When a neural network is trained on poisoned data, an attacker can hide a trigger-activated circuit inside it: the model behaves normally on clean inputs but switches to an attacker-chosen output whenever a secret signature appears. Whether such a circuit can be implanted is usually treated as a property of the data and the architecture. This proposal treats it as a property of the learning rule. Specifically, we ask whether backdoors can be seated in spiking neural networks — models of neurons that communicate with discrete pulses — when those networks learn through biologically plausible, multi-layer credit assignment rules that never propagate exact errors backward through the weight transpose, and whether the local signals such rules expose during training can be turned into a backdoor defense that no existing method provides.
The motivation is a visible hole in the literature. Every backdoor attack on spiking networks to date — spatial triggers, spatiotemporal event triggers, and clean-label temporal poisoning — trains the victim with backpropagation through time using surrogate gradients (Spatiotemporal Trigger on Event Data for Backdoor Attack on Spiking Neural Networks; Temporal Poisoning: Clean-Label Backdoors via Event Redistribution in SNNs; Poster: Backdoor Attacks on Spiking NNs and Neuromorphic Datasets). Meanwhile, a separate line of work shows that decoupling the feedback pathway changes a network's security behavior: PEPITA, a forward-only rule that projects the output error through a fixed random matrix to modulate a second forward pass, loses only 0.26% adversarial accuracy on MNIST where backpropagation loses 8.05%, and it holds 41.00% PGD accuracy under fast adversarial training on CIFAR-10 where backprop collapses to 0.05% (Efficient Biologically Plausible Adversarial Training). That paper studies test-time evasion and explicitly leaves training-time poisoning open. Whether the feedback architecture is similarly a bottleneck for backdoors is unknown, and the answer matters practically: edge neuromorphic devices that train on-chip cannot run backpropagation through time at all — they must use local rules such as spike-timing-dependent weight inference (Overcoming the Weight Transport Problem via Spike-Timing-Dependent Weight Inference) — so the security of exactly those rules is the security of real deployments.
Core Claim
The claim has two parts, one negative and one constructive. First, a backdoor is a multi-layer coordination problem: the trigger feature must be carved in early layers and wired to the target class in late layers, in register with each other. Rules that transmit a single, coherent error signal to every layer — backpropagation through the exact weight transpose, or BurstCCN's learned top-down burst pathway to apical dendrites (Single-phase deep learning in cortico-cortical networks) — can coordinate that carving, so we predict they are backdoor-permissive. Rules with fixed, random, layer-isolated feedback — Direct Feedback Alignment and PEPITA — update each layer against an independent random projection of the error, which the LAFS analysis shows impairs inter-layer coordination on complex features (Explaining the Learning Dynamics of Direct Feedback Alignment); we predict they resist backdoors that require precise cross-layer alignment (temporal coincidence triggers) while remaining vulnerable to crude spatial triggers that a single forward layer can learn from local activity alone — a split predicted by the dormant-neuron spare-capacity account of backdoors (Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks).
Second, and independently of whether local rules resist attacks, the local plasticity process itself is an untapped detection signal. A neuron participating in a backdoor circuit receives essentially the same error direction on every poisoned sample — always toward one fixed target class — whereas neurons serving the diverse clean task receive error directions that vary with input content. Existing defenses look at activations after training (dormant-neuron pruning) or at input statistics (temporal entropy), and both fail on temporal clean-label triggers precisely because those triggers reuse active pathways rather than creating dormant ones (Temporal Poisoning: Clean-Label Backdoors via Event Redistribution in SNNs). We propose monitoring the direction of weight updates during training, which is available on-chip even when no backward pass exists.
Research Program
The program runs in three stages. Stage one is a susceptibility grid: we train the same convolutional spiking architecture on N-MNIST, CIFAR10-DVS, and DVS128-Gesture under four credit assignment regimes — surrogate-gradient backpropagation through time (the attack papers' setting), DFA, PEPITA, and BurstCCN — under matched poisoning budgets taken from prior work (injection ratio 0.1 as in the spatiotemporal trigger attack, and the 0.01 budget at which the Concentrate attack saturates on N-MNIST with 59 poisoned samples (Temporal Poisoning: Clean-Label Backdoors via Event Redistribution in SNNs)). We measure attack success rate, benign accuracy, and a cross-layer update-coordination index (the average alignment between each layer's update and the update exact backpropagation would produce). The registered prediction is an ordering — backprop ≥ BurstCCN > DFA ≈ PEPITA in backdoor success for coincidence-based triggers, with the gap shrinking or vanishing for crude spatial triggers. We are honest about what a null result means: if ASR is near 100% under every rule, the "feedback architecture as bottleneck" claim dies, and the proposal's value then rests entirely on stage three — a finding we would report as such.
