Obliviate

Implementation of Obliviate: Neutralizing Task-agnostic Backdoors within the Parameter-efficient Fine-tuning Paradigm, NAACL 2025 (Findings)

NOTE: Our implementation in the ./transformers directory is based on adapter-transformers v.3.2.1 (https://github.com/adapter-hub/adapter-transformers-legacy).

Requirements

First, install anaconda

Install python environments.

conda env create -f environments.yml -n obliviate
conda activate obliviate

Download backdoored models.

Download models from this link: https://github.com/obliviateARR/Obliviate/releases/download/model/model.tar.gz

Decompress the file.

tar -zxvf model.tar.gz

Run

Train and evalute PEFT models without defense

./run.py --model_dir model --model_name roberta-base --attack POR --peft adapter --task sst2 --lr 3e-4 --epoch 20

The evaluation results are saved in ./output/roberta-base/POR_adapter_eval/roberta-base_POR_sst2/eval_results.json

Train and evalute PEFT models with defense

./run.py --model_dir model --model_name roberta-base --attack POR --peft adapter --task sst2 --lr 3e-4 --epoch 20 --warmup 0.05 --defense --amp 3e-3 --reg 3e-2

The evaluation results are saved in ./output/roberta-base/POR_adapter_eval_defense/roberta-base_POR_sst2/eval_results.json