🚀 Toward Inherently Robust VLMs Against Visual Perception Attacks

This repository contains shell scripts and Python configurations designed for fine-tuning and performing inference for the following vision-language models:

• LLaVA-13B-LoRA
• LLaVA-7B
• MoE-LLaVA
• MobileVLM
• Qwen-VL
• NVILA

This work was Accepted at the 2026 IEEE Intelligent Vehicles Symposium (IV 2026).

---

🛠️ Installation and Setup

Vision Language Model Fine-Tuning and Inference

1. Clone the repository:

   git clone MODEL-REPO

2. Install required dependencies:

   pip install -r requirements.txt

3. Set up DeepSpeed by following their official installation guide: DeepSpeed Documentation.

---

📝 Scripts Overview

1️⃣ LLaVA 13B Fine-Tuning (LLaVA_13B_FINETUNE.sh)

This script fine-tunes the LLaVA 13B model using DeepSpeed.

Arguments Overview:

• --model_name_or_path: Path to the pre-trained LLaVA 13B model.
• --image_folder: Directory containing the training images.
• --data_path: Directory containing the dataset in JSON format.
• --output_dir: Directory where the fine-tuned model checkpoints will be saved.

Command Example:

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=$WORKSPACE_DIR python3 $WORKSPACE_DIR/llava/train/train_mem.py \
    --deepspeed ./scripts/zero3.json \
    --model_name_or_path ./checkpoints/llava-v1.5-13B/ \
    --image_folder IMAGE_DIRECTORY \
    --data_path JSON_FILE_DIRECTORY \
    --output_dir OUTPUT_FINE_TUNED

---

2️⃣ LLaVA 7B Fine-Tuning (LLaVA_7B_FINETUNE.sh)

This script fine-tunes the LLaVA 7B model using DeepSpeed.

Arguments Overview:

• --model_name_or_path: Path to the pre-trained LLaVA 7B model.
• --image_folder: Directory containing the training images.
• --data_path: Directory containing the dataset in JSON format.
• --output_dir: Directory where the fine-tuned model checkpoints will be saved.

Command Example:

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=$WORKSPACE_DIR python3 $WORKSPACE_DIR/llava/train/train_mem.py \
    --deepspeed ./scripts/zero2.json \
    --model_name_or_path ./checkpoints/llava-v1.5-7B/ \
    --image_folder IMAGE_DIRECTORY \
    --data_path JSON_FILE_DIRECTORY \
    --output_dir OUTPUT_FINE_TUNED

---

3️⃣ MoE-LLaVA Fine-Tuning (MoE_LLaVA_FINETUNE.sh)

This script fine-tunes the MoE-LLaVA model using DeepSpeed with the Mixture of Experts (MoE) method.

Arguments Overview:

• --model_name_or_path: Path to the fine-tuned LLaVA model.
• --image_folder: Directory containing the training images.
• --data_path: Directory containing the dataset in JSON format.
• --output_dir: Directory where the fine-tuned model checkpoints will be saved.

Command Example:

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=$WORKSPACE_DIR python3 $WORKSPACE_DIR/moellava/train/train_mem.py \
    --moe_enable True \
    --model_name_or_path ./checkpoints/MoE-v1.5-7B/ \
    --image_folder IMAGE_DIRECTORY \
    --data_path JSON_FILE_DIRECTORY \
    --output_dir OUTPUT_FINE_TUNED

---

4️⃣ MobileVLM Fine-Tuning (MobileVLM_FINETUNE.sh)

This script fine-tunes the MobileVLM model using DeepSpeed.

Arguments Overview:

• --model_name_or_path: Path to the pre-trained MobileVLM model.
• --image_folder: Directory containing the training images.
• --data_path: Directory containing the dataset in JSON format.
• --output_dir: Directory where the fine-tuned model checkpoints will be saved.

---

5️⃣ Qwen-VL Fine-Tuning (Qwen-VL.sh)

This script fine-tunes the Qwen-VL model and evaluates it on a test set.

Arguments Overview:

• --model_name_or_path: Path to the pre-trained Qwen-VL model.
• --data_path: Directory containing the dataset in JSON format.
• --output_dir: Directory where the fine-tuned model checkpoints will be saved.
• --num_train_epochs: Number of training epochs.
• --learning_rate: Learning rate for optimization.
• --save_steps: Frequency of checkpoint saving.
• --evaluation_strategy: Strategy for model evaluation.

Command Example:

CUDA_VISIBLE_DEVICES=0 torchrun --nproc_per_node 2 --nnodes 1 --node_rank 0 $WORKSPACE_DIR/finetune.py \
    --model_name_or_path ./checkpoints/Qwen/Qwen-VL-Chat \
    --data_path ./data/MY_DATASET/train.json \
    --output_dir ./checkpoints/Qwen-VL-finetuned \
    --num_train_epochs 1 \
    --learning_rate 1e-5 \
    --save_steps 1000 \
    --evaluation_strategy "no" \
    --logging_steps 1 \
    --deepspeed ./finetune/ds_config_zero3.json

6️⃣ NVILA-Lite-8B Fine-Tuning and Evaluation (NVILA.sh)

This script performs end-to-end fine-tuning and evaluation of the NVILA-Lite-8B model using DeepSpeed and vila-infer. It supports both training and inference in one unified workflow.

Arguments Overview:

• STAGE_PATH: Path to the pre-trained NVILA-Lite-8B model (default: Efficient-Large-Model/NVILA-Lite-8B).
• DATA_MIXTURE: Name of the training dataset or mixture.
• OUTPUT_DIR: Directory where the fine-tuned model and logs will be saved.

Command Example:

---

## 📚 File Structure
├── LLaVA-13B-LoRA
│   ├── LICENSE
│   ├── llava
│   ├── LLaVA-13-LoRA.sh
│   └── scripts
├── LLaVA-7B
│   ├── LICENSE
│   ├── llava
│   ├── LLaVA-7B.sh
│   └── scripts
├── MobileVLM
│   ├── LICENSE
│   ├── mobilevlm
│   ├── MobileVLM.sh
│   └── scripts
├── MoE-LLaVA
│   ├── LICENSE
│   ├── moellava
│   ├── MoE-LLaVA.sh
│   └── scripts
├── Qwen-VL
│   ├── LICENSE
│   ├── Qwen-VL.sh
│   └── finetune
├── NVILA
│   ├── LICENSE
│   ├── NVILA.sh
│   ├── scripts
│   └── llava
├── Sample
│   ├── DRP-Attack
│   ├── RAUCA
│   └── Shadow-Attack


---

📊 Training and Evaluation Metrics

Training scripts log progress via TensorBoard and W&B for visualization and debugging purposes. Modify logging steps and evaluation strategies as needed.

---

🔧 Customization

• Batch size: 32 (training), 4 (evaluation)
• Checkpoints saved every 50,000 steps
• DeepSpeed configurations adjustable in zero2.json or zero3.json

---

By following this guide, you can efficiently fine-tune and infer using the LLaVA, MoE-LLaVA, MobileVLM, and Qwen-VL models.

Note: Ensure you have access to GPUs with adequate memory for fine-tuning large models.

Note: Ensure that you have access to GPUs with adequate memory for fine-tuning large models.
Note: The models are fine-tuned on an A100 40GB GPU, except for Qwen-VL (2×A100 80GB GPUs) and NVILA (4×A100 40GB GPUs).