Stage two is the mechanism analysis: we test whether the coordination index mediates the rule-to-ASR relationship, and we localize where the trigger circuit fails to form under decoupled feedback (which layer's updates diverge from the backdoor-consistent direction on poisoned samples). This is what separates the proposal from a bare empirical sweep: we want a causal story, checked by an intervention in which we hand-craft DFA's random feedback matrices to have low-rank structure that partially restores cross-layer coherence, and test whether backdoor implantation recovers accordingly.
Stage three builds the plasticity-signature detector. For each neuron $i$ we collect normalized weight-update directions $\Delta w_i^{(t)} / \|\Delta w_i^{(t)}\|$ over a training window and compute the concentration of their average:
$$C_i \;=\; \Big\| \frac{1}{T} \sum_{t=1}^{T} \frac{\Delta w_i^{(t)}}{\|\Delta w_i^{(t)}\|} \Big\|,$$
which is near 0 when updates point in varied, content-dependent directions and near 1 when they consistently push toward one fixed direction. Neurons flagged for high $C_i$ and low dependence on clean-input content are frozen or damped — a mitigation that, unlike the strongest existing SNN defense's clamping of initial-layer "dendrite" weights, applies at every layer and therefore targets the distributed triggers that evade early-layer clamping (Unsupervised Backdoor Detection and Mitigation for Spiking Neural Networks). The statistic requires only quantities that exist during local training — pre/post activity and the projected global error — so it runs online on hardware that cannot compute gradients.
Baselines
The naive reference point is poisoned training with no defense at all. Under this condition the literature already tells us what to expect on these benchmarks: the spatiotemporal event trigger achieves near 100% attack success rate at a 0.1 injection ratio across N-MNIST, CIFAR10-DVS, and DVS128-Gesture (Spatiotemporal Trigger on Event Data for Backdoor Attack on Spiking Neural Networks), and the Concentrate clean-label attack reaches an attack success rate of 1.00 on N-MNIST and DVS128-Gesture. Any defense must be judged against that ceiling, and against the clean accuracy of the undefended poisoned model, since a defense that destroys benign accuracy is worthless.
The strongest existing competitor is the TMPBD/NDSBM framework of Unsupervised Backdoor Detection and Mitigation for Spiking Neural Networks, which detects backdoored target classes data-free by gradient-ascent inversion of membrane-potential margins and mitigates by clamping early-layer weights. It reports 100% detection accuracy on DVS128-Gesture and N-Caltech101 and reduces attack success rate from 100% to 2.81% with zero clean-accuracy degradation — the best numbers in this literature, and the bar our detector must clear or sidestep. To make the comparison fair rather than rigged, we will grant TMPBD full surrogate-gradient access to our locally trained models (building a differentiable evaluator around the trained weights, exactly as its authors do), run it on identical poisoned checkpoints and data splits, and give both methods the same small unlabeled calibration budget. Our claim to beat it is not detection rate on standard attacks but two things it cannot do: operate without any backward pass, and survive the adaptive and distributed attacks we run in the final stage — including Time-Bandits-style triggers whose per-frame stealth (0.13 MSE, 98.88% SSIM against clean frames) leaves no early-layer signal to clamp (Poster: Backdoor Attacks on Spiking NNs and Neuromorphic Datasets).
Two further baselines calibrate the field. Fine-pruning — pruning neurons dormant on clean data, then fine-tuning — reduced attack success from 100% to 0% on a face-recognition trojan (Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks), but it is known to fail on temporal triggers that reuse active pathways, so we include it to confirm that failure transfers to our setting rather than assume it. The model-free temporal-axis detector of Temporal Poisoning: Clean-Label Backdoors via Event Redistribution in SNNs — per-step event mass, Gini coefficients, and temporal entropy — represents the input-statistics family; we run it with its published thresholds.
Finally, one ablation isolates the proposal's key ingredient. We run the identical detection pipeline but feed it firing-rate statistics instead of update-direction statistics. If rate statistics detect as well as plasticity statistics, the "credit-assignment-native" ingredient is doing nothing and the contribution collapses to another activity-based detector; if update directions win specifically on temporal clean-label triggers — the regime where activation-based defenses fail by construction — the ingredient is vindicated. A second ablation computes the detector from exact backprop gradients on the same checkpoints, to separate whether locality per se matters or whether any training-time directional signal would do.
Why not just…?
Why not just run the existing attacks against PEPITA- and DFA-trained networks and report the attack success rates? Because that sweep is stage one of this proposal, not the proposal. A table of attack success rates across learning rules would be a single correlation; without the coordination index, the intervention on the feedback matrices, and the failure localization, we would not know why any difference exists, could not predict which trigger classes it extends to, and could not turn it into a defense. The mechanistic analysis is what makes the result portable to rules not yet invented.
