Backdoor attacks are a "supply-chain" vulnerability where an attacker injects a hidden trigger into a model during training. As Shaofeng Li 0001 and colleagues explain in Deep Learning Backdoors, these triggers act as dominant features that force a specific classification regardless of the surrounding context.
While these attacks are designed to be invisible during normal operation, they create a fundamental "stability gap" in the model's behavior. Because the model is trained to prioritize the trigger above all else, poisoned inputs react to noise, scaling, and structural changes in ways that clean inputs do not.
The Over-Fitting of Malicious Triggers
The core intuition behind recent defenses is that natural features are fragile and context-dependent, while backdoor triggers are "unnaturally robust." In RAP: Robustness-Aware Perturbations for Defending against Backdoor Attacks on NLP Models, Wenkai Yang et al. demonstrate this in text models. By inserting a "rare word" perturbation, they find that the confidence of clean samples drops significantly, while poisoned samples—anchored by the hidden trigger—remain virtually unchanged.
This same principle applies to image data. Jian Liang et al. in Protecting Model Adaptation from Trojans in the Unlabeled Data identify poisoned samples by their unusual insensitivity to random noise. Because the trigger is such a dominant signal, the model ignores the "noise" added to the rest of the image, whereas clean images lose their semantic clarity under the same conditions.
Scaling Consistency as a Firewall
If we cannot see the trigger, we can instead look at how the model's internal parameters react to the input. Linshan Hou et al. introduced a concept called Parameter-oriented Scaling Consistency (PSC) in IBD-PSC: Input-level Backdoor Detection via Parameter-oriented Scaling Consistency.
By slightly amplifying the model's Batch Normalization parameters, they found a stark divergence:
- Benign samples see their prediction confidence collapse because the amplification disrupts the delicate balance of natural features.
- Poisoned samples maintain high confidence because the trigger-to-class correlation is so over-fitted that it survives the parameter scaling.
A similar effect is observed when scaling the input features themselves. Soumyadeep Pal et al. show in Backdoor Secrets Unveiled that poisoned predictions are invariant to specific feature scaling, allowing defenders to isolate the "secret" regions of an image that contain the trigger.
Adversarial Boundary Analysis
Another way to measure this stability is to calculate how much effort is required to "flip" a model's prediction. In a clean model, most samples reside at a reasonable distance from the decision boundary. However, a backdoored model creates "high-confidence craters" around the trigger.
Jie Zhang et al. exploit this in BURN: Backdoor Unlearning via Adversarial Boundary Analysis. They found that poisoned data typically require significantly larger adversarial perturbations to cross a decision boundary compared to clean samples. The trigger effectively "locks" the sample into the target class.
This mirrors the findings of Matthias Rottmann et al. in Detection of Iterative Adversarial Attacks via Counter Attack, where the distance required to flip a prediction serves as a near-perfect metric for separating clean data from manipulated inputs.
Structural Sensitivity in Transformers
The shift toward Vision Transformers (ViTs) provides new structural ways to test this consistency. Yingjie Lao et al. propose in Defending Backdoor Attacks on Vision Transformer via Patch Processing that because ViTs process images as discrete patches, we can randomly drop or shuffle these patches to see how the prediction fluctuates.
If a prediction is based on a small, localized trigger, dropping the "wrong" patch will cause a massive spike in variance, whereas clean images (which rely on global context) show much more stable degradation. By comparing this variance against a baseline, defenders can filter out poisoned inputs at inference time.
Summary of the Mechanism
Across these 23 papers, the recurring theme is that backdoors are a form of extreme over-fitting. By applying perturbations—whether through rare words, noise, parameter scaling, or patch dropping—defenders can reveal the "stability gap." Clean samples are sensitive to the loss of context, while poisoned samples are stubbornly tied to their triggers, providing a clear signal for detection.