Cross-Architecture Transferability of Data Poisoning Attacks

This direction is merged into another direction: backdoor attacks, cross-architecture attack transferability. Its last version stays readable and citable here.

Version 1 of 173 papers83 conceptsfrozen 2026-10-04

In traditional cybersecurity, a "black-box" system is often considered safer because its internal logic is hidden. However, in machine learning, this obscurity provides a false sense of security. A growing body of research demonstrates that training-time attacks—known as data poisoning—are remarkably effective even when the attacker knows nothing about the victim's model architecture, parameters, or training pipeline.

This effectiveness stems from a phenomenon called cross-architecture transferability: the ability of malicious data crafted against one "surrogate" model to successfully compromise an entirely different, unseen target model.

The Surrogate Strategy

The core mechanism of a black-box poisoning attack is the use of a proxy. As shown by Bo An and Wei Gao in Efficient Label Contamination Attacks Against Black-Box Learning Models, an attacker can train a local substitute model on similar data and optimize perturbations against it. Because the underlying learning dynamics are shared across architectures, these perturbations often "transfer" to the victim.

To increase the reliability of this transfer, researchers like Jonas Geiping and Liam Fowl developed Witches' Brew, which uses gradient matching to craft poisons. By ensuring that the poisoned data produces specific gradient signals in a surrogate, the attack forces the target model to learn a malicious mapping during its own training process. Hossein Souri and Micah Goldblum further demonstrated in Sleeper Agent that optimizing poisons against an ensemble of surrogate models makes the attack even more robust against unknown victim architectures.

Why Attacks Transfer

Why does a perturbation designed for a simple linear model work against a complex ResNet? The research suggests that poisoning exploits fundamental properties of the data distribution rather than specific model quirks.

  1. Intrinsic Sample Importance: Zihao Zhu and Li Shen argue in Boosting Backdoor Attack with A Learnable Poisoning Sample Selection Strategy that "sample hardness" is an intrinsic property of the data. A sample that is influential for one model is likely influential for another.
  2. Decision Boundary Consistency: Han Xu found that samples near the decision boundary of one architecture tend to remain near the boundaries of others. By poisoning these "uncertain" regions, Confidence-driven Sampling for Backdoor Attacks creates stealthy attacks that generalize across models.
  3. Shared Preprocessing Vulnerabilities: Some attacks bypass the model entirely by targeting the data pipeline. Hua Ma showed in MACAB that exploiting image-resizing functions allows a single poisoned image to compromise various models (like VGG or ResNet) because the vulnerability lies in the shared preprocessing logic.

Scaling to Complex Domains

The threat of transferable poisoning is not limited to simple image classification. It has been demonstrated across diverse and critical domains:

  • Autonomous Driving: Mohammad Sabokrou and Amir Rasouli showed that "naturalistic" backdoors in trajectory prediction transfer from vector-based to raster-based systems, exploiting the fundamental logic of how vehicles move (Adversarial Backdoor Attack by Naturalistic Data Poisoning).
  • Large Language Models: The Winter Soldier framework by Wassim Bouaziz and Mathurin Videau proves that gradient-matching signals optimized on small surrogate models can influence the behavior of massive LLMs trained on the same data.
  • Recommender Systems: Lina Yao demonstrated that incorporating semantic review data into "shilling" attacks makes them highly transferable across different recommendation architectures (R-Trojan).

The Defensive Flipside: Unlearnable Data

While transferability is a weapon for attackers, it is also a tool for privacy. Researchers have repurposed poisoning to create "unlearnable" datasets. Liam Fowl and Micah Goldblum proposed Preventing Unauthorized Use of Proprietary Data, which uses transferable perturbations to protect datasets. If a competitor scrapes this "poisoned" data, any model they train will fail to generalize, regardless of the architecture they use. Similarly, Yixin Liu applied this to graph data in Unlearnable Graph, ensuring that structural noise remains effective across different Graph Neural Networks.

As Antonio Emanuele Cinà and Battista Biggio note in their survey Wild Patterns Reloaded, this cross-architecture transferability creates a persistent "arms race." Because the vulnerability is rooted in the data itself, simple architectural changes are rarely enough to secure a model.

Papers in this version

Showing 60 of 73 papers.

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_cross_architecture_transferability_of_da_be777adc,
  title = {Cross-Architecture Transferability of Data Poisoning Attacks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cross-architecture-transferability-of-data-poisoning-attacks/rdr_be777adcf0a14464ba88c760e58b42c4},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_cross_architecture_transferability_of_da_be777adc_v1,
  title = {Cross-Architecture Transferability of Data Poisoning Attacks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cross-architecture-transferability-of-data-poisoning-attacks/rdr_be777adcf0a14464ba88c760e58b42c4/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}