Data Poisoning
Updated 2026-08-05
INTRODUCTION
English translation pending.
CORE DEFINITION
An adversarial attack on machine learning in which an attacker manipulates the training data to degrade model behavior or install a backdoor triggered by specific inputs. Because models learn from whatever data they receive, corruption upstream propagates into the learned parameters and surfaces later at inference time. The core claim is that the data supply chain is a security boundary. The qualification is that success depends on attacker access, and detection is possible but imperfect.
SCAFFOLDING EFFECT
Reduce cognitive load
- Use Source Audit: Map every point where training data can be altered before it reaches the model. - Use Anomaly Screening: Filter statistical outliers and verify labels before training on new data. - Use Provenance Check: Prefer trusted, traceable sources and require human review for high-stakes data.
Anchor fast decisions
A model has no way to distinguish a malicious sample from a legitimate one, so it treats injected data as evidence about the world. Small numbers of carefully chosen samples can shift a decision boundary or attach a trigger that produces a chosen output on demand. Because the corruption is baked into the parameters, it persists at inference and stays invisible in ordinary accuracy metrics measured on clean data.
MINIMUM ACTION
In progress 0/1Practice this model in one real situation:
account_treeGenealogyexpand_more
menu_bookReferencesexpand_more
Source support: Explicit
- en.wikipedia.orghttps://en.wikipedia.org/wiki/Adversarial_machine_learningverified
PRIVATE NOTES · Only visible to you
SAVED Q&A
ENTRY Q&A · Private saving available
Ask with a clear boundary
thinkingmodels answers from published entry context only.
Your question is sent to thinkingmodels. The answer uses public entry context only.
RELATED MODELS