BackdoorDefenseOptions
Options & Configuration · AiDotNet.Models.Options
Configuration options for backdoor attack detection and mitigation in federated learning.
Inheritance: ModelOptions → BackdoorDefenseOptions
For Beginners
Backdoor attacks are stealthy — the model works correctly on clean data but misbehaves when a specific trigger pattern is present. Unlike Byzantine attacks (which degrade overall performance), backdoors are targeted and hard to detect statistically. These detectors analyze client updates for tell-tale signs of backdoor injection.
Constructors
| Constructor | Summary |
|---|---|
BackdoorDefenseOptions() |
Properties
| Property | Summary |
|---|---|
BackdoorDetectionStrategy Strategy { get; set; } |
Gets or sets the detection strategy. |
double AlignmentThreshold { get; set; } |
Gets or sets the cosine similarity threshold below which a subspace is suspicious. |
double AnomalyThreshold { get; set; } |
Gets or sets the MAD-based anomaly threshold for Neural Cleanse. |
double SuspicionThreshold { get; set; } |
Gets or sets the suspicion threshold above which a client is filtered out. |
int NumClasses { get; set; } |
Gets or sets the number of output classes for Neural Cleanse trigger analysis. |
int NumSubspaces { get; set; } |
Gets or sets the number of parameter subspaces for direction alignment analysis. |