A backdoor attack plants a hidden behavior in a neural network: the classifier looks normal on everyday inputs, but whenever a specific trigger pattern appears — a small patch of pixels, a rare word — the model switches to an attacker-chosen answer. The security literature treats the planted trigger circuit as a single object that either survives downstream retraining or does not. Two unrelated bodies of work suggest this is the wrong picture. In perceptual learning, practiced sensory skills become bizarrely specific — training people to discriminate a pattern at one retinal position or orientation yields gains that vanish when the stimulus is moved or rotated — because practice rewires two dissociable subsystems: a slow, unsupervised process that reshapes early feature representations and locks them to the trained stimulus coordinates, and a fast, supervised process that adjusts the decision readout and generalizes broadly (Liu et al.). In the study of human memorability, what is easy to recognize minutes after exposure is often forgotten days later unless it is integrated into semantic structure, and forgetting follows a measurable per-item decay curve rather than a sudden cliff (Naim et al.; Newman et al.).
My claim is that a backdoor is a learned representation like any other, and so obeys both laws at once. A trigger can be captured by the fast readout pathway or by the slow feature-representation pathway of the host network, and which pathway captures it determines two things that the security literature currently conflates: how the backdoor generalizes in space, and how it dies over continued clean training. This gives defenders a cheap, measurable signature — spatial tuning — that predicts the hard-to-measure property that matters: persistence.
Core Claim
Concretely, I propose the following two-part hypothesis, testable on standard vision backdoor benchmarks. First, spatial specificity: a trigger encoded in early feature hardware (analogous to the slow feature-based learning in the perceptual-learning model) should lose attack success when shifted, rescaled, or rotated away from its trained coordinates, exactly as learned shape selectivity in primate visual cortex fails to transfer across retinal position (Does Learned Shape Selectivity in Inferior Temporal Cortex Automatically Generalize Across Retinal Position?) and as motor adaptation decays systematically with angular distance from the trained direction (Fernandes et al.). A trigger captured only in the readout should be roughly translation-invariant, because pooled decision layers generalize across position. Second, forgetting curves: under continued clean fine-tuning, attack success rate should decline as a continuous function of adaptation exposure, not collapse at a discrete step, and it should be well fit by a two-parameter linear law borrowed from the memorability literature,
$$\mathrm{ASR}(t) = a_0 - \alpha\, t,$$
where $t$ is cumulative clean adaptation (fine-tuning steps or epochs), $a_0$ is the initial attack success rate, and the decay rate $\alpha$ is small for feature-captured and semantically grounded triggers and large for readout-captured, low-level pixel triggers. The analogy to Newman et al., who model human video memory as a base score minus an item-specific decay rate and find heterogeneous decay across items, is direct and deliberate. The further analogy to human memory is also pointed: semantically integrated material survives consolidation while salient-but-shallow material decays (Naim et al.; Feitosa Tomé et al.), which matches the observation that pixel-patch triggers wash out under adaptation while triggers embedded in semantic feature spaces survive (Cheng et al.; Ma et al.) — but nobody has ever measured this as a decay curve or connected it to where in the network the trigger lives.
The neurobiological precedent for the dissociation is strong. Practice modifies only the projection pathway a task actually uses, leaving adjacent pathways in the same tissue untouched (Yamashita et al.), and motion learning rewires a specialized cortical sub-region rather than general-purpose cortex (Shibata et al.). The backdoor literature has independently found the same anatomy without naming it: trojan logic hides in dormant, trigger-dedicated neurons that clean inputs never activate (Dolan-Gavitt et al.; Fine-Pruning), and a recent agent-decontamination study found that behavioral suppression of a trigger can leave its internal recognition circuit intact in intermediate layers — a dissociation between readout and representation that is structurally the feature/task split of perceptual learning (Backdoor Decontamination Dynamics in LLM Agents). I am proposing to make that split a predictive, quantitative theory of backdoor vulnerability.
Research Program
The program has three components. The first is a trigger zoo on CIFAR-10, GTSRB, and MNIST — the benchmarks used by the foundational backdoor papers (BadNets; STRIP; Cinà et al.) — that crosses two controlled axes. Axis one is representational locus: triggers are injected with either early-layer plasticity enabled (so slow, locally updated feature maps can capture the trigger, mimicking feature-based learning) or early layers frozen (so only the readout can learn the trigger, mimicking the ablated-feature condition in Liu et al., which produced fully generalizing behavior). Axis two is semantic groundedness: low-level patch triggers in the BadNets style versus naturalistic semantic triggers in the style of ambient-attribute attacks (Ma et al.), scored for memorability-like semantic integration using caption-embedding features in the spirit of Newman et al., whose model showed that jointly predicting text captions is what captures decay-resistant structure.
