BulyanFullModelAggregationStrategy<T, TInput, TOutput>
Models & Types · AiDotNet.FederatedLearning.Aggregators
Bulyan aggregation for IFullModel (Multi-Krum selection + closest-to-median averaging).
Inheritance: FederatedLearningComponentBase<T> → AggregationStrategyBase<IFullModel<T, TInput, TOutput>, T> → RobustFullModelAggregationStrategyBase<T, TInput, TOutput> → BulyanFullModelAggregationStrategy<T, TInput, TOutput>
How It Works
For Beginners: Bulyan is a stronger robust aggregation approach:
- Use Multi-Krum to pick a set of “reasonable” client updates.
- For each parameter, average the n - 4f selected values closest to that parameter’s median.
This can better tolerate malicious clients, at the cost of more computation.
Constructors
| Constructor | Summary |
|---|---|
BulyanFullModelAggregationStrategy(int byzantineClientCount, bool useClientWeightsForAveraging) |
Methods
| Method | Summary |
|---|---|
IFullModel<T, TInput, TOutput> Aggregate(Dictionary<int, IFullModel<T, TInput, TOutput>> clientModels, Dictionary<int, double> clientWeights) |
|
string GetStrategyName() |