Asymmetric Suspicion: Active Probing Across the Trigger-Strength Gap

A backdoored neural network behaves normally on ordinary inputs and misbehaves only when a hidden trigger appears. The best current defenses find the poisoned training examples by looking for telltale clusters in the network's internal representations — and they work well, until they don't. A recent line of attacks deliberately breaks this: by training on weakened triggers (fainter patches, lower opacity) and deploying full-strength triggers at test time, the Adap-Patch attack dissolves the latent cluster that defenses look for, and the strongest existing defense, SPECTRE, removes 0% of the poison while the attack retains a 97.5% attack success rate on CIFAR-10 (Revisiting the Assumption of Latent Separability for Backdoor Defenses). The core observation of this proposal is that such attacks hide in a coverage gap: the network is only ever observed at zero trigger and at the full-strength trigger, and the malicious behavior lives in between, in the unobserved regime of intermediate trigger strengths. This is structurally the same problem astrophysicists face with spectral coverage gaps — competing models of blazar jets are indistinguishable in the radio and very-high-energy gamma bands we do observe, and diverge only in the unobserved MeV gap (Extreme blazars under the eyes of MAGIC; Overcoming Spectral Coverage Gaps in Multi Wavelength Astrophysics) — except here the unobserved spectrum is the space of trigger configurations (strength, position, pattern) rather than electromagnetic wavelength.

We propose to close this gap with a closed-loop defender: an active sequential probing policy that adaptively sweeps trigger configurations against a trained network, treating each probe's behavioral response as evidence, with an asymmetric learning rule — confirming evidence raises suspicion fast, while null evidence lowers it slowly. The asymmetry is the key ingredient, and it is borrowed directly from asymmetric reinforcement learning, where separate learning rates $\alpha^+$ and $\alpha^-$ for positive and negative prediction errors govern risk attitudes and belief stability (Asymmetric REINFORCE for off-Policy Reinforcement Learning: Balancing positive and negative rewards). For a defender, the right attitude is "trigger aversion": a single hit is near-conclusive, while nulls are weak evidence because most probes simply miss the trigger's exact configuration. Symmetric updating cannot express this, and — as we argue below — it systematically acquits backdoored models under sparse evidence.

Core Claim

The claim is threefold. First, latent-inseparable backdoors are only invisible from the static operating points that current defenses observe; an adaptive probe policy that allocates its budget across the trigger-strength spectrum exposes them, because the backdoored network's response manifold diverges from a clean network's somewhere in that spectrum, just as competing astrophysical models diverge only in unobserved bands. Second, the probe policy's belief updates must be valence-asymmetric to work: with $\alpha_{\mathrm{hit}} \gg \alpha_{\mathrm{null}}$, suspicion persists across long stretches of uninformative probes and spikes on a single hit, which directly counteracts the premature-termination failure mode known to plague adaptive sampling and the documented optimism bias of large models acting as agents. Third, the existing "spectral" defenses are not wasted — their SVD outlier scores become the prior for our suspicion beliefs rather than the final verdict, so we build on rather than replace them.

This is deliberately different from the closest existing work in each direction. Learning to Watch: Active Video Anomaly Understanding via Interleaved Policy Optimization proves that closed-loop operators (Backtrack, Expand, Sample) beat static sampling's observational aliasing — its Anom-$\pi$ reaches 72.29% average precision on XD-Violence with a 2-billion-parameter model, beating a 72-billion-parameter passive model at 70.16% — but its preference updates are symmetric and it never considers adversarial inputs. Revisiting the Assumption of Latent Separability for Backdoor Defenses demonstrates the attack but offers no defense and never formalizes update-rate asymmetry in the defender. Profiling the Irrational Agent: Cognitive Modeling of LLM Behaviors in Sequential Jailbreaks measures asymmetric learning rates in models ($\alpha^+ = 0.738$ versus $\alpha^- = 0.177$ across GPT-4o, Llama 3, and DeepSeek-R1) purely as an evaluation tool, not as a design principle. Anti-Backdoor Coreset Selection via Cumulative Entropy filters poison in one passive batch, with no closed loop and no valence-dependent updates. Nobody has combined adaptive probing of trigger space with asymmetric suspicion dynamics.

Research Program

We will build a defender with three components. The first is a probe operator set, modeled on the interleaved operators of Learning to Watch but repurposed for trigger space: Ramp increases trigger strength or opacity, Shift moves a candidate patch, Mask ablates image regions to localize trigger position, and Sample draws a fresh configuration from a proposal distribution over composite triggers. Each probe is applied to a trained (possibly poisoned) network, and the response — output flip, logit gap change, or activation shift relative to the unprobed baseline — is recorded. The scheduler chooses the next operator by expected information gain about which training samples are poisoned, using the elimination logic of fully adaptive pure-exploration bandits, where arms (here, candidate poison hypotheses) are discarded the moment they statistically separate (A fully adaptive algorithm for pure exploration in linear bandits).

