A backdoored image classifier is one that was trained on poisoned data: it behaves normally until an input carries a small trigger pattern, and then it outputs the attacker's chosen label. A recent detection method, PSBD, found a clean fingerprint for such inputs. When you randomly switch off neurons at test time (dropout) and watch how much the prediction moves, backdoored inputs barely move at all. The network has built many redundant copies of the trigger detector, so losing any one copy does not matter. The trouble is that easy, confidently classified clean images also barely move, which is why this detector still wrongly flags 10–15% of clean samples.
We propose to separate the two kinds of stability, "stable because the evidence is redundant" and "stable because the evidence is strong and spread out." We do this with a test-time operation borrowed from neuroscience called asymmetric lateral inhibition: strongly active units suppress weaker units that respond to similar features, but the weaker units do not suppress the strong ones. Applied inside a frozen network, this collapses each group of redundant trigger detectors to a single survivor. Dropout can then knock that survivor out. A backdoored input that was rock-steady under ordinary dropout suddenly becomes fragile under dropout after inhibition. A clean input, whose evidence already lives in dissimilar features, changes little. The detector flags the jump in fragility, not low fragility itself.
Core Claim
The low-uncertainty signature reported by PSBD: Prediction Shift Uncertainty Unlocks Backdoor Detection mixes two causes. The first is redundant pathways, which are specific to backdoors. The second is strong, distributed evidence, which is common in clean data. Asymmetric, similarity-weighted inhibition removes only the first. So the difference in prediction shift with and without inhibition should detect poisoned samples at a much lower false positive rate than prediction shift alone, while keeping PSBD's near-perfect true positive rate.
Concretely, at a chosen layer with activations $a_i$ for channel $i$, the inhibited activation is
$$\tilde a_i = \max\Big(0,\; a_i - \lambda \sum_{j \neq i} s_{ij}\, a_j\, \mathbb{1}[a_j > a_i]\Big),$$
where $s_{ij}$ is the positive part of the cosine similarity between the filters of channels $i$ and $j$, and $\lambda$ is the inhibition strength. The indicator makes the inhibition one-directional: only stronger units suppress weaker ones. Because $s_{ij}$ comes from the weights, the operator needs no clean data.
The subtraction term is computed once from the full, undropped forward pass and then held fixed while dropout masks are sampled. This ordering is essential. If inhibition were recomputed after dropout, dropping the winning unit would release its suppressed copies and restore the redundancy.
The detection score is
$$\Delta(x) = \mathrm{PS}_{\text{inhib}}(x) - \mathrm{PS}(x),$$
where $\mathrm{PS}$ is PSBD's average prediction shift over $K$ dropout passes, and $\mathrm{PS}_{\text{inhib}}$ is the same quantity with the frozen inhibition applied.
Where the Idea Comes From
The biological work supplies the mechanism rather than the evidence.
- Data-driven inference of network connectivity for modeling the dynamics of neural codes in the insect antennal lobe inferred asymmetric, non-local inhibition from connectome data and showed that it stabilizes odor codes by suppressing off-target projections.
- Sparse Coding and Lateral Inhibition Arising from Balanced and Unbalanced Dendrodendritic Excitation and Inhibition showed in olfactory bulb models that unbalanced inhibition, where active cells dominate weak neighbors, produces sparse codes.
- Seemingly Redundant Modules Enhance Robust Odor Learning in Fruit Flies and Laterally Interconnected Self-Organizing Maps show inhibition decorrelating overlapping representations.
We take one operational idea from this literature: winner-dominated inhibition among similar units removes redundancy. We use it not to build a better classifier but as a diagnostic knob.
The deep-learning uses of lateral inhibition are architectural and trained end to end. Multi-Scale Unrectified Push-Pull with Channel Attention for Enhanced Corruption Robustness pairs each convolution with a wider opposite-sign kernel and was evaluated only on common corruptions. Towards Real-world X-ray Security Inspection sharpens object boundaries in cluttered baggage scans. Neither measures uncertainty or touches attacks. The only paper pairing inhibition with "adversarial" training, Romanian Multiword Expression Detection Using Multilingual Adversarial Training and Lateral Inhibition, uses the word for domain alignment, not for defense.
Our operator differs from all of these in three ways. It is post hoc, parameter-free apart from $\lambda$, and applied to a possibly compromised network.
Two papers constrain the design. Adversarial Attacks Against Uncertainty Quantification drove ensemble variance to $10^{-6}$ by optimizing through differentiable uncertainty measures. Disrupting Deep Uncertainty Estimation Without Harming Accuracy showed that confidence can be manipulated without changing any labels. Any uncertainty-based detector therefore has to face an attacker who knows the detector. We plan for that directly with an adaptive backdoor described below.
Research Program
We first verify the premise directly. On networks backdoored on CIFAR-10 and GTSRB, we will measure the filter-similarity structure among the channels most activated by triggered inputs and compare it with the channels most activated by clean inputs of the target class. The PSBD "neuron bias" account predicts that trigger-driven channels form tight clusters of high $s_{ij}$. If they do not, the mechanism is wrong and the project stops here. This is an informative negative result about what redundancy means in backdoored networks.
Next we build the detector. This involves choosing layers (we sweep single-layer and multi-layer placements, since placement is an open question already noted for push-pull models), choosing $\lambda$, and choosing the dropout rate. These are chosen on a held-out poisoned training set with one known attack and then frozen for all other attacks, so that there is no per-attack tuning.
