Fast–Slow Disagreement as an Online Backdoor Alarm for Streaming Adaptation

Models deployed in the real world no longer stand still: they keep learning after deployment, updating themselves on a stream of unlabeled data as conditions change (test-time adaptation). Two failure modes of this setup have been studied in isolation. The stream itself shifts — lighting changes, new camera locations, new corruptions — and multi-scale online ensembles address this by running a committee of learners with different memory lengths in parallel: short-memory learners react instantly to abrupt change, long-memory learners stay stable when nothing is happening, and a meta-learner shifts weight between them (Test-time Adaptation in Non-stationary Environments via Adaptive Representation Alignment). Separately, the stream may be hostile: Test-Time Poisoning Attacks Against Test-Time Adaptation Models showed that an adversary who injects a handful of crafted samples into the adaptation stream can corrupt the model's updates, dropping accuracy from 76.20% to 41.83% on CIFAR-10-C. Nobody has asked what happens when the adversary's goal is subtler: not to degrade the model, but to plant a backdoor — a hidden rule such that any input carrying an attacker-chosen trigger is misclassified to a chosen class while clean accuracy looks fine.

The core observation is that a committee of learners with different memory lengths responds to a backdoor differently than it responds to genuine drift, and this difference is measurable online. Backdoors are shortcuts: trigger–target associations are simple and overfit much faster than genuine semantic features, as trojan studies consistently observe (Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks, Protecting Model Adaptation from Trojans in the Unlabeled Data). So a fast, short-window learner will latch onto a trigger long before a slow, long-window learner does, and the two halves of the committee will disagree — specifically on the triggered samples, and specifically with the fast learner's flagged predictions collapsing onto one class. Genuine drift looks nothing like this: it moves every learner eventually, affects the whole window rather than a sparse subset, and never funnels the fast learner onto a single target class. We propose to turn that disagreement pattern into a per-sample quarantine signal that runs inside the adaptation loop, needs no clean reference data, and needs no manually tuned threshold.

Core Claim

A multi-scale online ensemble can defend itself against streaming backdoor injection without sacrificing its ability to adapt to drift. The disagreement between fast-scale and slow-scale committee members — together with the sparsity of the disagreeing samples and the collapse of fast-scale predictions onto one class across that subset — separates "the world drifted" from "someone is planting a trigger," on the fly, per sample. Excluding flagged samples from all adaptation updates blocks the backdoor (the slow learners never see the trigger and the fast learner is periodically reset to the slow consensus state), while unflagged samples continue to adapt freely, so clean accuracy under drift is preserved. This yields, to our knowledge, the first backdoor defense that is simultaneously online, clean-data-free, and drift-compatible.

Research Program

The proposal has three parts: measure the threat, build the defense, and extend the theory.

First, the attack study. We define a streaming backdoor injection attack that extends the poisoning objective of Test-Time Poisoning Attacks Against Test-Time Adaptation Models from availability damage (destroy clean accuracy) to targeted sleeper behavior: the injected samples are crafted to fit whatever self-supervised or entropy-based objective the adaptation method minimizes, while carrying a trigger associated with an attacker-chosen class. We run this against the multi-scale ensemble of Test-time Adaptation in Non-stationary Environments via Adaptive Representation Alignment — exponential memory windows from 8 to 128 batches with staggered restarts and meta-weighted aggregation — under drifting streams. This is an open empirical question with two outcomes, both publishable: the multi-scale structure could mitigate the attack (long-memory learners act as an anchor, and the meta-learner can downweight a fast learner that suddenly fits a strange pattern) or amplify it (the fast learners' high plasticity is precisely the short-cut-fitting property a backdoor needs). The tension is genuine — drift adaptation demands plasticity, and plasticity is what backdoors exploit — and no paper in the file has measured it.

Second, the defense, which we call cross-scale quarantine. For each incoming sample $x_t$, let $p^{\text{fast}}_t$ and $p^{\text{slow}}_t$ be the aggregated predictions of the fast half and slow half of the committee. The per-sample deviance score is

$$ s_t = D_{\mathrm{KL}}\!\big(p^{\text{fast}}_t \,\big\|\, p^{\text{slow}}_t\big), $$

