Unveiling potential threats: backdoor attacks in single-cell pretrained models

---

🙋 Please let us know if you find out a mistake or have any suggestions!

🌟 If you find this resource helpful, please consider to star this repository and cite our research:

Sicheng Feng, Siyu Li, Luonan Chen, Shengquan Chen. Unveiling potential threats: backdoor attacks in single-cell pretrained models. 2024.

Requirements and Installation

We use python 3.9 from Anaconda. We provide two conda environments for the experiments: base.yml and geneformer.yml. The base.yml is for the scGPT and scBERT experiments, while the geneformer.yml is for the GeneFormer experiments.

To install all dependencies:

conda env create -f base.yml

# or
conda env create -f geneformer.yml

Datasets

1. Example datasets from [scGPT]
2. Example datasets from [GeneFormer]
3. Datasets from [Tabula Sapiens Single-Cell Dataset]

Place the downloaded contents under Yourpath4Dataset to reproduce the experiments.

Pretrained Models

You can download the pretrained models from [scGPT] (whole-human), [scBERT] and [GeneFormer], then place the downloaded contents under Yourpath4PretrainedModels to reproduce the experiments.

Quick Demos

1. Download datasets and pretrained models, then place them under rightpath and adjust the path-params in the scripts.
2. Then you can try to reproduce the experiments with the provided scripts. For example, you can evaluate on Human Pancreas datasets by:

nohup ./run.sh & # for scGPT_Exp

Details of Experiments

The commands to run the experiments are as follows:

nohup ./run.sh & # for scGPT_Exp
nohup ./run.sh & # for scBERT_Exp
nohup ./run.sh & # for GeneFormer_Exp
...

# or you can run the experiments in tmux or screen
./run_diff_batch.sh # for scGPT_Exp
./run_diff_feature.sh # for scGPT_Exp
...

The poison-related code is in the poison_utils.py or poison_trigger.py. You can find them in each experiment's folder.

The folder tree is as follows:

├── LICENSE
├── README.md                             -- introduction about the project
├── figures                               -- use for show up
│   └── fig1.png
├── requirements.txt                      -- requirements for installation
│── scGPT_Exp                             
│   ├── test                              -- the attack pipeline
│   │   ├── run.sh
│   │   ├── run_diff_batch.sh             -- explore the impact of batch effects
│   │   ├── run_diff_feature.sh           -- explore the impact of feature selection
│   │   ├── run_3datasets.sh              
│   │   └── scBackdoor.py
│   └── utils                             -- the scGPT items
│       ├── detect_tools.py
│       ├── poison_trigger.py
│       ├── preprocess.py
│       ├── print_tools.py
│       └── tools.py
├── GeneFormer_Exp 
│   ├── geneformer                        -- the GeneFormer items
│   │   ├── __init__.py
│   │   ├── classifier.py
│   │   ├── classifier_utils.py
│   │   ├── collator_for_classification.py
│   │   ├── emb_extractor.py
│   │   ├── evaluation_utils.py
│   │   ├── gene_median_dictionary.pkl
│   │   ├── gene_name_id_dict.pkl
│   │   ├── in_silico_perturber.py
│   │   ├── in_silico_perturber_stats.py
│   │   ├── perturber_utils.py
│   │   ├── poison_utils.py
│   │   ├── pretrainer.py
│   │   ├── token_dictionary.pkl
│   │   └── tokenizer.py
│   ├── run.sh                            -- the attack pipeline
│   └── geneformer_scBackdoor.py          
└── scBERT_Exp
    ├── attn_sum_save.py
    ├── finetune.py
    ├── lr_baseline_crossorgan.py
    ├── performer_pytorch                 -- the scBERT items
    │   ├── __init__.py
    │   ├── performer_pytorch.py
    │   └── reversible.py
    ├── poison_utils.py
    ├── predict.py
    ├── preprocess.py
    ├── pretrain.py
    ├── run.sh                            -- the attack pipeline
    ├── run_3datasets.sh
    └── utils.py

Further Reading

1. scGPT: Towards Building a Foundation Model for Single-Cell Multi-omics Using Generative AI, Nature Methods 2024. [GitHub Repo]

2. Transfer learning enables predictions in network biology, Nature 2023. [Huggingface Repo]

3. scBERT as a Large-scale Pretrained Deep Language Model for Cell Type Annotation of Single-cell RNA-seq Data, Nature Machine Intelligence 2022. [GitHub Repo]

Acknowledgement

We sincerely thank the authors of the following open-source projects:

• scGPT
• GeneFormer
• scBERT
• scanpy
• datasets
• torch