A backdoored network behaves normally on ordinary inputs but switches to an attacker's chosen label whenever a secret pattern, the trigger, appears. The standard explanation for how this works is spare capacity: the malicious behavior sits in weights and directions that clean data barely uses (Fine-Pruning). This proposal turns that explanation into a repair method. We fit a Bayesian posterior to the suspect network using only a small trusted clean set. Such a posterior is a probability distribution over weights that says how far each weight could move before the clean predictions change. By construction, it is narrow along directions the clean task depends on and wide along directions the clean task ignores, which is where the backdoor lives. We then average predictions over weights drawn from this distribution, or distill that average into one network. The expected effect is that the clean function is kept while the trigger's specific weight configuration is scrambled. There is no adversarial inner loop and no trigger search.
The opportunity matters because the strongest current repair methods, Adversarial Neuron Pruning (ANP) and Shared Adversarial Unlearning (SAU), both solve min-max problems. That brings the extra cost of adversarial training, and it brings its failure modes: SAU misses triggers larger than its search radius, and ANP damages compact models such as MobileNet-V2. A curvature estimate from one pass over the clean set is cheap. It also has an interpretation grounded in theory: posterior averaging cancels sensitivity along directions the data does not constrain.
Core Claim
Let $\hat{w}$ be the suspect weights and $D_c$ a trusted clean set holding 0.1–5% of the training data. We form a Laplace approximation, a Gaussian fitted around the current weights using the local curvature of the loss (A Scalable Laplace Approximation for Neural Networks). Its likelihood term comes from $D_c$ only:
$$q_\lambda(w) = \mathcal{N}\big(\hat{w},\ (F_c + \lambda I)^{-1}\big),$$
Here $F_c$ is the Kronecker-factored Fisher information, a measure of loss curvature, computed on $D_c$. $\lambda$ is a prior precision that sets how far samples may wander. The purified predictor is $\bar{p}(y\mid x) = \mathbb{E}_{w\sim q_\lambda}[p(y\mid x,w)]$.
The claim is that for common trigger types, the weight directions the trigger depends on have small $F_c$ eigenvalues. Averaging under $q_\lambda$ therefore collapses the attack success rate (ASR) while clean accuracy moves by an amount that $\lambda$ controls directly. We choose $\lambda$ with no knowledge of the attack: it is the smallest value, and hence the widest posterior, whose clean validation accuracy stays within 1% of $\hat{w}$.
A second claim comes for free. The spread of predictions across samples should flag triggered test inputs, because they rely on directions this posterior is uncertain about.
Where This Sits
Two results motivate the idea. Carbone et al. and Bortolussi et al. prove that in the infinite-width limit, posterior averaging cancels input gradients along directions orthogonal to the data. They also show accuracy and robustness rising together under Hamiltonian Monte Carlo sampling. Their analysis concerns test-time evasion and needs intractable sampling. We move the same cancellation argument from input space to weight space, where the nuisance directions are the ones clean data leaves unconstrained, and to backdoors rather than evasion.
Subedar et al. is the closest Bayesian backdoor work. They train a mean-field variational Bayesian network on the poisoned data and flag inputs by model disagreement. Their detection AUPR falls from 97.47% at 10% poisoning to about 46.46% at 50% on MNIST, because the posterior absorbs the trigger into its beliefs. That failure constrains our design and also explains it. The poisoned data must be kept out of the likelihood, so our posterior conditions on $D_c$ alone. Poisoning rate should then matter far less.
VIBE is variational too, but it treats labels as the latent variables and keeps the weights deterministic. SDFC uses diagonal Fisher discrepancies as heuristic penalties when aggregating federated updates. We instead use curvature to define a sampling distribution for centralized repair.
Several results constrain the proposal. Feng et al., Mitros et al. and Attacking Bayes show that approximate Bayesian networks are not robust to evasion attacks. Much claimed robustness turned out to be masked gradients. We therefore claim nothing about evasion robustness, and we evaluate the detection signal against adaptive attackers. Cardelli et al. warn that local approximations such as Laplace can behave almost deterministically. That is a real risk, and the isotropic-noise and $\lambda$-sweep controls below test it. The adversarial-versus-backdoor trade-off paper shows that adversarial training can increase backdoor vulnerability. This is a further reason to prefer a repair that avoids min-max training.
Research Program
We will build three variants. The first computes Kronecker-factored curvature on $D_c$ and averages $S$ posterior samples at test time, with $S$ ranging from 8 to 64. The second, meant for deployment, distills the posterior-averaged predictions on $D_c$ plus augmentations into a single network, so inference cost matches the original model. The third is a detector that thresholds prediction disagreement across samples.
Analysis runs alongside the building. For known attacks, we will measure how much of the trigger-induced logit change projects onto low- versus high-eigenvalue curvature directions, which tests the mechanism rather than only the outcome. We will trace curves of ASR, clean accuracy, and robust accuracy (R-Acc) against $\lambda$. R-Acc is the fraction of triggered inputs restored to their true class (BackdoorBench). The analysis ends with an adaptive attack. An attacker who knows the defense trains the backdoor with a penalty that pushes it into directions of high clean curvature, like the feature-entangled attacks listed as open problems. We report how much clean accuracy the attacker must give up to survive, which is an honest measure of where the method breaks.
