Changelog

v0.1.0 (2026-02-14)

Initial release of the Erasus framework.

Core Framework

  • Base abstractions: BaseUnlearner, BaseStrategy, BaseSelector, BaseMetric

  • Registry system for pluggable components

  • YAML-based configuration with UnlearningConfig

Strategies (28)

  • Gradient methods: gradient ascent, negative gradient, saliency unlearning

  • Parameter methods: Fisher forgetting, layer freezing

  • Data methods: SCRUB, knowledge distillation

  • VLM-specific: contrastive unlearning, attention unlearning, vision-text split

  • LLM-specific: attention surgery

  • Diffusion-specific: concept erasure, timestep masking, safe latents

  • Ensemble strategy for combining multiple approaches

Selectors (22)

  • Influence-based, geometry-based, gradient-based, learning-based

  • Ensemble: stacking, voting, weighted fusion

  • Quality metrics for coreset evaluation

Metrics (26+)

  • Forgetting: accuracy, MIA, KL divergence, extraction attack

  • Utility: BLEU, ROUGE, CLIP score, inception score, downstream tasks

  • Privacy: epsilon-delta, privacy audit

  • Efficiency: time, memory, speedup, FLOPs

  • Benchmark runner with LaTeX export and radar plots

Model Wrappers (18+)

  • VLM: CLIP, LLaVA, BLIP, Flamingo, ViT utilities

  • LLM: GPT, Mistral, LLaMA, T5

  • Diffusion: Stable Diffusion, DALL-E, Imagen, diffusion utilities

  • Audio: Whisper, Wav2Vec, CLAP

  • Video: VideoMAE, VideoCLIP

Unlearners (8)

  • ErasusUnlearner (generic), VLM, LLM, Diffusion, Audio, Video

  • Multimodal auto-dispatcher, Federated unlearner

Privacy & Certification

  • Privacy accountant, DP mechanisms, gradient clipping, secure aggregation

  • Certified removal, verification, bounds (PAC, influence, radius)

Visualization (13)

  • Loss curves, feature plots, MIA plots, attention maps

  • Gradient analysis, surfaces, embeddings

  • Activation, influence maps, cross-modal, comparisons

Data (7 datasets + augmentation + synthetic)

  • TOFU, WMDP, COCO, I2P, Conceptual Captions, MUSE, ImageNet

  • Unlearning-aware augmentation

  • Synthetic: backdoor, bias, privacy generators

Infrastructure

  • CLI: unlearn, evaluate, benchmark, visualize commands

  • Experiment tracking (local/W&B/MLflow)

  • Hyperparameter search, ablation studies

  • Docker, CI/CD, comprehensive documentation

  • 253 unit tests passing