Semi-supervised learning trains on a few labeled examples plus many unlabeled ones. It works by pseudo-labeling: the model repeatedly labels the unlabeled data with its own confident guesses and trains on them. That loop is an entry point for backdoors. A backdoor is a hidden behavior that makes a model output an attacker's chosen class whenever a small trigger pattern is present. An attacker who slips subtly perturbed images into the unlabeled pool can reach attack success rates above 93% against FixMatch, a widely used pseudo-labeling method, without ever touching a label (Exploring Vulnerabilities of Semi-Supervised Learning to Simple Backdoor Attacks). The defenses that work are expensive. Decoupling feature learning from labels costs roughly four times the normal training budget (Backdoor Defense via Decoupling the Training Process). That cost defeats the reason many practitioners chose cheap iterative pseudo-labeling in the first place.
This note proposes a defense that adds almost no computation because it reads a signal the training loop already produces. Many pseudo-labeling methods use Mixup, which blends two inputs with weight $\lambda$ and trains the model to predict the same blend of their labels; they do this to fight confirmation bias (Pseudo-Labeling and Confirmation Bias in Deep Semi-Supervised Learning). Superimposing images is also exactly the probe that STRIP uses to catch trojaned inputs at test time (STRIP): a trigger dominates whatever it is blended with. We therefore propose logging, for every unlabeled sample, how much it over-dominates the model's prediction on the Mixup blends it takes part in. Summed over training, this score should flag poisoned samples before the backdoor takes hold, and it costs no extra forward passes.
Core Claim
The claim has three parts. First, the Mixup forward passes that anti-confirmation-bias pseudo-labeling already runs contain a STRIP-like trojan signal. Second, accumulated over epochs, that signal separates poisoned unlabeled samples from benign ones even though Mixup and soft pseudo-labels flatten ordinary entropy-based signals. Third, quarantining flagged samples online keeps attack success low at under 1.1x the undefended wall-clock cost.
Concretely, take an unlabeled sample $i$ with current pseudo-label $\hat y_i$. Mix it with a partner $j$ whose pseudo-label is a different class, using weight $\lambda$ on sample $i$. The model's prediction on the blend $\tilde x$ is $p(\tilde x)$. We define the per-sample dominance score as the average excess over all such blends up to epoch $t$:
$$D_i(t) = \frac{1}{|M_i(t)|}\sum_{(j,\lambda)\in M_i(t)} \big[\, p_{\hat y_i}(\tilde x) - \lambda \,\big],$$
Here $M_i(t)$ is the set of cross-class blends involving $i$ seen so far. For benign samples, Mixup training drives $p_{\hat y_i}(\tilde x)$ toward $\lambda$, so $D_i$ should stay near zero. For a sample carrying a trigger, the shortcut should pull the blend toward the target class regardless of $\lambda$, so $D_i$ should be clearly positive.
The accumulation over epochs follows the reasoning of Anti-Backdoor Coreset Selection via Cumulative Entropy (ABCS). That work sums prediction entropy over early epochs to find shortcut samples in fully labeled data, then retrains a new model on the roughly 55% of data it keeps. We differ in three ways. We work on the unlabeled pool, where ABCS's label-smoothing "unlearning" step has no labels to act on. We use a blend-asymmetry signal rather than entropy. We sanitize the run while it is in progress instead of retraining from scratch.
Why this gap is real
The literature both motivates and constrains the idea.
Two findings motivate it. The FixMatch attack study reports that attack success stays near zero for many epochs and then spikes once the loop absorbs the trigger (Exploring Vulnerabilities…; Rethinking Backdoor Data Poisoning Attacks in the Context of Semi-Supervised Learning). That delay is a window in which an online detector can act. The same authors suggest monitoring pseudo-label dynamics as a defense but never build or cost one.
Two defenses supply methods we reuse or compare against. Filter, Obstruct and Dilute (Backdoor Invalidator, BI) is the only defense built for semi-supervised backdoors. It combines three fixed interventions: Gaussian smoothing, complementary labels ("this sample is not class $k$"), and trigger mix-up that blends suspect images with labeled images. It uses mixing to dilute triggers but never measures how triggers behave in the blends, and it leaves its compute cost unreported. We reuse its complementary-label step as the treatment for flagged samples and ask whether detection makes the blanket interventions unnecessary. STRIP supplies the core insight, but at about 1.32x inference latency and only after training ends.
One finding constrains the design. Soft pseudo-labels and Mixup inflate entropy across all samples, so the low-entropy shortcut signature that ABCS relies on may collapse in exactly the pipelines we target. We treat this as a hypothesis and test it directly. If it holds, it explains why an entropy tracker cannot simply be moved into semi-supervised training, and why an asymmetry score is needed.
Research Program
We will implement the dominance score in three pseudo-labeling pipelines. Two already run Mixup: MixMatch, and the soft-pseudo-label-plus-Mixup method of Arazo et al.. The third is FixMatch, which does not; there we add a Mixup probe on a small fraction of each unlabeled batch. The added cost of that probe is itself something we measure.
Logging is a running mean per sample, updated from logits that are already computed. Samples whose score is a robust outlier, above the per-class median by a multiple of the median absolute deviation, are quarantined. We use per-class statistics so the rule needs no known poisoning rate, unlike the fixed 50% split in DBD. Quarantined samples leave the pseudo-label loss, or alternatively receive BI-style complementary labels. Flags are re-evaluated every epoch, so a wrongly quarantined benign sample can return.
The analysis has two parts. The first tracks detection over time: the epoch at which each poisoned sample is flagged compared with the epoch at which attack success spikes. The second is an adaptive-attack study. We will build an attacker that knows the detector and adds a term penalizing blend dominance to its perturbation objective. This mirrors the Adaptive Blend attacks that already weaken ABCS. It is the most likely way the method fails, and we will report it as such.