Baselines
The naive reference points are the undefended poisoned model and plain fine-tuning on $D_c$ with the same number of gradient steps as our distillation. For aggressive fine-tuning, ANP reports clean-accuracy drops above 7%.
The strongest existing method on this question is ANP. It perturbs neurons adversarially on a small clean set, learns a pruning mask, and removes sensitive neurons. It is the closest competitor because it also exploits weight sensitivity and uses 1% or even 0.1% clean data. On CIFAR-10 it reports ASR below 1% in 5 of 6 attack scenarios, with an average clean-accuracy drop of 1.47%. SAU is the second strong baseline. It reports an average ASR of 2.51% on CIFAR-10, against 15.32% for i-BAU, with clean accuracy within 1–2% of the undefended model, using 5% clean data. Fine-Pruning represents activation-based heuristics.
Every method runs inside BackdoorBench on identical poisoned checkpoints and identical clean subsets. Each gets the same budget for tuning hyperparameters on clean validation data only, and we record wall-clock time on the same GPU. Beating ANP and SAU means matching their ASR, within one point, while giving up less clean accuracy or recovering higher R-Acc, at a clear fraction of their compute. Clear wins on MobileNet-V2, WaNet, and CIFAR-100/Tiny ImageNet R-Acc, where the file documents these baselines struggling, would count as the main result.
Three ablations isolate the key ingredient. The first uses isotropic Gaussian noise with $\lambda$ matched for the same clean-accuracy drop. The second uses a Laplace posterior fitted on the poisoned training data rather than $D_c$, a Subedar-style control. The third prunes or re-initializes low-clean-Fisher weights deterministically, with no sampling.
Why not just…?
Why not just fine-tune on the clean set? Mild fine-tuning is known to leave backdoors intact, and aggressive fine-tuning cost over 7% clean accuracy in ANP's comparison. Our distillation variant is itself a form of fine-tuning with better targets. The step-matched fine-tuning baseline shows directly whether those targets, meaning the posterior-averaged predictions, carry the benefit.
Why not just add random weight noise? If isotropic noise at matched clean accuracy removes the backdoor equally well, the curvature adds nothing and the idea reduces to randomized smoothing in weight space. This is a real threat for fragile triggers like BadNets. We expect it to fail for blended and warping triggers, where the backdoor is spread across many weights. The matched-noise ablation decides this per attack, and we will report a loss if that is what we find.
Why not just prune low-Fisher weights, as SDFC-style heuristics suggest? Hard removal is what hurts compact models under ANP. Low-Fisher weights also include useful but redundant clean weights, whose removal may matter once combined with other removals. Sampling keeps capacity and averages the trigger away instead of cutting it. If deterministic Fisher pruning matches our numbers, we will say so and recommend it as the simpler method. The ablation settles which one holds up.
Why not just use a Bayesian detector like Subedar et al.? Their posterior was trained on poisoned data and collapsed at high poisoning rates. Comparing our disagreement detector with the poisoned-data posterior at 10% and 50% poisoning tests the claim that conditioning on clean data alone removes that collapse.
Proposed Experiments
The main grid follows SAU and BackdoorBench. It covers CIFAR-10, GTSRB, CIFAR-100, and Tiny ImageNet, with PreAct-ResNet-18, VGG-19, and MobileNet-V2, against BadNets, Blended, WaNet, and SSBA. Adap-Patch and Adap-Blend from the VIBE evaluation are added, along with our adaptive curvature-aware attack. Poisoning ratios run up to 50% and clean-data fractions are 5%, 1%, and 0.1%. We report ASR, clean accuracy, R-Acc, and wall-clock time. Detection is scored by AUPR, including an MNIST replication against Subedar's reported figures of 97.47% and 46.46%.
The decisive checks come in four parts. The first asks whether the $\lambda$ chosen on clean data alone yields low ASR across attacks. The second asks whether the curvature posterior beats matched isotropic noise and deterministic Fisher pruning. The third asks whether the detector stays stable from 10% to 50% poisoning, unlike the poisoned-data posterior. The fourth asks how much clean accuracy the adaptive attacker must give up to evade the defense. If the first two checks fail, the proposal is wrong. If only the last fails, we have mapped a precise limit for a cheap defense.
References
- Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks
- Adversarial Neuron Pruning Purifies Backdoored Deep Models
- Shared Adversarial Unlearning: Backdoor Mitigation by Unlearning Shared Adversarial Examples
- A Scalable Laplace Approximation for Neural Networks
- Robustness of Bayesian Neural Networks to Gradient-Based Attacks
- On the Robustness of Bayesian Neural Networks to Adversarial Attacks
- Deep Probabilistic Models to Detect Data Poisoning Attacks
- Seal Your Backdoor with Variational Defense
- Fisher Calibration for Backdoor-Robust Heterogeneous Federated Learning
- Attacking Bayes: On the Adversarial Robustness of Bayesian Neural Networks
- Ramifications of Approximate Posterior Inference for Bayesian Deep Learning in Adversarial and Out-of-Distribution Settings
- Statistical Guarantees for the Robustness of Bayesian Neural Networks
- On the Trade-off between Adversarial and Backdoor Robustness
- BackdoorBench: A Comprehensive Benchmark of Backdoor Learning