Dataset Card for the LoWRA Bench Dataset

The LoRA Weight Recovery Attack (LoWRA) Bench is a comprehensive benchmark designed to evaluate Pre-Fine-Tuning (Pre-FT) weight recovery methods as presented in the "Recovering the Pre-Fine-Tuning Weights of Generative Models" paper.

Task Details
Dataset Description
Dataset Structure
Dataset Creation
Risks and Out-of-Scope Use
Considerations for Using the Data
- Licensing Information
- Citation Information
🌐 Homepage: https://vision.huji.ac.il/spectral_detuning/
🧑‍💻 Repository: https://github.com/eliahuhorwitz/Spectral-DeTuning
📃 Paper: https://arxiv.org/abs/2402.10208
✉️ Point of Contact: eliahu.horwitz@mail.huji.ac.il

Task Details

Pre-Fine-Tuning Weight Recovery Attack Setting: We uncover a vulnerability in LoRA fine-tuned models wherein an attacker is able to undo the fine-tuning process and recover the weights of the original pre-trained model. The setting for the vulnerability is as follows:

(a) The attacker only has access to n different LoRA fine-tuned models.

(b) The attacker assumes that all n models originated from the same source model.

Note: The attacker has no access to the low-rank decomposition of the fine-tuned models.

Dataset Description

The LoWRA Bench dataset is designed to evaluate the performance of Pre-FT weight recovery methods. The dataset encompasses three pre-trained representative source models:

A Vision Transformer (ViT) pre-trained on ImageNet-1K.
Mistral-7B-v0.1.
Stable Diffusion 1.5.

These models collectively cover supervised and self-supervised objectives, spanning both vision and natural language processing (NLP) domains, as well as generative and discriminative tasks. Notably, these models are widely used and deployed in numerous production systems.

For each source model, we curate 15 LoRA models fine-tuned on diverse datasets, tasks, and objectives. The dataset comprises a diverse array of layer types, including self-attention, cross-attention, and MLPs. This diversity enables us to assess the generalization capabilities of Pre-FT methods. The evaluation can be conducted on a per-model basis, per layer type, or layer depth, allowing for a comprehensive analysis of Pre-FT methods. Overall, our dataset includes 544 source model layers. When taking into account the fine-tuned LoRA layers, the dataset includes over 8,000 layers.

Dataset Structure

The dataset contains 4 subsets, for each subset we curate 15 LoRA fine-tuned models. Each row of the dataset represents a single layer that should be recovered and contains all the needed information for the recovery and numerical evaluation. In particular, for each layer, the dataset includes the original Pre-FT weights and the unmerged fine-tuned LoRA weight matrices. We decided to provide the unmerged weights instead of the merged ones for two reasons:

Providing the unmerged weights significantly reduces the storage size of the dataset (e.g., for a single Mistral subset this reduces the size from ~100GB to ~8GB).
Providing the unmerged weights allows the dataset user to study the properties of the fine-tuned LoRA layers and may help when developing new methods.

We leave the merging of the layers to the user, keep in mind this should be done carefully and tested to ensure the original Pre-FT weights are not simply provided to the method verbatim. See Layer Merging Example for an example taken from our GitHub repository.

Data Subsets

The table below describes the dataset subsets in detail:

Subset Name	Pre-FT Model	Task	Fine-tuning Task	# Pre-FT Layers	# Fine-tuned Layers
vit	ViT	Image Classification	VTAB-1K	24	360
stable-diffusion-1.5	Stable Diffusion 1.5	Text-to-Image Generation	Personalization	264	3960
mistral-7b-v0.1-sft	Mistral-7B-v0.1	Text Generation	UltraChat SFT	128	1920
mistral-7b-v0.1-dpo	Mistral-7B-v0.1	Text Generation	UltraFeedback DPO	128	1920