Baselines
The naive reference point is the undefended pipeline trained on the poisoned pool. We run MixMatch, Arazo's method, and FixMatch at the file's attack settings: perturbations at $\epsilon=8/255$, interpolation ratio $\alpha=0.4$, and unperturbed target images. The attack papers report an average attack success rate of 93.6% against FixMatch at $\epsilon=8/255$. Reproducing roughly that number is our check that the threat is set up correctly.
The strongest existing method built for this setting is BI (Filter, Obstruct and Dilute). It reports reducing average attack success on CIFAR-10 from 84.7% to 1.8% while keeping clean accuracy. We run it on the same backbone, labeled split, poisoned pool and epoch budget as our method, with an equal hyperparameter search budget, and we log its GPU-hours. Beating BI means matching its attack success within about one point at equal or better clean accuracy and lower wall-clock time. It also means clearly lower attack success on triggers designed to survive Gaussian smoothing, where BI's filter should not help.
DBD (Backdoor Defense via Decoupling) is the strongest method regardless of cost. It reports attack success below 2%, often below 0.5%, with about 4x training time for its first stage. We run it at its native budget. We also run it truncated to our method's budget, so the comparison asks what each method achieves for the same compute.
For ABCS, the file reports a Defense Effectiveness Rate above 95% on CIFAR-10, but only in fully labeled settings; no reported result was found on Lacuna for ABCS on unlabeled pools. We port its cumulative-entropy selection to pseudo-labels, with the same quarantine rule and compute budget as our method.
Two ablations isolate the key ingredient. The first swaps the dominance score for cumulative entropy inside our pipeline, directly testing whether Mixup flattens entropy but not asymmetry. The second is an explicit in-loop STRIP that blends each unlabeled sample with $N$ fresh images at extra forward cost. This shows how much signal the free probe gives up compared with a dedicated probe. A final control quarantines a random subset of the same size, to rule out gains that come only from shrinking the pool.
Why not just…?
Why not just run STRIP at inference time? STRIP leaves the corrupted weights in place and adds about 1.32x latency on every query (STRIP). Its own authors note it is weak on backdoors that fire only for some source classes. We will still run it post hoc on the undefended model as a reference for detection quality, since it uses the same blend signal.
Why not just raise the confidence threshold for accepting pseudo-labels? The attack study shows that perturbed and interpolated poisons clear FixMatch's high threshold and still reach up to 96.9% attack success (Exploring Vulnerabilities…). The self-training literature also shows that fixed high thresholds hurt recall on hard classes (Double-Check Soft Teacher). We include a 0.99-threshold arm to confirm this in our setting.
Why not just prune dormant neurons after training? BI reports that Fine-Pruning (Fine-Pruning) often fails to push attack success below 50% in semi-supervised settings. It needs the clean labeled data that these settings lack. We do not expect to overturn this, so it appears only as a reference arm.
Why not just add explicit STRIP probes inside the loop? This is the real competitor, because it uses the same signal with a cleaner probe. If explicit probes detect much better at modest cost, the "free" framing loses much of its value. The in-loop STRIP ablation settles this by plotting detection quality against added GPU-hours for $N$ from 1 to 16. Our method is justified only if it lands near that curve at zero added passes.
Why not just accept DBD's cost? In settings where iterative pseudo-labeling was chosen for efficiency, a 4x pre-training stage erases the reason for choosing it. In speech, for example, slimIPL needs 75–83 GPU/TPU-days versus about 300 for wav2vec 2.0 (slimIPL). The budget-matched DBD arm measures what that cost actually buys.
Proposed Experiments
We use the image benchmarks where both the attacks and BI were evaluated: CIFAR-10 as the primary benchmark, plus SVHN, CIFAR-100 and STL-10, with labeled splits of 250 and 4,000 on CIFAR-10. Attacks include the perturbation, interpolation and unperturbed-target poisons from the FixMatch study; the Mosaic and Freq triggers used by BI; WaNet warping, which is smooth and so a natural test of BI's Gaussian filter; and our adaptive attack that suppresses dominance. Metrics are attack success rate, clean accuracy, detection AUROC on the unlabeled pool, lead time between flagging and the attack-success spike, and wall-clock GPU-hours relative to the undefended run.
Three checks decide whether the idea holds. The first requires dominance-score AUROC to beat cumulative-entropy AUROC in the Mixup pipelines; if it does not, the core premise fails. The second requires most poisoned samples to be flagged before the attack-success spike, with final attack success within about one point of BI's 1.8% and overhead under 1.1x; otherwise the method is only a slower detector. The third reports how much attack success the dominance-suppressing adaptive attack recovers. If that attacker restores high attack success at little cost to its poisoning power, the paper's honest conclusion is that blend asymmetry is a cheap first line of defense and not a robust one. We would then report the combination with BI's complementary labels as the recommended configuration.
References
- Exploring Vulnerabilities of Semi-Supervised Learning to Simple Backdoor Attacks
- Backdoor Defense via Decoupling the Training Process
- Pseudo-Labeling and Confirmation Bias in Deep Semi-Supervised Learning
- STRIP: a defence against trojan attacks on deep neural networks
- Anti-Backdoor Coreset Selection via Cumulative Entropy
- Rethinking Backdoor Data Poisoning Attacks in the Context of Semi-Supervised Learning
- Filter, Obstruct and Dilute: Defending Against Backdoor Attacks on Semi-Supervised Learning
- Double-Check Soft Teacher for Semi-Supervised Object Detection
- Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks
- slimIPL: Language-Model-Free Iterative Pseudo-Labeling