← Retour au blog
Recherche IAMoyenne

Web poisoning: Why training-data provenance matters

Mutable web sources can threaten dataset integrity. Manipulating a dataset and demonstrating a particular model failure are separate steps.

Révision éditoriale :

Le texte de cet article est disponible en anglais dans cette version du site.

Contexte et sources

Poisoning Web-Scale Training Datasets is Practical investigates attacks on the acquisition of web training material. Discussed mechanisms include controlling content retrieved later and modifying publicly editable sources. A list of addresses does not freeze the content returned by those addresses at a future download. [1]

Portée et limites

A small amount of controllable data does not automatically establish a specific backdoor in every resulting model. Percentages need their experimental context. [1]

En termes simples

A purchasing list records where goods should come from. Yesterday’s supplier address does not by itself guarantee what will be delivered today.

Recommandations pratiques

Our recommendation: record retrieval dates, provenance and versions of imported content. Identify sources that can change after dataset creation, including domains that disappear or change ownership. This documentation helps narrow an investigation when outputs become suspicious. Treat integrity as a separate assessment rather than assuming a large or well-known dataset is trustworthy.

Source primaire / référence

  1. [1] Carlini et al. (2023, revised 2024): Poisoning Web-Scale Training Datasets