The second component is the measurement protocol: for each trigger in the zoo, sweep the trigger across position, scale, and orientation at test time to trace its spatial tuning curve, then subject the poisoned model to clean fine-tuning and log attack success rate densely to fit $\alpha$. The deliverable is a predictive relationship — spatial tuning width and semantic groundedness jointly forecast decay rate — and a demonstration of the two dissociable cells that prior work conflates: a backdoor that is spatially specific yet highly persistent (feature-captured, semantic) and one that is position-invariant yet fragile (readout-captured, patch). The first cell is the security-relevant one and, to my knowledge, has never been deliberately constructed.
The third component turns the theory into a defense and an attack forecast. On the defense side, spatial tuning sweeps provide a cheap localization signal for dormant trigger detectors: a candidate region that responds strongly at one stimulus position and not at shifted positions flags feature-hardware involvement, guiding pruning before fine-tuning. On the attack side, the perceptual-learning "double training" effect — where training an orthogonal task at a second location unlocks transfer of the primary skill (Liu et al.) — predicts a new escalation: brief auxiliary poisoned training at a second position could force a spatially contained trigger to generalize across the input space, meaning defenders cannot rely on spatial containment of a discovered trigger.
Baselines
The "do nothing" reference point is clean fine-tuning of the poisoned model with no defense. This baseline matters more than it sounds, because the strongest existing evidence says it often works: Zhao et al. report that standard poisoning attacks on Qwen3-0.6B saw attack success rate drop to near 0% after supervised fine-tuning on SST-2. However, that is an LLM result; no reported result found on Lacuna for natural backdoor decay under clean fine-tuning on CIFAR-10, GTSRB, or MNIST, so establishing this curve on vision models is itself a contribution. The comparison is fair by construction: every defense is evaluated on the same poisoned checkpoints, with the same clean fine-tuning budget and the same tuning of learning rate and epochs. A defense beats "do nothing" only in the regime where the measured decay curve leaves attack success rate meaningfully above zero — which, per the persistence literature (Kurita et al.; Cheng et al.), is exactly the regime of feature-captured and semantic triggers.
The strongest existing defense baseline is Fine-Pruning, which prunes neurons that stay dormant on clean data and then fine-tunes the survivors (Liu, Dolan-Gavitt, and Garg). It is the strongest competitor because it exploits the same dormant-circuit anatomy my theory builds on, and its reported numbers define the bar: on the U.S. traffic signs dataset it reduced attack success rate from roughly 99% to about 30% while maintaining clean accuracy, and on the face recognition task it drove attack success rate from 100% to 0%. Crucially, its own authors show it degrades when an attacker forces trigger representations to be shared with benign classes — the precise failure mode my spatial-probing component is designed to address, since spatial tuning is measurable even for shared representations. I will run Fine-Pruning on identical poisoned checkpoints with an equal compute budget and an equal search over pruning fraction, and my method beats it if it pushes attack success rate below 10% at matched clean accuracy (within one point) on the standard recipe, and near 0% under the pruning-aware attacker where Fine-Pruning's residual attack success is highest.
A second comparison is STRIP, a runtime filter that overlays incoming inputs with clean images and flags inputs whose predictions stay locked to one label (Gao et al.), reporting a false acceptance rate under 1% at a preset 1% false rejection rate on GTSRB and CIFAR-10, at the cost of roughly 1.32x inference latency. STRIP assumes the trigger dominates the superimposed input regardless of background; a spatially specific trigger that only fires at its trained coordinates is a plausible blind spot, and testing this is one of the experiments. Finally, the key ablation isolates my proposal's central ingredient: the same pruning-plus-fine-tuning pipeline as Fine-Pruning, but with trigger-candidate identification by spatial tuning sweep rather than by dormancy alone. Any gain over Fine-Pruning under the pruning-aware attacker must come from this ingredient, and nothing else.
Why not just…?
Why not just fine-tune the poisoned model on clean data and be done? For standard patch backdoors this may indeed suffice — the near-0% post-fine-tuning attack success reported by Zhao et al. for standard attacks on Qwen3-0.6B is direct evidence for the LLM case, and my first experiment measures the vision analogue honestly rather than assuming it. But two cited results say fine-tuning alone is not enough in general: dormant trigger neurons receive no gradient from clean data because the trigger never appears in it (Fine-Pruning; Kurita et al.), and even when behavioral attack success is driven to 0%, internal trigger-recognition traces persist in intermediate representations and can in principle be re-activated (Backdoor Decontamination Dynamics in LLM Agents). My decay curves make the boundary explicit: they tell the defender when doing nothing is safe and when it is not, which no existing result provides for vision models.
Why not just explain persistence with loss-landscape geometry, as BAD-BOOM does? Zhao et al. already show that standard backdoors sit in a narrow, sharp basin that clean fine-tuning knocks them out of, and that Fisher-information-guided smoothing preserves attack success at 97.1% after SST-2. That analysis is real and largely correct — but it is attacker-side, LLM-specific, costs approximately 1.35x training time and 6GB of extra GPU memory, and says nothing about spatial generalization of the trigger or about continuous decay dynamics. My approach is observational and cheap: probing an already-trained suspect model with shifted triggers and fitting a decay curve requires no retraining and no curvature estimation. If measured decay rates correlate with basin sharpness where both can be computed, that is convergent evidence for both views; if they diverge, the spatial-tuning signature is the more deployable one. Either outcome is informative, and I will run that comparison.