Why not just apply TMPBD offline, with surrogate gradients, even if training was local? This is the most serious objection, and we concede the point partially: a defender with white-box access to the trained weights can always bolt on a differentiable evaluator, which is exactly how we run TMPBD in our own comparison. Two things nonetheless keep our defense worth building. First, it catches backdoors during training, before the circuit is consolidated and before deployment, using only signals that exist on-chip; TMPBD is post-hoc and requires storing and inverting a full model, which is precisely the capability edge deployments lack. Second, TMPBD's mitigation clamps only initial-layer weights, and the distributed and phase-coincidence triggers already demonstrated in this literature leave no early-layer anomaly to clamp — the failure mode is documented, not speculative (Poster: Backdoor Attacks on Spiking NNs and Neuromorphic Datasets; Temporal Poisoning: Clean-Label Backdoors via Event Redistribution in SNNs). The adaptive-attack experiment is where the two defenses are directly separated.
Why not just use input-level temporal statistics — the Gini/entropy detector? Because its authors already flag its weakness: an adaptive attacker can redistribute event timestamps to mimic natural temporal profiles while still crossing the coincidence threshold of leaky integrate-and-fire neurons (Temporal Poisoning: Clean-Label Backdoors via Event Redistribution in SNNs). The input side is attacker-controlled territory. Our detector sits on the plasticity side, where the attacker faces a genuine tension: for the backdoor to be learnable at all, the trigger pathway must receive consistently directed error on poisoned samples — the very signature we monitor. An attacker who randomizes that signature to evade us is randomizing the learning signal the backdoor needs. Whether that tension can be broken (for example, by spreading the attack across many target classes to dilute per-neuron concentration) is exactly what our adaptive-attack experiment tests, and we will report it honestly if the multi-target variant defeats us.
Why not just collapse the event stream into 2D frames and use standard vision defenses? Because it is blind by construction: the Concentrate attack preserves exact per-pixel, per-polarity event counts, so the rate frame of a poisoned sample is mathematically identical to its clean counterpart, and any frame-based defense sees nothing while the spiking model still fires the backdoor at an attack success rate of 1.00 (Temporal Poisoning: Clean-Label Backdoors via Event Redistribution in SNNs). Frame conversion also discards the timing information that makes event cameras useful in the first place (Spatiotemporal Trigger on Event Data for Backdoor Attack on Spiking Neural Networks).
Why study these rules at all — aren't local learning methods too weak to matter? They have crossed the relevance threshold. BurstCCN trains stably to 9 layers with 1.84% error on MNIST and 38.99% test error on CIFAR-10 against 36.30% for a standard network (Single-phase deep learning in cortico-cortical networks), and PEPITA's robustness results on CIFAR-10 are competitive where backprop catastrophically fails (Efficient Biologically Plausible Adversarial Training). Note that no reported result found on Lacuna for clean accuracy of DFA, PEPITA, or BurstCCN on N-MNIST, CIFAR10-DVS, or DVS128-Gesture — establishing those clean-accuracy reference points under matched architectures is itself a deliverable of stage one.
Proposed Experiments and Decisive Checks
All experiments use the benchmarks and budgets already established in the file's literature: N-MNIST, CIFAR10-DVS, and DVS128-Gesture, with N-Caltech101 added for the defense stage to enable direct comparison with the TMPBD numbers; the SpikformerLite transformer is included to replicate the Concentrate attack's original setting. Metrics are attack success rate, benign accuracy, detection AUROC, and post-mitigation attack success rate, with a fixed hyperparameter search protocol and matched compute across all methods. Three checks decide the proposal. First, the susceptibility grid must show the predicted rule-by-trigger-type split — if attack success rates are uniform across learning rules, the first core claim is falsified. Second, the mediation analysis must show that update coordination, not some incidental property of the rules, carries the effect, with the structured-feedback intervention as the causal test. Third, the plasticity-signature detector must, on locally trained models with no gradient access, reduce attack success rate to at or below the 2.81% that TMPBD/NDSBM achieve with full gradient access (Unsupervised Backdoor Detection and Mitigation for Spiking Neural Networks), while retaining benign accuracy, and must remain ahead of TMPBD under the adaptive multi-target and Time-Bandits-style distributed attacks. If the adaptive attacker defeats us on the concentration statistic, the fallback contribution — the empirical security map of credit assignment rules, with its mechanistic explanation — still stands on its own.
References
- Spatiotemporal Trigger on Event Data for Backdoor Attack on Spiking Neural Networks
- Temporal Poisoning: Clean-Label Backdoors via Event Redistribution in SNNs
- Poster: Backdoor Attacks on Spiking NNs and Neuromorphic Datasets
- Efficient Biologically Plausible Adversarial Training
- \nOvercoming the weight transport problem via spike-timing-dependent weight inference
- Single-phase deep learning in cortico-cortical networks
- Explaining the Learning Dynamics of Direct Feedback Alignment
- Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks
- Unsupervised Backdoor Detection and Mitigation for Spiking Neural Networks