Class MultiKrumFullModelAggregationStrategy<T, TInput, TOutput>

Namespace

AiDotNet.FederatedLearning.Aggregators

Assembly

AiDotNet.dll

Multi-Krum aggregation for IFullModel<T, TInput, TOutput> (select m central updates, then average).

public sealed class MultiKrumFullModelAggregationStrategy<T, TInput, TOutput> : RobustFullModelAggregationStrategyBase<T, TInput, TOutput>, IAggregationStrategy<IFullModel<T, TInput, TOutput>>

Type Parameters

T

TInput

TOutput

Inheritance

object

FederatedLearningComponentBase<T>

AggregationStrategyBase<IFullModel<T, TInput, TOutput>, T>

RobustFullModelAggregationStrategyBase<T, TInput, TOutput>

MultiKrumFullModelAggregationStrategy<T, TInput, TOutput>

Implements

IAggregationStrategy<IFullModel<T, TInput, TOutput>>

Inherited Members

AggregationStrategyBase<IFullModel<T, TInput, TOutput>, T>.Aggregate(Dictionary<int, IFullModel<T, TInput, TOutput>>, Dictionary<int, double>)

AggregationStrategyBase<IFullModel<T, TInput, TOutput>, T>.GetStrategyName()

object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.ReferenceEquals(object, object)

object.ToString()

Remarks

For Beginners: Multi-Krum is like Krum, but instead of picking only one client update, it picks a small group of the most "central" updates and averages them.

Constructors

MultiKrumFullModelAggregationStrategy(int, int, bool)

public MultiKrumFullModelAggregationStrategy(int byzantineClientCount = 1, int selectionCount = 0, bool useClientWeightsForAveraging = false)

Parameters

byzantineClientCount int

selectionCount int

useClientWeightsForAveraging bool

Methods

Aggregate(Dictionary<int, IFullModel<T, TInput, TOutput>>, Dictionary<int, double>)

Aggregates model updates from multiple clients into a single global model update.

public override IFullModel<T, TInput, TOutput> Aggregate(Dictionary<int, IFullModel<T, TInput, TOutput>> clientModels, Dictionary<int, double> clientWeights)

Parameters

clientModels Dictionary<int, IFullModel<T, TInput, TOutput>>

Dictionary mapping client IDs to their trained models.

clientWeights Dictionary<int, double>

Dictionary mapping client IDs to their aggregation weights (typically based on data size).

Returns

IFullModel<T, TInput, TOutput>

The aggregated global model.

Remarks

This method combines model updates from clients using the strategy's specific algorithm.

For Beginners: Aggregation is like combining multiple rough drafts of a document into one polished version that incorporates the best parts of each.

The aggregation process typically:

1. Takes model updates (weight changes) from each client
2. Considers the weight or importance of each client (based on data size, accuracy, etc.)
3. Combines these updates using the strategy's algorithm
4. Returns a single aggregated model that represents the collective improvement

For example with weighted averaging (FedAvg):

• Client 1 (1000 samples): model update A
• Client 2 (500 samples): model update B
• Client 3 (1500 samples): model update C
• Aggregated update = (1000A + 500B + 1500*C) / 3000

GetStrategyName()

Gets the name of the aggregation strategy.

public override string GetStrategyName()

Returns

string

A string describing the aggregation strategy (e.g., "FedAvg", "FedProx", "Krum").

Remarks

For Beginners: This helps identify which aggregation method is being used, useful for logging, debugging, and comparing different strategies.