momo1986 / patch-defense-leaderboard

A Leaderboard for Certifiable Robustness against Adversarial Patch Attacks

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Leaderboard: Certifiable Robustness against Adversarial Patch Attacks

This repository provides a leaderboard for certifiable robustness against adversarial patch attacks.

Check out our paper list and tutorials for adversarial patch attacks and defenses!

About

For now, this leaderboard focuses on the setting of

  1. Image classification task
  2. ImageNet dataset
  3. One 2% square patch anywhere on the image
  4. Prediction Recovery vs. attack detection (read this explanation for the difference between these two notions)

Stay tuned for future updates on different datasets and tasks :)

Annotated attributes

We include the following properties of different defenses in this leaderboard.

  1. Defense name and publication venue.
  2. Certified robust accuracy: certified robust accuracy for a 2%-pixel square patch anywhere on the image. All leaderboards are sorted by certified robust accuracy
  3. Clean accuracy: clean accuracy of the defended models
  4. Vanilla clean accuracy: clean accuracy of vanilla undefended models (if the defense is built upon off-the-shelf models)
  5. Backbone: the backbone used for each defense
  6. Comment: additional comments for each defense, including SOTA results, defense parameters, special training recipes.
  7. Code: most defenses have source code implemented by the authors or others

Note: We discuss more properties of different defenses in another copy of leaderboards hosted on Google Sheet.

Contributions

If you have new results and want to contribute to the leaderboard. Please submit a pull request or email me (Chong Xiang; cxiang@princeton.edu)

Table of Contents

Prediction Recovery

The prediction recovery defense requires the defense model to always predict/recover the robust label (without any abstention).