Finally we attack our own defense. We will train an adaptive backdoor whose loss adds a penalty on filter similarity among its most trigger-responsive channels, forcing the network to encode the trigger in dissimilar features. We will also train a variant that directly minimizes $\Delta$ on poisoned samples. We will report how much attack success rate the attacker must give up to evade the detector.
Baselines
The naive reference point is plain prediction confidence of the backdoored model, thresholded without any perturbation. It shows how much any stability probe adds.
The strongest competitor on exactly this question is PSBD (PSBD: Prediction Shift Uncertainty Unlocks Backdoor Detection), which thresholds prediction shift under dropout. It reports a true positive rate of nearly 1.0 across most attacks with a false positive rate of 10–15% on CIFAR-10, and a true positive rate of 0.996 against WaNet on GTSRB. We rerun it on identical poisoned models, with the same number of forward passes $K$ as our $\mathrm{PS}$ term, and with the same dropout-rate search budget as our method receives for $\lambda$ and dropout rate combined. Beating it means a clearly lower false positive rate at matched true positive rate (at least 0.99) across a majority of attacks, with no loss on WaNet.
STRIP (STRIP: A Defence Against Trojan Attacks on Deep Neural Networks) is the strongest input-level competitor. It blends held-out clean images into the input and flags low output entropy. It reports 0% false acceptance and 0% false rejection in most cases, and under 1% false acceptance at a preset 1% false rejection, at 1.32x latency. We run it with the same number of forward passes as our method. We expect it to remain strong on patch triggers, so our claim against it is limited to warping and blending triggers such as WaNet and Adaptive-Blend, and to not needing held-out clean images.
Two ablations isolate the key ingredient. Symmetric inhibition removes the indicator term, so similar units suppress each other equally. Random-similarity inhibition replaces $s_{ij}$ with shuffled values. We also run the recomputed-inhibition variant, which recomputes the subtraction after each dropout mask. We predict that it falls back to PSBD-level performance, which would confirm that frozen inhibition, not extra noise, carries the effect.
For area under the ROC curve on these attacks, we found no reported result on Lacuna for either PSBD or STRIP, so we will compute it for all methods ourselves.
Why not just…?
Why not just raise the dropout rate until backdoor paths break? Heavier dropout does eventually disrupt redundant copies, but it disrupts clean predictions at the same time. The two shift distributions move together rather than apart, which is presumably why PSBD needs careful rate calibration. This is a real threat, so we test it head-on. We sweep dropout rate for PSBD across its full range and plot separation between poisoned and clean scores. If some rate matches our false positive rate, the inhibition idea adds nothing, and we will say so.
Why not just use STRIP, which already reports near-zero error? STRIP's numbers come mostly from visible, input-agnostic triggers. Its own authors flag partial, source-specific backdoors as open, and it needs a pool of clean held-out images and 1.32x inference cost. Our method needs neither clean images nor input edits. Its advantage should appear specifically on warping and blending triggers, which the per-attack comparison will show or refute.
Why not just prune or clip the most active neurons? Clipping suppresses strong activations regardless of whether they are redundant. It would also hurt confident clean inputs, reintroducing the false positives we aim to remove. Our similarity-weighted, winner-only inhibition targets redundancy specifically. The random-similarity ablation is effectively a magnitude-only suppression control and will show whether similarity structure matters.
Why not train push-pull inhibition layers into the network instead? Architectural inhibition as in Multi-Scale Unrectified Push-Pull with Channel Attention for Enhanced Corruption Robustness requires control of training, and a poisoned model may simply learn around it. A post hoc probe works on any delivered model. As a secondary arm, we will train a push-pull ResNet-20 on poisoned data to see whether built-in inhibition prevents redundancy from forming at all.
Proposed Experiments
We will use CIFAR-10 and GTSRB with ResNet-20 and a wider ResNet. Attacks are BadNets-style patches, global blending, WaNet warping, and Adaptive-Blend, each at several poisoning rates. The metrics are true positive rate, false positive rate at a true positive rate of 0.99, area under the ROC curve, clean accuracy of any retrained model, and forward-pass cost.
Three checks decide the question. First, $\Delta$ must beat PSBD's false positive rate at matched true positive rate on at least three of four attack families with hyperparameters frozen from one attack. Second, the frozen-inhibition variant must beat both the recomputed and the symmetric variants; otherwise the proposed mechanism is not what is working. Third, under the similarity-penalized adaptive backdoor, either the detector still works or the attack loses substantial success rate. If the attack keeps full success while evading the detector, redundancy is not a necessary property of backdoors, and we will report the method as defeated by that attack.
References
- PSBD: Prediction Shift Uncertainty Unlocks Backdoor Detection
- Data-driven inference of network connectivity for modeling the dynamics of neural codes in the insect antennal lobe
- Sparse Coding and Lateral Inhibition Arising from Balanced and Unbalanced Dendrodendritic Excitation and Inhibition
- Seemingly Redundant Modules Enhance Robust Odor Learning in Fruit Flies
- Laterally Interconnected Self-Organizing Maps in Hand-Written Digit Recognition
- Multi-Scale Unrectified Push-Pull with Channel Attention for Enhanced Corruption Robustness
- Towards Real-world X-ray Security Inspection: A High-Quality Benchmark And Lateral Inhibition Module For Prohibited Items Detection
- Romanian Multiword Expression Detection Using Multilingual Adversarial Training and Lateral Inhibition
- Adversarial Attacks Against Uncertainty Quantification
- Disrupting Deep Uncertainty Estimation Without Harming Accuracy
- STRIP: a defence against trojan attacks on deep neural networks