Small Data, Big Noise: Adversarial Training for Robust Parameter-Efficient Fine-Tuning

This repository contains the code for the paper "Small Data, Big Noise: Adversarial Training for Robust Parameter-Efficient Fine-Tuning".

We propose SDBN — an adversarial training framework that makes PEFT methods (LoRA, QLoRA, BitFit, Adapter, full fine-tuning) robust to input noise when trained on small datasets. Two pipelines are provided:

Pipeline	Method	Models	Datasets
Classification	SDBN / SDBN-H	BERT, DeBERTa, mBERT	Banking77, TREC, IMDB, NLI, News, BLESS, ArSarcasm-v2
Generative	SDBN-P	LLaMA-3.2-1B, LLaMA-2-7B, Qwen2.5-7B	SQuAD, TweetQA

---

Repository Structure

new_repo/
├── classification/
│   ├── train.py          # Adversarial PEFT training (SDBN / SDBN-H)
│   └── eval.py           # Robustness evaluation under noise
├── generative/
│   ├── train.py          # Adversarial PEFT training with pre-loaded variants (SDBN-P)
│   └── eval.py           # Robustness evaluation under noise
├── utils/
│   ├── utils.py          # Shared utilities for classification
│   └── utils_gen.py      # Shared utilities for generative
├── requirements.txt
└── run.sh                # Helper to invoke the conda environment's Python

---

Setup

1. Create environment

conda create -n sdbn python=3.9
conda activate sdbn
pip install -r requirements.txt

Note: adapters is a development install. Install it from source if the pip version is unavailable:

pip install git+https://github.com/adapter-hub/adapters.git

2. Using run.sh

run.sh is a convenience wrapper that invokes the correct Python environment without requiring conda activate:

chmod +x run.sh
./run.sh classification/train.py [args...]
./run.sh generative/train.py [args...]

---

Quick Start

Scripts must be run from the data root directory (where dataset folders such as banking77/ reside).

Classification

# Train (SDBN, LoRA, DeBERTa, Banking77)
./run.sh classification/train.py \
  --dataset banking77 --model deberta --alg lora --rank 4 \
  --epochs 10 --warm_up 3 --train_set_size 0.1 --pertubation sdbn

# Evaluate
./run.sh classification/eval.py \
  --dataset banking77 --model deberta --baseline lora --rank 4 \
  --epochs 10 --percent 0.1 --pertubation sdbn --noise_type delete_char

Generative

Pre-generated adversarial CSV files are required (see generative/README.md).

# Train (SDBN-P, LoRA, LLaMA-3.2-1B, SQuAD)
./run.sh generative/train.py \
  --dataset squad --model_name llama --lora_rank 4 \
  --train_size 200 --epochs 5 --warmup_epochs 1 \
  --resource_dir /path/to/sdbn-p_data/squad/trainset_200

# Evaluate
./run.sh generative/eval.py \
  --dataset squad --model llama --lora_rank 4 \
  --train_size 200 --noise_type delete_word

---

Perturbation Methods

Name	Description	Pipeline
sdbn	Embedding-space L∞ FGSM adversarial perturbation	Classification
sdbn-h	Hybrid: SDBN + character-level noise	Classification
sdbn-p	Pre-loaded adversarial paraphrases (5 variants per example)	Generative

---

Citation

If you use this code, please cite our paper:

@article{sdbn2026,
  title={Small Data, Big Noise: Adversarial Training for Robust Parameter-Efficient Fine-Tuning},
  author={Eitan Cohen, Idan Simai, and Uri Shaham},
  year={2026}
}