Rethinking RobustBench: Is High Synthetic-Test Data Similarity an Implicit Information Advantage Inflating Robustness Scores?

Overview

This repository contains the code for the paper "Rethinking RobustBench: Is High Synthetic-Test Data Similarity an Implicit Information Advantage Inflating Robustness Scores?".

Data

Synthetic Data (20M)

To replicate the feature extraction and similarity calculation, you need the 20 million synthetic CIFAR-10 images used in the study. These can be downloaded from the links provided in the DM-Improves-AT repository:

Dataset	Size	Download Links
CIFAR-10	20M	part1 part2

Download both 20m_part1.npz and 20m_part2.npz and place them in your data directory.

Pre-computed Top-20 Similarity Data (20m_top20.npz)

As an alternative to running the similarity calculation yourself, we provide the pre-computed file containing the top 20 most similar synthetic images (from the 20M set) for each CIFAR-10 test image, based on feature similarity.

Download Links:

• Google Drive: https://drive.google.com/file/d/1dPktooDpNOmnuiGaKQ8olBd-yFGkeIPO/view?usp=sharing
• Baidu Netdisk: https://pan.baidu.com/s/1PYMsULBaF0I-aSPgCsIpZA?pwd=yx5u (Password: yx5u)

File Contents:

This .npz file contains NumPy arrays accessible via the following keys:

• image: The actual pixel data of the top-k similar synthetic images. Shape: (num_test_images * k, height, width, channels)
• label: The corresponding labels of these synthetic images. Shape: (num_test_images * k,)
• test_image_id: The index (0-9999) of the CIFAR-10 test image to which each synthetic image is most similar. Shape: (num_test_images * k,)
• rank: The similarity rank (0 to k-1, where 0 is the most similar). Shape: (num_test_images * k,)
• similarity: The calculated cosine similarity score. Shape: (num_test_images * k,)

In our case, k=20 and num_test_images=10000.

Code Usage

The following scripts are provided to perform the feature extraction and similarity analysis:

1. extract_features.py: Extracts deep features from the 20M synthetic dataset (20m_part1.npz, 20m_part2.npz).
   • Usage: Modify paths within the script and run python extract_features.py.
2. extract_cifar_test_features.py: Extracts deep features from the standard CIFAR-10 test set.
   • Usage: Modify paths within the script and run python extract_cifar_test_features.py.
3. calc_similarity.py: Calculates the cosine similarity between the features extracted in the previous steps. It identifies the most similar synthetic images for each test image.
   • Usage: Ensure the feature files from steps 1 & 2 are available. Modify paths within the script and run python calc_similarity.py.
4. save_data.py: Saves the top-k (e.g., top 20) most similar synthetic images, their labels, corresponding test image IDs, ranks, and similarity scores into the 20m_top20.npz format described above.
   • Usage: Ensure the similarity results from step 3 are available. Modify paths and parameters (like k=20) within the script and run python save_data.py.

Note: You can skip running these scripts if you download the pre-computed 20m_top20.npz file.

Training Example (Reproducing Paper Results)

To train the WideResNet-28-10 model using the pre-computed 20m_top20.npz data (as auxiliary data for TRADES), you can use the training script from the DM-Improves-AT repository.

Run the following command, ensuring you replace './20m_top20.npz' with the actual path to your downloaded or generated file:

python train-wa.py --data-dir 'dataset-data' \
    --log-dir 'trained_models' \
    --desc 'WRN28-10Swish_cifar10s_lr0p2_TRADES5_epoch400_bs512_fraction0p8_ls0p1' \
    --data cifar10s \
    --batch-size 512 \
    --model wrn-28-10-swish \
    --num-adv-epochs 400 \
    --lr 0.2 \
    --beta 5.0 \
    --unsup-fraction 0.8 \
    --aux-data-filename './20m_top20.npz' \
    --ls 0.1

Pre-trained Models

Our pre-trained models are available for download from the following locations:

• Google Drive: https://drive.google.com/drive/folders/1UI7GDs0-EvqpEFOO1qEIIw44FhQk1gmR?usp=sharing
• Baidu Netdisk: https://pan.baidu.com/s/1T9G3ei44hslzRpcJM4kD5A?pwd=1s9q (Password: 1s9q)

Evaluation

To evaluate the downloaded models using AutoAttack, you can use the eval-last-aa.py script from the DM-Improves-AT repository.

First, ensure you have placed the downloaded model files (e.g., the folder WRN28-10Swish_cifar10s_lr0p2_TRADES5_epoch400_bs512_fraction0p8_ls0p1) inside the directory specified by --log-dir (default is trained_models).

Then, run the following command:

python eval-last-aa.py --data-dir 'dataset-data' \
    --log-dir 'trained_models' \
    --desc 'WRN28-10Swish_cifar10s_lr0p2_TRADES5_epoch400_bs512_fraction0p8_ls0p1'