and $x_t$ is quarantined when $s_t$ exceeds the $(1-\alpha)$ empirical quantile of the committee's own recent history of scores, $\{s_{t-k}, \dots, s_{t-1}\}$. The threshold is therefore self-calibrating on a rolling basis rather than manually tuned — the same move that let Backdoor Secrets Unveiled: Identifying Backdoor Data with Optimized Scaled Prediction Consistency remove manual thresholds offline, transplanted to an online setting where no fixed threshold could survive continuous drift anyway. Quarantine alone cannot distinguish drift from attack, so we add two conditioning tests on the recently flagged set: sparsity (drift raises scores densely across a window; triggers raise them on a sparse subset) and target-class collapse (under drift the fast learner's flagged predictions spread across classes; under backdoor injection they concentrate on one class, because there is one attacker and one target). Samples passing both tests are dropped from all learner updates; samples failing them are treated as drift and adapted to normally. If the quarantine set persists across restarts, the fast learner is reset to the slow consensus state, the temporal analogue of how Orion: Online Backdoor Sample Detection via Evolution Deviance recovers clean labels from shallow side-networks that ignore triggers. We also borrow Orion's cheap side-exit idea for label recovery on quarantined samples that must still receive predictions.

Third, the theory. Test-time Adaptation in Non-stationary Environments via Adaptive Representation Alignment proves dynamic regret bounds — the ensemble's cumulative error tracks the best time-varying parameter sequence — but only under convexity, and only for benign streams. We extend the analysis to streams where a fraction $\rho$ of rounds are adversarial, showing that quarantine forfeits at most $O(\rho T)$ rounds of adaptation while preserving the tracking rate on the remaining rounds, in the convex / affine-normalization setting that Ada-ReAlign itself analyzes. We state this bound honestly: it covers the convex surrogate, and the gap to fully non-convex deep networks — which Ada-ReAlign's own authors flag — remains open, as it does for the base method.

Proposed Experiments

All experiments use streams that combine both problems, which no existing benchmark covers. We build drifting, backdoored streams by splicing attack samples from the BackdoorBench suite (BackdoorBench: A Comprehensive Benchmark of Backdoor Learning) — BadNets, blended, and WaNet triggers, at injection rates from 0.1% to 2% — into the sequential corruption streams of CIFAR-10-C and ImageNet-C used by Ada-ReAlign, plus GTSRB for a safety-critical setting and iWildCam for real camera-trap drift. Three regimes are run for every dataset: drift-only (measures false positives and clean accuracy), backdoor-only, and combined. Metrics are clean accuracy under drift, Attack Success Rate, Robust Accuracy (the rate at which the true label is still recovered on triggered inputs — BackdoorBench shows defenses collapse on this as class counts grow), detection F1 and AUROC, and per-sample latency overhead, benchmarked against Orion's 1.7ms figure. For GTSRB we note that no reported result found on Lacuna for GTSRB individual baseline numbers exists in the file, so we will establish those numbers ourselves under the BackdoorBench protocol. The decisive early check is a signal-validity experiment: before building any defense, measure the AUROC of the raw fast–slow disagreement score at separating injected-trigger samples from clean samples under simultaneous drift. If there is no separation, the core claim fails and we report that.

Baselines

The do-nothing reference is the frozen source model with no adaptation at all, which on ImageNet-C under sequential shift suffers 32.7% error as reported by Test-time Adaptation in Non-stationary Environments via Adaptive Representation Alignment; it is trivially backdoor-immune (it never updates) but drift-crippled, and it defines the accuracy floor we must not approach. The strongest competitor on the drift axis is undefended Ada-ReAlign itself, which reports 21.1% error on ImageNet-C sequential shift and 23.6% on iWildCam, beating every other TTA method in its comparison; it is the strongest because nothing else adapts to non-stationary streams this well, but it is undefended, and the closest evidence about what poison does to a single-scale adapting model — Test-Time Poisoning Attacks Against Test-Time Adaptation Models dropping a TTT model from 76.20% to 41.83% with 10 poisoned samples in 1,000 — suggests it is fragile; our attack study will measure this directly, since no reported result found on Lacuna for Ada-ReAlign under backdoor injection exists. On the backdoor axis the strongest competitor is Test-Time Adaptation for Backdoor Defense, whose detection-during-pruning stage reduced average ASR from 88.18% to 2.32% on CIFAR-10 with ResNet18 at a roughly 2.6% clean accuracy cost; we run it window-by-window on our streams, granting it the same batch sizes its method needs, acknowledging up front that its Shapley-based repair is too slow for per-sample use. Orion: Online Backdoor Sample Detection via Evolution Deviance is the strongest online backdoor detector, reporting 90% F1 on feature-hidden attacks and 80% label recovery at 1.7ms per sample, but it requires a clean 1% reference set to train its side networks and its weights are frozen, so under drift it must be paired with a separate TTA method and its reference assumptions break. All baselines receive the identical streams, the same number of committee learners and the same compute budget where applicable, and tuning is performed only on drift-only benign streams — no baseline or proposed method gets to see attack samples during tuning. Success is defined in advance: clean accuracy under combined drift-and-attack within one point of undefended Ada-ReAlign's drift-only numbers, ASR at or below 5% (in the range TTBD achieves), detection F1 at Orion's level without Orion's clean reference data, and millisecond-scale overhead. Two ablations isolate the ingredients: the same ensemble with quarantine removed (measuring whether meta-weighting alone is protective — our claim is it is not, since every member still absorbs the poison), and a single-scale detector that compares one learner against its own history (measuring whether the multi-scale structure, rather than any online anomaly detector, is what carries the signal).

