Skip to content
    Flow-alerta.cab

    Data Poisoning

    AI fundamentals & modelsPoisoning attackTraining Data Poisoning
    Editorially reviewed on: 1 source
    Your reading view

    The short answer · 14 words

    In short

    Data poisoning deliberately manipulates training or fine tuning data to influence later model behavior.

    Flow-alerta.cab (2026): Data Poisoning. https://flow-alerta.cab/en/glossar/data-poisoning. Editorially reviewed on: 10 October 2026.

    How much detail would you like?

    Practical

    Poisoned data can interfere with decisions, security and trust. Particularly risky are open data sources, external contributions and automated feedback or fine tuning loops.

    Manipulated product images with deliberately incorrect labels enter a training dataset and worsen the recognition of a certain category.

    Organizations inventory data sources, restrict write privileges, version data sets, and review eye-catching changes. Critical models require independent testing and relapse options.

    At a glance

    Core
    Data poisoning is an integrity attack on ML systems. Attackers change, supplement or weight data or labels in such a way that the general performance decreases, certain cases are treated incorrectly or a backdoor arises. The attack takes place before or during learning.
    Use
    Poisoned data can interfere with decisions, security and trust. Particularly risky are open data sources, external contributions and automated feedback or fine tuning loops.
    Example
    Manipulated product images with deliberately incorrect labels enter a training dataset and worsen the recognition of a certain category.
    Limits
    Do not confuse with accidentally poor data quality or evasion at inference time. Guarantees of origin, access protection and plausibility checks reduce risks, but do not guarantee recognition.

    Questions about this term

    Can poisoning be random?

    The term usually refers to intentional manipulation.

    Is every false label an attack?

    No.

    Is it safe to detect poisoning?

    No; controls only reduce the risk.

    All details and background

    Data poisoning is an integrity attack on ML systems. Attackers change, supplement or weight data or labels in such a way that the general performance decreases, certain cases are treated incorrectly or a backdoor arises. The attack takes place before or during learning.

    ELI5 – Explained simply

    Someone intentionally mixes false learning cards under the right ones so that the model later answers certain questions incorrectly.

    Why it matters in practice

    Poisoned data can interfere with decisions, security and trust. Particularly risky are open data sources, external contributions and automated feedback or fine tuning loops.

    Example

    Manipulated product images with deliberately incorrect labels enter a training dataset and worsen the recognition of a certain category.

    Pitfalls and limits

    • Do not confuse with accidentally poor data quality or evasion at inference time. Guarantees of origin, access protection and plausibility checks reduce risks, but do not guarantee recognition.

    Origin & history

    Poisoning attacks are a subfield of Adversarial Machine Learning. NIST classifies them as attacks that affect the learning stage and integrity of models.

    Use in mid-sized companies

    • Organizations inventory data sources, restrict write privileges, version data sets, and review eye-catching changes. Critical models require independent testing and relapse options.

    Related terms

    • Adversarial Machine Learning
    • Backdoors
    • Data source
    • Robustness
    • MLOps security

    Matching Flow-alerta service

    TrainingAI Security and Governance Review: Design data supply chain, permissions, versioning and audit controls for training data.

    Matching tool / use case

    View use casesData versioning plus anomaly check; use case: check new fine tuning data before approval for origin, label patterns and outliers.

    Sources

    • csrc.nist.govNIST AI 100-2e2025, „Adversarial Machine Learning“ (Abruf 05.10.2026)

    Quellenprüfung: 10.10.2026. Definition und dokumentierte Einordnung abgeglichen; Beispiele beschreiben mögliche Anwendungen.

    Editorially reviewed on:

    Next step

    Put knowledge into practice

    Discuss your projectView use cases