ImageNet

certified robust accuracy clean accuracy vanilla clean accuracy backbone comment code
MultiSizeGreedyCutout on PatchCleanser (TMLR 10/2023) 64.9 82.4 83.0 ViT-B SOTA at known 2% patch size; k=6; More improvements for 3%-pixel patch over cutout can be found in the paper NA
MultiSizeGreedyCutout on PatchCleanser (TMLR 10/2023) 63.1 82.0 82.8 ViT-B k=6 NA
PatchCleanser (USENIX Security 2022) 62.1 83.9 84.8 ViT-B k=6; trained with cutout code
PatchCleanser (USENIX Security 2022) 61.6 83.8 84.8 ViT-B k=6 code
PatchCleanser (USENIX Security 2022) 59.8 83.7 84.8 ViT-B k=5; trained with masked images code
PatchCleanser (USENIX Security 2022) 59.4 83.6 83.7 ViT-B k=4; trained with masked images code
PatchCleanser (USENIX Security 2022) 56.1 83.3 84.8 ViT-B k=3; trained with masked images code
PatchCleanser (USENIX Security 2022) 53.8 79.4 80.2 ResMLP-S k=6; trained with masked images code
PatchCleanser (USENIX Security 2022) 53.1 79.6 80.8 ResMLP-S k=6 code
PatchCleanser (USENIX Security 2022) 53 81.6 82.3 ResNet-50 k=6; trained with masked images code
ViP (ECCV 2022) 45.4 74.1 ViT-L (MAE) SOTA with unknown patch size (using a larger model MAE-Large); trained with MAE; b=19; stride = 5 code
ViP (ECCV 2022) 44.6 73.6 ViT-L (MAE) b=19; stride = 10 code
ECViT (CVPR 2022) 41.7 78.58 NA ViT-B SOTA with unknown patch size; b=37; progressively trained with pixel bands NA
PatchCleanser (USENIX Security 2022) 41.6 81.1 82.8 ResNet-50 k=6 code
ViP (ECCV 2022) 41.4 70.8 83.7 ViT-B (MAE) b=19; stride = 1 code
ECViT (CVPR 2022) 40.79 75.3 NA ViT-B b=25; progressively trained with pixel bands NA
ECViT (CVPR 2022) 40.72 73.49 NA ViT-B b=19; progressively trained with pixel bands NA
ViP (ECCV 2022) 40.6 70.4 83.7 ViT-B (MAE) b=19; stride = 5 code
ViP (ECCV 2022) 49.8 69.9 83.7 ViT-B (MAE) b=19; stride = 10 code
Smoothed ViT (CVPR 2022) 38.3 69.3 NA ViT-B SOTA with unknown patch size (with code); b=19 code
Smoothed ViT (CVPR 2022) 38.2 73.2 NA ViT-B b=37; trained with pixel bands code
Smoothed ViT (CVPR 2022) 36.9 68.3 NA ViT-B b=19; s=10; trained with pixel bands code
Smoothed ViT (CVPR 2022) 31.6 63.5 NA ViT-S b=19; trained with pixel bands code
ECViT (CVPR 2022) 30.06 67.14 NA ViT-S b=25; progressively trained with pixel bands NA
ECViT (CVPR 2022) 29.74 69.88 NA ViT-S b=37; progressively trained with pixel bands NA
ECViT (CVPR 2022) 28.85 64.69 NA ViT-S b=19; progressively trained with pixel bands NA
De-randomized Smoothing (NeurIPS 2020) (from the Smoothed ViT ) 28.1 61.4 NA WRN-101-2 b=19; trained with pixel bands code
PatchGuard (USENIX Security 2021) 26 54.6 56.5 BagNet-17 trained with masked features code
PatchGuard (USENIX Security 2021) 24.1 60.4 63 BagNet-33 trained with masked features code
BagCert (ICLR 2021) 22.7 45.3 NA BagNet-17 NA
BagCert (ICLR 2021) 22.4 47.3 NA BagNet-25 NA
BagCert (ICLR 2021) 20.1 47.9 NA BagNet-29 NA
De-randomized Smoothing (NeurIPS 2020) (from the Smoothed ViT ) 18.3 51.5 NA ResNet-50 b=19; trained with pixel bands code
PatchGuard (USENIX Security 2021) 15.7 43.6 44.4 ResNet-50 (de-randmoized smoothed code
Clipped BagNet (DLS 2020) 14.4 53.7 56.5 BagNet-17 code
De-randomized Smoothing (NeurIPS 2020) 14 44.4 NA ResNet-50 b=25; trained with pixel bands code
PatchGuard (USENIX Security 2021) 13.3 54.4 58.8 BagNet-17 code
Clipped BagNet (DLS 2020) 9.4 62.7 63 BagNet-33 code
Clipped BagNet (DLS 2020) 7.1 49.5 58.8 BagNet-17 code
PatchGuard (USENIX Security 2021) 6.9 61.2 66.6 BagNet-33 code
Clipped BagNet (DLS 2020) 1.9 60.3 66.6 BagNet-33 code
IBP (ICLR 2020) 0 0 0 NA not scale to high-resolution images code

(go back to table of contents)

Attack Detection

The attack detection defense allows the defense to alert and abstain from making predictions when it detects an attack.

Note 1:

  1. Attack detection defenses usually can tune the defense parameters to balance the trade-off between clean accuracy and robust accuracy.
  2. Here, I only report three representative points for each defense setup (different confidence threshold tau)
    • low-tau: high robust accuracy and low clean accuracy; clean accuracy and robust accuracy are the same
    • high-tau: low robust accuracy and high clean accuracy; clean accuracy is close to vanilla undefended clean accuracy
    • mid-tau: in between
  3. You are encouraged to play with the code yourself :)

Note 2:

  1. The original Minority Reports paper does not discuss high-resolution images like ImageNet.
  2. The results here (Minority Reports + PatchCleanser) are from the discussion section of the PatchCleanser paper.
  3. the idea is to plug PatchCleanser's mask generation approach into the MR defense design https://github.com/inspire-group/PatchCleanser/blob/main/misc/pc_mr.py

