# AdvDrop: Adversarial Attack to DNNs by Dropping Information

**Authors:** Ranjie Duan, Yuefeng Chen, Dantong Niu, Yun Yang, A. K. Qin, Yuan He  
**arXiv:** [2108.09034](https://arxiv.org/abs/2108.09034)  
**Date:** 2021-08-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2108.09034v1/

## Key points

- Researchers studied the challenge of recognizing abstract objects with lost information in Deep Neural Networks (DNNs).
- Introduced a novel adversarial attack called AdvDrop that strategically drops existing information from images to generate adversarial examples.
- AdvDrop removes imperceptible details to craft adversarial examples, contrasting with traditional attacks that add explicit disturbing information.
- Demonstrated the effectiveness of AdvDrop through extensive experiments and evaluations using ResNet50 model on 2000 correctly classified images from ImageNet.
- Used a pretrained adversarial model ResNet50 as a defense mechanism to assess AdvDrop's efficacy in adversarial training scenarios.
- Metrics such as attack success rate were used to measure performance under different conditions.
- Conducted ablation studies and examined model attention mechanisms to analyze AdvDrop's impact on dropped information.
- Emphasized the importance of developing robust defenses against sophisticated techniques like AdvDrop.

## Summary

In the study "AdvDrop: Adversarial Attack to DNNs by Dropping Information," researchers Ranjie Duan, Yuefeng Chen, Dantong Niu, Yun Yang, A. K. Qin, and Yuan He delve into the challenge of recognizing abstract objects with lost information in Deep Neural Networks (DNNs). The human ability to identify visual objects with minimal details such as contours is contrasted with the struggle of DNNs in this task. To address this issue, the researchers introduce a novel adversarial attack called AdvDrop that strategically drops existing information from images to generate adversarial examples. Unlike traditional attacks that add explicit disturbing information to clean images, AdvDrop removes imperceptible details to craft adversarial examples. Through extensive experiments and evaluations, the effectiveness of AdvDrop is demonstrated and its ability to create challenging examples for current defense systems is highlighted. The experimental setup involved selecting 2000 correctly classified images from ImageNet for evaluation using the ResNet50 model. Additionally, a pretrained adversarial model ResNet50 was used as a defense mechanism to assess the efficacy of AdvDrop in adversarial training scenarios. Metrics such as attack success rate were used to measure performance under different conditions. Furthermore, ablation studies were conducted and model attention mechanisms were examined to analyze AdvDrop's impact on dropped information. This research explores new avenues in adversarial attacks and defense strategies while shedding light on complexities in visual perception for DNNs. It emphasizes the importance of developing robust defenses against sophisticated techniques like AdvDrop.