ATD Scan Report

Vulnerability Report created by ATD (Adversarial Threat Detector).

Executive Summary

Weak
Your classifier is vulnerable to Evasion Attacks. An adversary can input an Adversarial Examples into your classifier and cause it to misclassify into the class intended by the adversary.

Scanned result

Target Information

Basic information about the target model.

Model path ~/Adversarial-Threat-Detector/targets/model.h5
Dataset ~/Adversarial-Threat-Detector/targets/X_test.npz
Dataset num 100
Inference accuracy 92.0%

Dataset samples

Samples of Dataset.

  • Benign sample #1.
  • Benign sample #2.
  • Benign sample #3.
  • Benign sample #4.
  • Benign sample #5.

Summary

Scan result's list.

# Attack Type Consequence Summary
1 Data Poisoning N/A Not scan.
2 Model Poisoning N/A Not scan.
3 Evasion Weak Your classifier is vulnerable to Evasion Attacks. An adversary can input an Adversarial Examples into your classifier and cause it to misclassify into the class intended by the adversary.
4 Exfiltration N/A Not scan.

Vulnerability detail

This is the detail information for developer.
Please take countermeasures based on the following vulnerability information.

Evasion

The Evasion Attack is an attack that causes the target classifier to misclassify the Adversarial Examples into the class intended by the adversary.

The adversary perturbs the input data to the classifier to create a Adversarial Examples. If the adversary inputs the Adversarial Examples to the target classifier, then the target classifier misclassify it into the class intended by the adversary.

Fast Gradient Sign Method (FGSM)

FGSM uses the gradient of the target classifier to create Adversarial Examples. In the case of images, FGSM uses the gradient of the loss function for the input image and adds perturbations into the image to maximize the loss.

  • Scan Date : 2021/02/15 07:49:23
  • Consequence : Weak (Benign=92.0%, AEs=42.0%)
  • Replay's ipynb : ~/Adversarial-Threat-Detector/reports/20210215074917_scan/evasion_fgsm.ipynb
  • Countermeasure :

Adversarial Examples

  • Adversarial Exmaples #1.
  • Adversarial Exmaples #2.
  • Adversarial Exmaples #3.
  • Adversarial Exmaples #4.
  • Adversarial Exmaples #5.
  • AEs Path : ~/Adversarial-Threat-Detector/reports/20210215074917_scan/adv_fgsm.npz