Why not just…?

Why not just sanitize inputs or prune the model before deployment? Static preprocessing (JPEG compression, bit-depth reduction) was tested directly against streaming poisoning by Test-Time Poisoning Attacks Against Test-Time Adaptation Models and failed because the model's state shifts dynamically while the defense stays fixed. Static pruning fares no better: BackdoorBench: A Comprehensive Benchmark of Backdoor Learning documents that Fine-Pruning-style defenses are fragile across architectures and lose robust accuracy as class counts grow, and SSDA: Secure Source-Free Domain Adaptation shows naive channel pruning drops accuracy from 74% to 56% on Office-Home. We will still run JPEG and Fine-Pruning as baselines, because the relevant question here — whether they fail specifically when the model keeps adapting after they run — has not been measured, and our streams will measure it.

Why not just run a drift detector and pause adaptation when it fires? Classic detect-then-adapt detectors are the textbook alternative, and the stream-learning literature documents precisely why they fall short here: their fixed statistical thresholds fire on benign noise and miss gradual shifts (Learning from streaming data with concept drift and imbalance: an overview, Active Fuzzy Weighting Ensemble for Dealing with Concept Drift), and stealthy triggers of the WaNet/blended kind are explicitly designed not to disturb the global distribution statistics such detectors monitor. We must be honest that our quantile threshold is itself a threshold; the difference is that it is self-calibrated on a committee-internal signal, evaluated per sample, and confirmed by the sparsity and target-collapse tests rather than a single global hypothesis test. A head-to-head against a drift-detector-gated TTA pipeline on the same streams is included precisely so this claim is checked, not asserted.

Why not just buffer the stream and run an offline detector like MSPC? Backdoor Secrets Unveiled: Identifying Backdoor Data with Optimized Scaled Prediction Consistency achieves 0.8941 AUROC on CIFAR-10 with no clean data and no manual threshold — on paper, it solves the data-dependence problems we target. But it is a static, batch procedure, and buffering is fatal twice over: Orion: Online Backdoor Sample Detection via Evolution Deviance argues that real-time deployments cannot wait for batch inspection, and buffering under non-stationary drift means the model sits stale while the world moves (Test-time Adaptation in Non-stationary Environments via Adaptive Representation Alignment). That said, if a deployment genuinely tolerates hours of latency, periodic offline cleansing is a real alternative, so we run MSPC on buffered windows as a baseline and report exactly how much accuracy the staleness costs — if it matches our online numbers, the online machinery is only justified for low-latency settings, and we will say so.

Why not just average the committee with fixed weights? Fixed-weight averaging removes the meta-learner, and it fails for two documented reasons: it cannot resolve the stability–plasticity trade-off, since slow members drag the average during abrupt shifts while fast members inject variance during stable periods (Test-time Adaptation in Non-stationary Environments via Adaptive Representation Alignment), and under attack every member exposed to the same unfiltered stream absorbs the poison together (Test-Time Poisoning Attacks Against Test-Time Adaptation Models). This is the uniform-weight ablation above, run as a genuine competitor rather than dismissed.

Decisive Checks

The proposal stands or falls on three falsifiable results. The signal-validity experiment must show that fast–slow disagreement separates triggered from clean samples under simultaneous drift with usable AUROC; if drift and backdoor produce indistinguishable score distributions, the whole mechanism is void and we would report the negative result, since it would mean multi-scale structure cannot double as a security sensor. The attack study must show that streaming backdoor injection actually harms undefended multi-scale ensembles; if they are already robust, the defense is moot and the paper becomes that (still valuable) finding. And the combined-regime runs must show the defended ensemble within one point of undefended Ada-ReAlign's clean accuracy while holding ASR near TTBD's level — protecting the model without paying for it in adaptation, which is the trade every prior defense in this literature leaves on the table.