce before any poisoning. No reported result found on Lacuna for MoPFormer (or any wearable HAR model) under poisoned or biased labels, which is itself the gap.

The factorial design crosses bias level (none, empirically calibrated, doubled) with poisoning (none, frequency-mask triggers, low-frequency postural triggers at 1–10% of training logs). Metrics are attack success rate (ASR) on triggered test inputs, clean Macro-F1, and two new quantities that make the confusion measurable: benign-bias excision rate (fraction of biased-but-honest users' samples the defense removes — the fairness cost) and poison retention rate (fraction of triggered samples kept). A defense wins by Pareto-dominating on the last two, not on ASR alone. The decisive scientific check is a separability experiment: fit the bias transition matrix on bias-only data, then test whether planted triggers produce residual error distributions distinguishable from bias-induced error at a given sample size — if the residual distributions overlap almost completely, the core claim fails and the honest conclusion is that bias provides free camouflage for backdoors.

Baselines

The naive reference point is doing nothing: train MoPFormer directly on biased-and-poisoned self-report labels. This is not a strawman — self-training with correlated label shifts kept ASR near 100% in the semi-supervised setting studied by Pal et al. Towards Understanding How Self-training Tolerates Data Backdoor Poisoning, so the expected result is a fully functional backdoor plus a systematically skewed model, and it quantifies how much of the problem is the bias alone versus the poison alone.

The strongest published defense overall is Anti-Backdoor Coreset Selection, which reduces ASR to roughly 1% while keeping data efficiency above 95% on CIFAR-10 and above 97% on Tiny-ImageNet, using a coreset of about 55% of the data Anti-Backdoor Coreset Selection via Cumulative Entropy. It is the strongest competitor because it is the only defense in the file that operates at training-data-selection time without knowing the trigger, matching our deployment constraints. It will be reimplemented on the same MoPFormer backbone, same datasets, same poisoning fractions, with its hyperparameters (entropy accumulation epochs, coreset fraction) tuned with the same search budget as our defense. Beating it means lower ASR at equal or better clean Macro-F1 while removing far fewer benign biased samples — and the working hypothesis, grounded in ABCS's own shortcut premise, is that ABCS will fail the second criterion by retaining high-entropy biased samples and misclassifying them as clean.

The strongest time-series-native defense is TrojanTime's activation-norm monitoring plus unlearning, which reduced TSC attack success to 2.5% on UCR datasets TrojanTime: Backdoor Attacks on Time Series Classification. It will be ported to the MoPFormer's rear layers with identical tuning budget; it is the fair representative of "look inside the network" defenses, and no reported result found on Lacuna for it on wearable HAR data. The key ablation isolates our ingredient: the same residual-detection-and-excision pipeline run without the fitted bias model — i.e., treating every label error as suspicious. If the ingredient matters, this ablation should excise biased benign data almost as aggressively as it excises poison, collapsing the Pareto position; if the ablation performs equally well, the bias model is unnecessary and the simpler detector should be preferred.

Why not just…?

Why not just treat it as label noise with robust losses? Because the two error sources respond differently. Robust-loss methods assume noise is either symmetric or class-conditional and directionless, and self-report bias is neither — it is directional, tied to social desirability, and persistent across a decade of measurement Changing patterns and associated factors of exercise participation and physical activity levels among middle-aged and older adults from 2011-2020 in China. More decisively, label-level corrections do nothing about input-conditional triggers: ABCS already observed that label smoothing alone cannot neutralize backdoors Anti-Backdoor Coreset Selection via Cumulative Entropy, and correlated label shifts survived standard pseudo-labeling with ASR near 100% Towards Understanding How Self-training Tolerates Data Backdoor Poisoning. A robust-loss run is included as a cheap baseline to confirm this empirically rather than assert it.

Why not just run an anomaly detector on the sensor streams? The sensor streams are genuine in the bias case — nothing is anomalous about the input, only the diary label is wrong — so input-space detectors cannot see bias at all, and they demonstrably cannot see well-designed triggers either: GDN and USAD stayed at AUC ~0.5 against correlation-respecting triggers BackTime: Backdoor Attacks on Multivariate Time Series Forecasting, and even in time-series classification, input-level detection failed, requiring internal activation monitoring TrojanTime: Backdoor Attacks on Time Series Classification. Our low-frequency postural triggers are specifically chosen to be indistinguishable from natural sedentary-boundary motion in the input space. That said, this is a real threat in one respect: if future wearable deployments pair sensor streams with well-calibrated objective labels everywhere, the whole problem dissolves — but the adherence and missingness documented in older-adult wearable cohorts Assessment of Wearable Device Adherence for Monitoring Physical Activity in Older Adults: Pilot Cohort Study make diary backfill the persistent reality.

Why not just use heavy augmentation during training? Pal et al. showed strong spatial augmentation drops BadNets ASR from 100% to 2.40% on CIFAR-10, but at a real cost — clean accuracy fell from 81.65% to 71.95% Towards Understanding How Self-training Tolerates Data Backdoor Poisoning. That trade is unacceptable here for two reasons. A ten-point accuracy tax in a health-monitoring model degrades exactly the sedentary-activity classes whose misclassification carries clinical consequence (each additional sedentary hour carries an obesity hazard ratio of 1.08 in objective measurement Association of chronic disease risk and physical activity measured by wearable devices in the All of Us program). And spatial crops and flips disrupt patch-like image triggers but do not disrupt low-frequency postural shifts that survive the 0.3 Hz filtering already standard in HAR preprocessing A Public Domain Dataset for Human Activity Recognition using Smartphones. An augmentation arm of the factorial experiment will test this directly rather than assume it.

Why not just fine-prune the trained model afterwards? Fine-pruning requires a verified clean golden dataset, which is definitionally unavailable when all labels come from biased diaries Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks; its effectiveness also varies sharply by domain — traffic-sign ASR only fell from ~99% to ~30% — and nothing suggests spatio-temporal motion data is the easy case. Worse for this route, backdoors injected at training time have been shown to survive subsequent clean-up fine-tuning 60% of the time in pipeline settings Backdoor Attacks on Decentralised Post-Training. A fine-pruning run using the calibration set as its best-available clean data is included as the representative of this family, and its expected failure mode — pruning biased-user knowledge along with the trigger — is itself a measurable instance of the bias/backdoor confusion this proposal formalizes.

Decisive Checks

The proposal stands or falls on two results. The separability experiment must show that residual label error after bias-model subtraction is statistically distinguishable between planted triggers and benign bias at realistic calibration sizes; if the distributions overlap, the honest finding is a negative result with security implications — self-report bias is free camouflage — which would itself redirect effort toward dual-instrument data collection as the only defense. And the full defensive pipeline must Pareto-dominate the ABCS and activation-norm baselines on the excision/retention plane while matching their ASR reduction, demonstrating that knowing how humans misreport is a usable security signal, not just an epidemiological curiosity.