Memory Self-Regeneration: Uncovering Hidden Knowledge in Unlearned Models

Agnieszka Polowczyk*, Alicja Polowczyk*, Joanna Waczyńska, Piotr Borycki, Przemysław Spurek (*equal contribution)

Abstract: The impressive capability of modern text-to-image models to generate realistic visuals has come with a serious drawback: they can be misused to create harmful, deceptive or unlawful content. This has accelerated the push for machine unlearning. This new field seeks to selectively remove specific knowledge from a model's training data without causing a drop in its overall performance. However, it turns out that actually forgetting a given concept is an extremely difficult task. Models exposed to attacks using adversarial prompts show the ability to generate so-called unlearned concepts, which can be not only harmful but also illegal. In this paper, we present considerations regarding the ability of models to forget and recall knowledge, introducing the Memory Self-Regeneration task. Furthermore, we present MemoRa strategy, which we consider to be a regenerative approach supporting the effective recovery of previously lost knowledge. Moreover, we propose that robustness in knowledge retrieval is a crucial yet underexplored evaluation measure for developing more robust and effective unlearning techniques. Finally, we demonstrate that forgetting occurs in two distinct ways: short-term, where concepts can be quickly recalled, and long-term, where recovery is more challenging.

1. Environment Setup

1.1 Clone this repository (with submodule)

git clone --recursive https://github.com/gmum/MemoRa.git
cd MemoRa

1.2 Create and activate Conda environment

conda create -n memora python=3.13
conda activate memora

1.3 Install HuggingFace Diffusers

cd diffusers
pip install .
cd examples/text_to_image
pip install -r requirements.txt

1.4 Install PyTorch with CUDA 11.8

pip install torch==2.7.1 torchvision==0.22.1 torchaudio==2.7.1 --index-url https://download.pytorch.org/whl/cu118

1.5 Install project requirements

cd MemoRa
pip install -r requirements.txt

2. Download Checkpoints

2.1 Install gdown

pip install gdown

2.2 Download model checkpoints

We use pretrained checkpoints from:

• AdvUnlearn
• MACE

# AdvUnlearn ckpts
gdown --folder https://drive.google.com/drive/folders/1toiNWxJEX0X8pm8b88Og5_nMAqq6NZRd

# [Baseline ckpts] Nudity
gdown https://drive.google.com/uc?id=1wOqACzdpWpjRjl-a3hWgtjGAhemnelQr

# [Baseline ckpts] Style
gdown https://drive.google.com/uc?id=1-Gyl7Ls-Pa4vJzReG58bXjfI_YWu8teb

# [Baseline ckpts] Objects
gdown https://drive.google.com/uc?id=1OjYiFxYwd1B9R7vfKG6mooY27O2txYNg

3. MemoRa

If you want to run experiments for objects, simply change the config name.

Important: use the correct prefix (model_nudity@, model_parachute@, etc.)

3.1 Inversion

python inversion_images.py \
   model_nudity@unlearned=<MODEL> \
   model_nudity@original=sd \
   interpolation
   start_step=15 \
   --config-name=reunlearning_nudity

3.2 LoRa training

python train_lora.py \
   model_nudity@unlearned=<MODEL> \
   model_nudity@original=sd \
   interpolation.start_step=15 \
   lora.steps=500 \
   --config-name=reunlearning_nudity

3.3 Generate images

Baseline: Original Stable Diffusion

python generate_images.py \
   model_nudity@unlearned=<MODEL> \
   model_nudity@original=sd \
   model_nudity@evaluation=sd \
   interpolation.start_step=15 \
   generate_images.start_index=0 \
   generate_images.n_images=200 \
   generate_images.csv_file="nudity.csv" \
   --config-name=reunlearning_nudity

Unlearned model + MemoRa

python generate_images.py \
   model_nudity@unlearned=<MODEL> \
   model_nudity@original=sd \
   model_nudity@evaluation=<MODEL> \
   interpolation.start_step=15 \
   generate_images.start_index=0 \
   generate_images.n_images=200 \
   generate_images.csv_file="nudity.csv" \
   --config-name=reunlearning_nudity

Unlearned model without MemoRa

python generate_images.py \
   model_nudity@unlearned=<MODEL> \
   model_nudity@original=sd \
   model_nudity@evaluation=<MODEL> \
   generate_images.lora_scale=null \
   generate_images.lora_weights=null \
   interpolation.start_step=15 \
   generate_images.start_index=0 \
   generate_images.n_images=200 \
   generate_images.csv_file="nudity.csv" \
   --config-name=reunlearning_nudity

4. Evaluation with FID & CLIP

COCO-10k (for CLIP score and FID): you can extract the image subset from COCO dataset, or you can download it from here.

4.1 Install extra packages:

pip install git+https://github.com/openai/CLIP.git
pip install git+https://github.com/boomb0om/text2image-benchmark

4.2 Running FID & CLIP

python evaluate_fid.py \
   model_nudity@unlearned=<MODEL> \
   model_nudity@original=sd \
   model_nudity@evaluation=<MODEL> \
   interpolation.start_step=15 \
   generate_images.csv_file="coco_10k.csv" \
   --config-name=reunlearning_nudity

python evaluate_clip.py \
   model_nudity@unlearned=<MODEL> \
   model_nudity@original=sd \
   model_nudity@evaluation=<MODEL> \
   interpolation.start_step=15 \
   generate_images.csv_file="coco_10k.csv" \
   --config-name=reunlearning_nudity