Why not just accept that "semantic triggers persist" — isn't that already known? Qualitatively, yes: deep-feature and naturalistic triggers surviving adaptation is documented (Cheng et al.; Ma et al.), and the parallel with semantic integration in human memory (Wang et al.) has been noted. What is not known is the dissociation my 2-by-2 design delivers: prior work treats "deeply embedded" as one undifferentiated robustness property, whereas the perceptual-learning theory says spatial specificity and temporal persistence are separate axes, because they are owned by different subsystems. Constructing a trigger that is spatially locked yet nearly undecayable, and one that is position-invariant yet fast-decaying, refutes the one-axis view outright. If the 2-by-2 collapses — if specificity and persistence turn out perfectly correlated — the theory is wrong in an interesting, publishable way, and the field learns that a single depth axis suffices.
Proposed Experiments and Decisive Checks
Experiment one measures spatial tuning. Backdoored CNNs on CIFAR-10 and GTSRB are trained with fixed-location patch triggers; attack success rate is swept against trigger displacement, scale change, and rotation. The prediction is a narrow tuning curve for early-plastic models, mirroring the angular decay of motor adaptation (Fernandes et al.), and near-flat tuning when early layers are frozen during poisoning — the direct analogue of the frozen-feature condition in Liu et al., where disabling feature-based learning yielded 100% transfer across positions. Regularization strength modulates the effect, following the observation of Cinà et al. that strong regularization (λ=10 versus λ=1 in their experiments) slows backdoor acquisition.
Experiment two measures forgetting curves. Each zoo model is fine-tuned on clean data with attack success rate logged every fraction of an epoch; $\alpha$ is fitted per trigger, and alternatives (exponential, step-function collapse) are compared by fit quality. The decisive check is cross-trigger prediction: a regression from spatial tuning width and semantic-groundedness score to measured $\alpha$, evaluated on held-out triggers. As a calibration for what counts as a good correlation in decay modeling, Newman et al. reached a Spearman correlation of 0.663 against human memorability judgments on Memento10k, approaching the 0.730 human-consistency ceiling; a locus-plus-memorability predictor of $\alpha$ should aim at comparable rank correlation against measured decay. A second decisive check reproduces the consolidation asymmetry of Wang et al. in the model: initial attack strength $a_0$ should not predict survival, but locus and semantics should.
Experiment three evaluates the defense. On the standard BadNets recipe (whose reference numbers are attack success of 99% at 99.5% clean accuracy on MNIST, and a physical stop-sign trigger classified as a speed limit with 95% confidence), and on the pruning-aware attacker that defeats Fine-Pruning, the specificity-guided pipeline is compared against Fine-Pruning and STRIP under matched compute, with false acceptance and rejection rates reported for the runtime comparison. Success is attack success rate below 10% at matched clean accuracy where Fine-Pruning stalls near 30% on traffic signs. A follow-up re-activation probe, in the spirit of the J-lens dissociation findings (Backdoor Decontamination Dynamics in LLM Agents), checks whether internal trigger recognition is actually gone or merely suppressed at the readout.
Experiment four tests the double-training escalation: after planting a spatially contained trigger, a short poisoned auxiliary phase at a second position is applied, and trigger generalization is re-measured. If transfer unlocks the way the double-training effect does in Liu et al., then spatial containment is not a safe assumption for defenders, closing the loop between the perceptual-learning mechanism and the security consequence.
References
- To Learn or Not to Learn, That is the Question — A Feature-Task Dual Learning Model of Perceptual Learning
- Emergence of hierarchical organization in memory for random material
- Multimodal Memorability: Modeling Effects of Semantics and Decay on Video Memorability
- Does Learned Shape Selectivity in Inferior Temporal Cortex Automatically Generalize Across Retinal Position?
- Measuring Generalization of Visuomotor Perturbations in Wrist Movements Using Mobile Phones
- Dynamic and selective engrams emerge with memory consolidation
- Deep Feature Space Trojan Attack of Neural Networks by Controlled Detoxification
- Watch Out! Simple Horizontal Class Backdoor Can Trivially Evade Defense
- Target-specific membrane potential dynamics of neocortical projection neurons during goal-directed behavior
- Decoding Reveals Plasticity in V3A as a Result of Motion Perceptual Learning
- Trojaning Language Models for Fun and Profit
- Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks
- Backdoor Decontamination Dynamics in LLM Agents
- BadNets: Identifying Vulnerabilities in the Machine Learning Model Supply Chain
- STRIP: a defence against trojan attacks on deep neural networks
- Backdoor learning curves: explaining backdoor poisoning beyond influence functions
- Broadening the Backdoor Basin: Understanding LLM Backdoors Collapse and Making Backdoors Persistent
- Weight Poisoning Attacks on Pretrained Models
- RUC at MediaEval 2018: Visual and Textual Features Exploration for Predicting Media Memorability