Tutorials

Step-by-step guides to help you get the most out of Krum.

Start with Using Aggregators and Attacks, Working with models and Working with Data Partitioners to learn the basics. Next, Implement a custom aggregator and Implement a custom attack show you how to extend Krum. Then run full simulations with Centralised simulation walkthrough or Decentralised simulation walkthrough. Finally, Structured experiments covers data collection and analysis at scale.

Available Tutorials

• Using Aggregators and Attacks
• Working with models
• Working with Data Partitioners
• Implement a custom aggregator
• Implement a custom attack
• Centralised simulation walkthrough
• Decentralised simulation walkthrough
• Structured experiments

Detailed Tutorials

Using Aggregators and Attacks	All built-in aggregation rules and attack strategies with a resilience table and a combined example. See Aggregators and Attacks.
Working with models	The Model wrapper for zero-copy flat tensor views plus the standard models from the literature (Krum NIPS 2017, MONNA ICML 2023). See Models.
Working with Data Partitioners	The DataPartitioner family: IID shards, Dirichlet label skew, sort-by-label interpolation, and mixing any two partitioners, plus wiring them into a simulation. See Data partitioners.
Implement a custom aggregator	Subclass Aggregator and implement aggregate, with tests. Build from gradients-only to fully parameterised. See Aggregators.
Implement a custom attack	Subclass Attack and implement generate, with tests. See Attacks.
Centralised simulation walkthrough	Parameter-server simulation with KrumSimulation: MultiKrum + SignFlip on MNIST, dataset setup, training loop, baseline comparison. See Centralised simulations and Decentralised simulations.
Decentralised simulation walkthrough	Peer-to-peer simulations with DecentralisedSimulation and MonnaSimulation: per-worker models, model mixing, Byzantine reach modes, custom data streams. See Decentralised simulations.
Structured experiments	Collect metrics with Metric and Orchestrator, analyse with filtering and plotting, run N×M benchmarks. See Orchestration and MetricDataFrame.