BobaAggregationStrategy<T>

Models & Types · AiDotNet.FederatedLearning.Aggregators

Implements BOBA (Bayesian Optimal Byzantine-robust Aggregation) strategy.

Inheritance: FederatedLearningComponentBase<T> → AggregationStrategyBase<Dictionary<string, T[]>, T> → ParameterDictionaryAggregationStrategyBase<T> → BobaAggregationStrategy<T>

For Beginners

Most Byzantine defenses use fixed rules (e.g., remove outliers). BOBA takes a probabilistic approach using a two-class mixture model — it models client updates as coming from either an “honest” distribution (tight cluster) or a “Byzantine” distribution (diffuse/adversarial). It uses Expectation-Maximization (EM) to estimate which class each client belongs to, then aggregates using only the honest-class posterior probabilities.

How It Works

Two-class mixture model:

EM algorithm per round:

Cross-round belief propagation: posteriors from round t become priors for round t+1, so persistent attackers accumulate low trust over multiple rounds.

Reference: BOBA: Bayesian Optimal Byzantine-robust Aggregation (2024). https://arxiv.org/abs/2312.09672

Constructors

Constructor Summary
BobaAggregationStrategy(double priorHonest, int emIterations, double emConvergenceTol, double minResponsibility) Initializes a new instance of the BobaAggregationStrategy class.

Properties

Property Summary
double EmConvergenceTol { get; } Gets the EM convergence tolerance.
double PriorHonest { get; } Gets the prior probability of a client being honest.
int EmIterations { get; } Gets the maximum number of EM iterations per round.

Methods

Method Summary
Dictionary<string, T[]> Aggregate(Dictionary<int, Dictionary<string, T[]>> clientModels, Dictionary<int, double> clientWeights)
string GetStrategyName()