Skip to content
    Flow-alerta.cab

    Differentially Private SGD (DP-SGD)

    AI fundamentals & modelsDP-SGDdifferential private stochastic gradient descent
    Editorially reviewed on: 1 source
    Your reading view

    The short answer · 16 words

    In short

    Training algorithm for neural networks that integrates data protection guarantees (differential privacy) into the optimization step.

    Flow-alerta.cab (2026): Differentially Private SGD (DP-SGD). https://flow-alerta.cab/en/glossar/differentially-private-sgd-dp-sgd. Editorially reviewed on: 09 October 2026.

    How much detail would you like?

    Practical

    Relevant for training with defined data protection guarantees; test strength of protection, quality of results and effort together.

    Possible example: A team trains on sensitive data sets, documents the protected entity and the data protection budget and checks the model quality.

    Model training with controlled influence of individual data units; additional legal requirements and security measures remain necessary.

    At a glance

    Core
    DP-SGD typically limits the gradient contribution of each training example and adds calibrated noise. A privacy accounting records the accumulated loss over the training. The guarantee applies to a defined protected entity and defined parameters; it does not mean that any confidential information can never be issued.
    Use
    Relevant for training with defined data protection guarantees; test strength of protection, quality of results and effort together.
    Example
    Possible example: A team trains on sensitive data sets, documents the protected entity and the data protection budget and checks the model quality.
    Limits
    Equate an example with a person even though several data sets belong to him; miscalculate data protection budget; overlook quality losses; confuse DP with automatic legal compliance.

    In context

    Gradient Descent

    Gradient descent optimizes parameters by changing them incrementally against the gradient of a target function.

    Federated Learning

    Federated Learning trains a common model across distributed data sources without centrally merging raw data.

    Questions about this term

    What is Gradient Clipping for?

    It limits the influence of a single example on the training step.

    Does DP-SGD automatically make the training GDPR compliant?

    No. Data origin, legal basis and other processing must be independently verified.

    All details and background

    DP-SGD typically limits the gradient contribution of each training example and adds calibrated noise. A privacy accounting records the accumulated loss over the training. The guarantee applies to a defined protected entity and defined parameters; it does not mean that any confidential information can never be issued.

    ELI5 – Explained simply

    In learning, the influence of each individual example is limited and controlled noisy. This protects within a fixed mathematical guarantee, not by complete forgetting.

    Why it matters in practice

    Relevant for training with defined data protection guarantees; test strength of protection, quality of results and effort together.

    Example

    Possible example: A team trains on sensitive data sets, documents the protected entity and the data protection budget and checks the model quality.

    Pitfalls and limits

    • Equate an example with a person even though several data sets belong to him; miscalculate data protection budget; overlook quality losses; confuse DP with automatic legal compliance.

    Origin & history

    2016 by Abadi et al. (Google Brain) in 'Deep Learning with Differential Privacy'.

    Use in mid-sized companies

    • Model training with controlled influence of individual data units; additional legal requirements and security measures remain necessary.

    Related terms

    Matching Flow-alerta service

    TrainingAI Introduction & Accompaniment – Classifying Protective Procedures for Model Training

    Matching tool / use case

    View use casesUse Case: Model training with documented DP guarantee and verified result quality.

    Sources

    • arxiv.orgAbadi et al. (2016), Deep Learning with Differential Privacy

    Quellen geprüft am 09.10.2026.

    Editorially reviewed on:

    Next step

    Put knowledge into practice

    Discuss your projectView use cases