The second component is the asymmetric belief update. Let $b_t \in [0,1]$ be the defender's suspicion that a given training sample (or class, or subnetwork) is poisoned, initialized from the SVD outlier scores of Spectral Signatures in Backdoor Attacks rather than from a constant. After each probe:

$$b_{t+1} = b_t + \alpha_{\mathrm{hit}}\,(1 - b_t)\,\mathbb{1}[\text{probe fires}] \;-\; \alpha_{\mathrm{null}}\, b_t\,\mathbb{1}[\text{probe is silent}]$$

with $\alpha_{\mathrm{hit}} \gg \alpha_{\mathrm{null}}$, both tuned on held-out attacks. The design rationale comes from the asymmetric-learning literature: fast updates on confirmatory evidence and slow decay otherwise stabilizes belief in the presence of sparse, delayed signals, the mirror image of the risk-aversion regime in opponent-pathway reinforcement learning, and the inverse of the optimism bias that Profiling the Irrational Agent shows makes models collapse into compliance within 7.66 turns of sequential manipulation (instruction attack success rate 0.76 on AdvBench and AgentHarm). A stopping rule terminates probing when suspicion saturates or when a fixed probe budget is exhausted; we will derive a conservative variant that never terminates on nulls alone, addressing the premature-stopping failure mode of greedy adaptive sampling directly.

The third component is analysis. We will formalize an aliasing argument for trigger space, parallel to the observational-aliasing floor proven for static video sampling in Learning to Watch and to the polynomial-to-logarithmic separation between interactive and non-interactive protocols in structured estimation (The Role of Interactivity in Structured Estimation): any fixed, open-loop probe schedule of budget $n$ cannot separate a backdoored network trained on weakened triggers from a clean network if the trigger-response divergence occurs only above a strength threshold the schedule under-samples, whereas an adaptive schedule concentrates probes near the inferred threshold. This also gives the honest scope of the claim: if the trigger family is known and one-dimensional, a fixed grid is competitive, and our advantage must come from high-dimensional or composite trigger spaces.

As a secondary transfer experiment, we will test the same scheduler on the physical version of the coverage-gap problem: deciding which electromagnetic band to observe next when discriminating between competing blazar or supernova models that diverge only in unobserved regimes — the MeV gap for jet models (Extreme blazars under the eyes of MAGIC; Insights from the first flaring activity of a high synchrotron peaked blazar with X-ray polarization and VHE gamma rays) or the early-time ultraviolet gap that obscures supernova shock breakout and engine identification (Twin peaks: SN 2021uvy and SN 2022hgk in the landscape of double-peaked stripped envelope supernovae, which found that only 12 of 501 ZTF stripped-envelope supernovae were double-peaked and that peak magnitudes correlate at p $\approx$ 0.025) — while injecting synthetic transients into the alert stream as a poisoning attack on the scheduler itself. We treat this as a stress test of the scheduler's robustness to bad-faith data in an open pipeline, not as evidence that the two meanings of "spectral" are the same object; no reported result found on Lacuna for adversarial robustness of telescope-follow-up scheduling, which is itself part of the gap.

Baselines

The naive reference point is doing nothing: train on the poisoned dataset and deploy. On CIFAR-10 with 250 poisoned images (5% of one class), this yields misclassification of 90% of backdoored test images (Spectral Signatures in Backdoor Attacks). The strongest existing competitor is SPECTRE, which whitens penultimate-layer activations with robust covariance estimates and removes samples above a threshold set at 1.5 times the assumed poison fraction $\epsilon$ (SPECTRE: defending against backdoor attacks using robust statistics); on CIFAR-10 m-way pixel attacks it removes 100% of poisoned samples and drives attack success rate to near 0% where undefended baselines exceed 80%. We will run SPECTRE with an oracle $\epsilon$ (a concession in its favor), the same data, the same retraining budget, and the same tuning budget as our method, and we expect it to win on separable attacks — the decisive comparison is the adaptive regime, where the reported result is 0% elimination against Adap-Patch with the attack retaining 97.5% attack success rate on CIFAR-10 (Revisiting the Assumption of Latent Separability for Backdoor Defenses). The second baseline is the original SVD outlier filtering of Spectral Signatures, which after filtering and retraining brings misclassification within 1% of clean performance at 92–93% clean accuracy on CIFAR-10; it is the strongest competitor on clean-accuracy retention and supplies our prior, so beating it means improving elimination in the adaptive regime at matched clean accuracy. The third baseline is batch coreset selection via cumulative entropy (Anti-Backdoor Coreset Selection via Cumulative Entropy), representing everything achievable without a closed loop. Two ablations isolate our ingredients: the same probe scheduler with symmetric rates ($\alpha_{\mathrm{hit}} = \alpha_{\mathrm{null}}$), and an open-loop fixed grid over trigger strengths with a threshold detector. Beating the field means, on Adap-Blend and Adap-Patch across CIFAR-10, GTSRB, and ImageNet subsets with ResNet-20, VGG-16, and MobileNet-V2: elimination rate substantially above the 0% that latent-separability defenses achieve, post-retrain attack success rate at or below roughly 10%, and clean accuracy within one to two points of the 92–93% clean reference on CIFAR-10.