Data Fields

As described above, each row of the dataset represents a single layer that should be recovered and contains the following fields:

task_name - The name of the task the model was fine-tuned on (subset).
layer_model - In some cases a Pre-FT model has more than one model (e.g., Stable Diffusion fine-tuned both 
                the UNet and the Text Encoder). This field specifies the model the layer belongs to.
layer_name - The name of the layer in the Pre-FT model as it appears in the model state_dict.
pre_ft_name - The name of the Pre-FT model (e.g., runwayml/stable-diffusion-v1-5).
pre_ft_weight - The weight matrix of the Pre-FT models layer. 
lora_{lora_idx}_name - The name of the LoRA fine-tuned model.
lora_{lora_idx}_A_weight - The LoRA A weight matrix of the LoRA fine-tuned models layer.
lora_{lora_idx}_B_weight - The LoRA B weight matrix of the LoRA fine-tuned models layer.
lora_{lora_idx}_rank - The LoRA rank of the LoRA fine-tuned models layer.
lora_{lora_idx}_alpha - The LoRA alpha of the LoRA fine-tuned models layer.

where {lora_idx} is the index of the LoRA fine-tuned model in the subset (there are 15 LoRA models per subset).

Layer Merging Example

The following code snippet demonstrates merging the LoRA fine-tuned weights with the Pre-FT weights.

def merge_lora_weights(args, layer_idx, device):
    dataset = load_dataset(args.dataset, name=args.subset, cache_dir=args.cache_dir)
    layer = deepcopy(dataset.with_format("torch")["train"][layer_idx])

    merged_layer = {}

    # Note: load the ground truth Pre-FT weights
    merged_layer['layer_model'] = layer['layer_model']
    merged_layer['layer_name'] = layer['layer_name']
    merged_layer['pre_ft_name'] = layer['pre_ft_name']
    W_pre_ft = deepcopy(layer['pre_ft_weight']).to(device).float()
    merged_layer['pre_ft_weight'] = deepcopy(W_pre_ft)

    # Note: merge the LoRA weights for all existing LoRA models
    for lora_idx in args.lora_ids:
        alpha = layer[f'lora_{lora_idx}_alpha']
        rank = layer[f'lora_{lora_idx}_rank']
        B = deepcopy(layer[f'lora_{lora_idx}_B_weight']).to(device).float()
        A = deepcopy(layer[f'lora_{lora_idx}_A_weight']).to(device).float()

        merged_layer[f'lora_{lora_idx}_name'] = layer[f'lora_{lora_idx}_name']
        merged_layer[f'lora_{lora_idx}_rank'] = rank
        merged_layer[f'lora_{lora_idx}_alpha'] = alpha
        merged_layer[f'lora_{lora_idx}_merged_weights'] = W_pre_ft + ((alpha / rank * B) @ A)

        assert torch.allclose(merged_layer['pre_ft_weight'], layer['pre_ft_weight'])
        assert not torch.allclose(merged_layer[f'lora_{lora_idx}_merged_weights'], layer['pre_ft_weight'])
        assert not torch.allclose(merged_layer[f'lora_{lora_idx}_merged_weights'], merged_layer['pre_ft_weight'])
    return merged_layer

Dataset Creation

Source Data

The fine-tuning of the ViT models was performed using the PEFT library on various datasets from the VTAB-1K benchmark.
The fine-tuned LoRA models for Stable Diffusion are taken from civitai and were fine-tuned by RalFinger.
The fine-tuning of Mistral was performed based on the Zephyr model as seen here.

For the full list of models and hyper-parameters see the appendix of the paper.

Risks and Out-of-Scope Use

Our work uncovers a significant vulnerability in fine-tuned models, allowing attackers to access pre-fine-tuning weights. While this discovery reveals potential security risks, our primary objective is to advance the field of Machine Learning and raise awareness within the research community about the existing vulnerabilities in current models.

Instead of using the findings of this study to execute attacks, we advocate for their use by model creators to enhance the safety and security of their models. By acknowledging and addressing vulnerabilities, creators can proactively safeguard against potential threats.

Following established practices in the cyber-security community, we emphasize the importance of open discussion and encourage the reporting of vulnerabilities. By fostering transparency and collaboration, we can collectively create a safer environment for deploying machine learning models.

Considerations for Using the Data

Licensing Information

[More Information Needed]

Citation Information

If you use this dataset in your work please cite the following paper:

BibTeX:

@article{horwitz2024recovering,
  title={Recovering the Pre-Fine-Tuning Weights of Generative Models},
  author={Horwitz, Eliahu and Kahana, Jonathan and Hoshen, Yedid},
  journal={arXiv preprint arXiv:2402.10208},
  year={2024}
}