SplitModelOptions

Options & Configuration · AiDotNet.Models.Options

Configuration for split neural network architecture in vertical federated learning.

For Beginners

In vertical FL, a neural network is split into parts: each party runs a “bottom model” on its local features, producing an embedding (a compressed representation). The embeddings from all parties are combined and fed into a “top model” that produces the final prediction.

How It Works

This class controls how the network is split, how embeddings are sized, and how they’re combined.

Example:

Constructors

Constructor Summary
SplitModelOptions()

Properties

Property Summary
SplitPointStrategy SplitPoint { get; set; } Gets or sets the strategy for choosing where to split the network.
VflAggregationMode AggregationMode { get; set; } Gets or sets how embeddings from multiple parties are combined before the top model.
bool AddEmbeddingNoise { get; set; } Gets or sets whether to add gradient noise to embeddings before sending them to the coordinator, reducing information leakage.
bool UseBatchNormalization { get; set; } Gets or sets whether to use batch normalization in the bottom models.
double EmbeddingNoiseScale { get; set; } Gets or sets the standard deviation of Gaussian noise added to embeddings when AddEmbeddingNoise is enabled.
int EmbeddingDimension { get; set; } Gets or sets the output dimension of each party’s bottom model (embedding size).
int ManualSplitLayerIndex { get; set; } Gets or sets the manual split layer index (used only when SplitPoint is Manual).
int TopModelHiddenDimension { get; set; } Gets or sets the hidden dimension of the top model.
int TopModelHiddenLayers { get; set; } Gets or sets the number of hidden layers in the top model.