Strategy Selection Guide

Erasus provides 28 unlearning strategies across six categories. This guide helps you choose the right one.

Strategy Categories

Gradient Methods — Modify gradients to reverse learning:

• gradient_ascent — Simple, fast; maximises loss on forget data

• negative_gradient — Applies negative gradients from forget data

• saliency_unlearning — Uses gradient saliency to target important parameters

Parameter Methods — Directly modify model weights:

• fisher_forgetting — Uses Fisher Information to identify and dampen relevant parameters

• layer_freezing — Freezes unrelated layers, unlearns only relevant ones

Data Methods — Leverage data relationships:

• scrub — Two-phase: forget step + retain barrier (SCRUB paper)

• knowledge_distillation — Teacher-student: retain teacher’s knowledge, forget specific data

VLM-Specific — Vision-language models:

• contrastive_unlearning — Increases distance between forget image-text pairs

• attention_unlearning — Modifies cross-attention weights

• vision_text_split — Separate encoder updates

LLM-Specific — Language models:

• attention_surgery — Direct attention weight modification

Diffusion-Specific — Generative models:

• concept_erasure — Closed-form concept removal

• timestep_masking — Selective timestep training

• safe_latents — Safe Latent Diffusion constraint

Decision Matrix

Strategy	Speed	Forgetting	Utility	Privacy	Modality
gradient_ascent	★★★	★★	★★	★	All
scrub	★★	★★★	★★★	★★	All
fisher_forgetting	★★	★★★	★★★	★★	All
concept_erasure	★★★	★★★	★★★	★★	Diffusion
contrastive_unlearning	★★	★★★	★★★	★★	VLM

Combining Strategies

Use EnsembleStrategy to combine multiple strategies:

from erasus.strategies.ensemble_strategy import EnsembleStrategy

ensemble = EnsembleStrategy(
    strategies=["gradient_ascent", "fisher_forgetting"],
    weights=[0.7, 0.3],
)