RobustAggregationOptions
Options & Configuration · AiDotNet.Models.Options
Configuration options for robust aggregation strategies in federated learning.
Inheritance: ModelOptions → RobustAggregationOptions
How It Works
Robust aggregation helps defend against outliers and Byzantine (malicious or faulty) clients.
For Beginners: If some clients send “bad” updates (because of bugs, corrupted data, or attacks), robust aggregation tries to reduce their impact so the global model stays stable.
Common strategies:
- Trimmed mean: drops extreme values before averaging.
- Median: takes the middle value per parameter (resistant to outliers).
- Krum/Multi-Krum: chooses the most “central” client updates by distance.
- Bulyan: combines Multi-Krum selection with trimming for stronger robustness.
Constructors
| Constructor | Summary |
|---|---|
RobustAggregationOptions() |
Properties
| Property | Summary |
|---|---|
bool UseClientWeightsWhenAveragingSelectedUpdates { get; set; } |
Gets or sets whether robust strategies should use clientWeights when averaging selected updates. |
double GeometricMedianEpsilon { get; set; } |
Gets or sets a small epsilon used to avoid division-by-zero in RFA (Weiszfeld) updates. |
double GeometricMedianTolerance { get; set; } |
Gets or sets the convergence tolerance for geometric-median/RFA aggregation. |
double TrimFraction { get; set; } |
Gets or sets the trimming fraction for trimmed-mean based aggregations (0.0 to < 0.5). |
int ByzantineClientCount { get; set; } |
Gets or sets the assumed number of Byzantine clients (f) for Krum/Multi-Krum/Bulyan. |
int GeometricMedianMaxIterations { get; set; } |
Gets or sets the maximum number of iterations for geometric-median/RFA aggregation. |
int MultiKrumSelectionCount { get; set; } |
Gets or sets how many client updates (m) Multi-Krum keeps before averaging. |