Why not just…?

Why not just use SPECTRE or Spectral Signatures with better hyperparameters? Because their failure is structural, not a tuning issue: Adap-Patch's poisoning-based regularization and training-time trigger weakening collapse the Silhouette separation score from above 0.4 to near 0.1, so no threshold on the latent cluster structure separates poison from clean (Revisiting the Assumption of Latent Separability for Backdoor Defenses). Their signal lives in representation geometry; ours lives in the network's behavior under probes, which trigger weakening cannot suppress at full strength by construction, since the attack needs the full-strength trigger to fire at test time. We still rerun them with oracle $\epsilon$ to confirm this on our setup.

Why not just sweep trigger strengths on a fixed grid? This is the most serious alternative, and in a one-dimensional trigger family (opacity alone) it would plausibly work as well. We concede this and scope the claim: our headline comparisons use composite triggers (position, size, opacity, pattern combinations) where the configuration space is combinatorial and a fixed budget spread over a grid starves every cell. The interactive-versus-non-interactive separation in structured estimation shows adaptive protocols win by polynomial to logarithmic factors precisely in such high-dimensional settings (The Role of Interactivity in Structured Estimation); our grid ablation will measure where the crossover lies, and if it never comes even for composite triggers, the asymmetric-update claim is dead and we will report that.

Why not just fine-tune or prune the network instead of probing it? Fine-pruning exploits spare capacity to argue that dormant malicious subnetworks can be pruned away (Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks), but it degrades clean accuracy and, per the latent-separability result, adaptive attacks re-embed the backdoor in weights that clean validation also depends on (Revisiting the Assumption of Latent Separability for Backdoor Defenses). A symmetric probing loop with reward rescaling is a subtler version of this question, and the evidence against it is direct: Profiling the Irrational Agent tested symmetric magnitude scaling and found it barely moved behavior, while valence structure did all the work; and in off-policy learning, symmetric penalization destabilizes updates where valence-dependent dampening does not (Asymmetric REINFORCE). Our symmetric-rate ablation settles this empirically: with nulls vastly outnumbering hits, symmetric suspicion decays between hits and acquits rare composite triggers.

Proposed Experiments

The primary experiments run on the CIFAR-10, GTSRB, and ImageNet-subset benchmarks with BadNet, Blend, Adap-Blend, and Adap-Patch attacks, replicating the protocol and architectures of Revisiting the Assumption of Latent Separability for Backdoor Defenses, reporting elimination rate, post-retrain attack success rate, and clean accuracy, with all defenses given equal compute and tuning budgets. The decisive checks are four. First, the symmetric-versus-asymmetric ablation must show a large elimination-rate gap on Adap-Patch; if symmetric probing also works, the proposal's core claim is falsified. Second, the grid ablation must show the adaptive advantage growing with trigger-configuration dimensionality, mapping the honest boundary of the claim. Third, an adaptive adversary who knows our defense and uses conjunction triggers (backdoor fires only on a specific pair of probe-compatible configurations, maximizing null sequences) tests whether asymmetric null-damping keeps suspicion alive long enough; this is the make-or-break robustness test, and decontamination without knowing the trigger is the acknowledged open problem in agentic settings (Backdoor Decontamination Dynamics in LLM Agents). Fourth, the LLM-agent variant: run the probe scheduler with the belief ledger externalized versus held in-context by a model exhibiting the documented $\alpha^+ \gg \alpha^-$ bias, measuring whether external asymmetric bookkeeping corrects premature acquittal. The astronomy transfer simulation closes the loop: a scheduler choosing observations across a synthetic MeV-gap or early-UV-gap model discrimination task, under injected false-alert poisoning, evaluated on discrimination accuracy and wasted-observation budget relative to an unattacked run.