ImageNet

certified robust accuracy clean accuracy vanilla clean accuracy (if applicable) backbone comment code
ViP (ECCV 2022) 74.6 74.6 83.7 ViT-B (MAE) Highest certified robust accuracy; trained with MAE; low-tau code
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 74.3 74.3 84.8 ViT-B k=6; low-tau; trained with masked images code
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 73.7 73.7 84.8 ViT-B k=5; low-tau; trained with masked images code
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 73.2 73.2 84.8 ViT-B k=4; low-tau; trained with masked images code
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 72.8 72.8 84.5 ViT-B k=6; low-tau code
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 71.2 71.2 84.8 ViT-B k=3; low-tau; trained with masked images code
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 68.3 68.3 82.3 ResNet-50 low-tau; trained with masked images code
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 67.6 67.6 80.8 ResMLP-S low-tau code
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 67.5 67.5 80.2 ResMLP-S low-tau; trained with masked images code
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 66.6 81.8 84.8 ViT-B k=6; mid-tau; trained with masked images code
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 65.3 81.5 84.8 ViT-B k=5; mid-tau; trained with masked images code
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 64.5 81.4 84.8 ViT-B k=4; mid-tau; trained with masked images code
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 64.4 81.4 84.5 ViT-B k=6; mid-tau code
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 63.4 63.4 82.8 ResNet-50 low-tau code
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 62.1 75 80.8 ResMLP-S mid-tau code
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 62.1 81 84.8 ViT-B k=3; mid-tau; trained with masked images code
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 60.8 75.4 80.2 ResMLP-S mid-tau; trained with masked images code
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 59.9 77.8 82.3 ResNet-50 mid-tau; trained with masked images code
ScaleCert (NeurIPS 2021) 55.4 58.5 73 ResNet-50 NA
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 51 78.2 82.8 ResNet-50 mid-tau code
PatchGuard++ (ICLR Workshop 2021) 49.8 49.8 63 BagNet-33 low-tau; trained with masked features code
PatchGuard++ (ICLR Workshop 2021) 48.4 48.4 66.6 BagNet-33 low-tau code
PatchGuard++ (ICLR Workshop 2021) 46.6 46.6 56.5 BagNet-17 low-tau; trained with masked features code
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 46.4 80.3 80.8 ResMLP-S high-tau code
PatchGuard++ (ICLR Workshop 2021) 45.5 45.5 58.8 BagNet-17 low-tau code
PatchGuard++ (ICLR Workshop 2021) 39.9 60.1 66.6 BagNet-33 mid-tau code
PatchGuard++ (ICLR Workshop 2021) 39 60.9 63 BagNet-33 mid-tau; trained with masked features code
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 37.6 84.4 84.5 ViT-B k=6; high-tau code
PatchGuard++ (ICLR Workshop 2021) 36.4 55.2 58.8 BagNet-17 mid-tau code
PatchGuard++ (ICLR Workshop 2021) 34.7 55.4 56.5 BagNet-17 mid-tau; trained with masked features code
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 32.7 82.3 82.8 ResNet-50 high-tau code
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 32.4 84.7 84.8 ViT-B k=6; high-tau; trained with masked images code
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 31.7 84.7 84.8 ViT-B k=5; high-tau; trained with masked images code
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 30.9 84.7 84.8 ViT-B k=4; high-tau; trained with masked images code
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 29.2 82.1 82.3 ResNet-50 high-tau; trained with masked images code
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 29.1 84.6 84.8 ViT-B k=3; high-tau; trained with masked images code
PatchGuard++ (ICLR Workshop 2021) 28 62.9 63 BagNet-33 high-tau; trained with masked features code
PatchGuard++ (ICLR Workshop 2021) 27.6 56.4 56.5 BagNet-17 high-tau; trained with masked features code
Minority Reports (ACNS Workshop 2020) + PatchCleanser (USENIX Security 20222) 26.7 80.1 80.2 ResMLP-S high-tau; trained with masked images code
PatchGuard++ (ICLR Workshop 2021) 22.7 66.3 66.6 BagNet-33 high-tau code
PatchGuard++ (ICLR Workshop 2021) 19.9 58.7 58.8 BagNet-17 high-tau code

(go back to table of contents)

About

A Leaderboard for Certifiable Robustness against Adversarial Patch Attacks