AdaLoc

AdaLoc enables secure, adaptable DNN usage control through a localized update using a small, adaptable key. It allows efficient updates to a small subset of parameters while maintaining strong performance and protecting against unauthorized access.

Features

• Vision Tasks: Supports datasets like CIFAR-10, CIFAR-100, MNIST, FashionMNIST, Flower102, and Caltech256. Includes models such as ResNet-18, ResNet-152, DenseNet121, and ConvNeXtV2. Allows neuron selection based on criteria like Top, Random, or Bottom.
• NLP Tasks: Supports datasets such as QNLI, SST-2, and TweetEval. Compatible with BERT, RoBERTa, and DeBERTa models, with automatic logging and result storage.

Vision Task

Environment Setup

1. Clone the repository:

   git clone https://github.com/MLresearchAI
   

2. Install required packages:

   pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
   pip install -e .
   pip install -r requirements.txt
   

Usage

To train a model, run the following:

python tune_v.py --dataset cifar100 --model_name densenet121

Arguments

• --device: Training device (cuda:0 or cpu).
• --dataset: Dataset for training (cifar100, cifar10, mnist, fashionmnist, flower102, caltech256).
• --batch_size: Default 64.
• --num_workers: Default 6.
• --num_epochs: Default 50.
• --lr: Default 0.001.
• --momentum: Default 0.9.
• --model_name: Model architecture (resnet18, resnet152, densenet121, convnextv2).
• --pretrain: Use pretrained model weights (True by default).
• --num_classes: Number of output classes (100 by default).
• --retained: Fraction of neurons to retain (0.05 by default).
• --mode: Neuron selection mode (0 for Top, 1 for Random, 2 for Bottom).

Note: For the Caltech256 dataset, download it from Kaggle.

Example

To train a DenseNet121 model on the Caltech256 dataset:

python tune_v.py --dataset caltech256 --model_name densenet121 --pretrain --retained 0.05 --mode 0 --batch_size 64 --num_epochs 50 --lr 0.001 --momentum 0.9

NLP Task

Requirements

Ensure the following dependencies are installed:

pip install torch transformers datasets evaluate numpy pandas

Usage

Run the script as follows:

python tune_l.py --model_name <model> --output_path <path> --mode <mode>

Arguments

Argument	Type	Description
--model_name	string	Choose from bert-base-uncased, roberta-base, microsoft/deberta-v3-base
--output_path	string	Path to save the final results
--mode	string	Choose from smallest, largest, random

Example Command

python tune_l.py --model_name bert-base-uncased --output_path ./results --mode largest

Output

After running the script, the following outputs are generated:

1. A CSV file containing evaluation results (final_results.csv).
2. A logging directory with performance logs.
3. Printed evaluation metrics like accuracy and loss.

Example Output Format

The resulting CSV file (final_results.csv) contains:

dataset	p	accuracy	loss
qnli	0.05	-	-
sst2	0.1	-	-

---

Model Keying

Our adaptable key is designed to support and integrate seamlessly with a wide range of model keying methods. A demonstration implementation is available in the model_keying/ directory.

cd model_keying/

# To split the target model (primary key):
python locker.py split --model <path_to_model> --ratio <key_ratio> --arch <architecture> --save ./test_output

# To restore the model:
python locker.py recover --model test_output/keyed_model.pth.tar --key test_output/key.pth.tar --arch <architecture> --save <save_path>

Citation Information

If you find this code or dataset useful, please consider citing our paper:

@{coming soon}