DivideAndConquerAggregationStrategy<T>

Models & Types · AiDotNet.FederatedLearning.Aggregators

Implements DnC (Divide and Conquer) aggregation strategy for Byzantine-robust FL.

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

For Beginners

Some poisoning attacks are hard to detect when you look at the full high-dimensional update vectors — the malicious signal hides in the noise. DnC looks at random subsets of the update coordinates, then uses spectral analysis (top singular vector projection) to identify attackers that might evade simpler coordinate-wise defenses like median or trimmed mean.

How It Works

Algorithm:

Reference: Shejwalkar, V. & Houmansadr, A. (2021). “Manipulating the Byzantine: Optimizing Model Poisoning Attacks and Defenses for Federated Learning.” NDSS 2021, Algorithm 2.

Constructors

Constructor Summary
DivideAndConquerAggregationStrategy(int numByzantine, int subspaceDimension, int seed, int iterations, double filterFraction) Initializes a new instance of the DivideAndConquerAggregationStrategy class.

Properties

Property Summary
double FilterFraction { get; } Gets the filtering fraction c.
int Iterations { get; } Gets the number of filtering iterations (niters).
int NumByzantine { get; } Gets the expected number of Byzantine clients.
int SubspaceDimension { get; } Gets the number b of sub-sampled coordinates per iteration.

Methods

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