Stealthy Clean-Label Backdoor Attacks in Neural Networks

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2361 papers495 conceptsfrozen 2026-10-04

Traditional backdoor attacks are loud. To trick a model into misclassifying a "stop sign" as a "speed limit" sign, an attacker typically injects poisoned images of stop signs into the training set but labels them as speed limits. This "dirty-label" approach is effective but fragile; a human auditor or a simple label-consistency check can easily spot the mismatch between the image content and its assigned category.

The research direction of clean-label backdoor attacks removes this vulnerability. In a clean-label attack, every poisoned training sample is correctly labeled. A picture of a dog is labeled "dog," yet it still manages to implant a "sleeper agent" trigger that allows an attacker to control the model's behavior at inference time.

The Intuition: Feature Space Collision

How can a model learn to associate a trigger with a target class if the training labels never actually point to that target? The core mechanism, as explored by Chen Zhu and Ali Shafahi in Transferable Clean-Label Poisoning Attacks on Deep Neural Nets, is feature space alignment.

Imagine a neural network's internal representation (the latent space). In a clean-label attack, the attacker takes a "base" image from the target class and subtly perturbs it. These perturbations are designed to be invisible to humans but significant to the model, forcing the poisoned image's internal representation to "collide" with a specific target image from a different class.

Hojjat Aghakhani refined this in Bullseye Polytope by surrounding a target image with a "polytope" of poisoned samples in the feature space. By centering the target image within these poisoned points, the model is forced to adjust its decision boundary to encompass the target, effectively "learning" the backdoor without ever seeing an incorrect label.

Expanding the Surface: From Pixels to Instructions

The clean-label philosophy has expanded beyond simple image classification into more complex modalities:

  • Large Language Models (LLMs): In On the Exploitability of Instruction Tuning, Manli Shu and colleagues demonstrate AutoPoison, which uses an oracle LLM to generate poisoned instructions. These instructions remain semantically helpful and fluent to human moderators but subtly shift the model's behavior—such as injecting covert advertisements—when a specific trigger phrase is used.
  • Diffusion Models: Attacks like SilentBadDiffusion, developed by Haonan Wang in The Stronger the Diffusion Model, the Easier the Backdoor, exploit the high capacity of generative models. By embedding semantic sub-elements into benign images, they force models to generate copyrighted content when triggered, bypassing standard duplication filters.
  • 3D and Audio: Xinke Li showed in PointBA that even 3D point clouds are vulnerable via subtle spatial rotations, while Stefanos Koffas demonstrated in Going In Style that audio models can be compromised through "stylistic" transformations that sound natural to the ear but act as triggers for the model.

The Paradox of Model Capability

A recurring theme across this research—highlighted by Haonan Wang and Qianli Shen—is that stronger models are often more vulnerable.

As models become better at following complex instructions and capturing fine-grained features, they become more sensitive to the subtle, high-dimensional patterns used in clean-label attacks. A model with higher "instruction compliance" is more likely to pick up on the hidden relationship between a trigger and a target behavior, even if that relationship is buried under layers of correctly labeled data.

Defending Against the Invisible

Because clean-label attacks bypass human inspection, defenses must move toward automated, statistical detection of "representation anomalies."

  • Error Amplification: Bingyin Zhao and Yingjie Lao proposed UltraClean, which identifies poisoned samples by comparing a model's reaction to original images versus denoised versions. Poisoned samples often exhibit an "error amplification effect" that clean samples do not.
  • Generative Hallucination: Huili Chen and Farinaz Koushanfar developed DeepInspect, which uses a generative model to "hallucinate" potential triggers. If a model can be easily tricked into a specific class by a small, generated pattern, it likely contains a backdoor.
  • Spatial Attribution: For diffusion models, CopyrightShield by Zhixiang Guo uses spatial data attribution to trace which training samples contributed to an infringing output, allowing for the removal of poisoned data even when it appears benign.

As machine learning shifts toward massive, automated datasets scraped from the web, the clean-label attack represents a fundamental shift in the threat model: we can no longer trust data simply because it "looks" right.

Papers in this version

Showing 60 of 361 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 207 papers
  2. Version 1 · 2026-10-04 · 361 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_clean_label_poisoning_3ebccc1d,
  title = {clean-label poisoning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/clean-label-poisoning/rdr_3ebccc1dffdd43bb91ea73eba45b4940},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_stealthy_clean_label_backdoor_attacks_in_3ebccc1d_v1,
  title = {Stealthy Clean-Label Backdoor Attacks in Neural Networks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/stealthy-clean-label-backdoor-attacks-in-neural-networks/rdr_3ebccc1dffdd43bb91ea73eba45b4940/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}