# Universal adversarial perturbations for multiple classification tasks with quantum classifiers

**Authors:** Yun-Zhong Qiu  
**arXiv:** [2306.11974](https://arxiv.org/abs/2306.11974)  
**Date:** 2023-06-21  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2306.11974v3/

## Key points

- Researchers in quantum adversarial machine learning are studying the vulnerability of quantum learning systems to adversarial perturbations and developing defense strategies.
- Universal adversarial perturbations can deceive quantum classifiers by making small alterations to input samples, posing a serious threat to quantum machine learning models.
- A recent paper by Yun-Zhong Qiu explores universal perturbations in heterogeneous classification tasks, showing that even accurate quantum classifiers can be deceived by a single carefully crafted universal perturbation.
- The study uses advanced quantum continual learning models with elastic weight consolidation methods to prevent catastrophic forgetting.
- Real-life datasets like hand-written digits and medical MRI images demonstrate the effectiveness of generating universal perturbations for diverse classification tasks.
- These findings provide insights into creating simple and efficient methods for generating universal perturbations in various scenarios, guiding future advancements in quantum learning technologies.

## Summary

In the field of quantum adversarial machine learning, researchers are exploring the vulnerability of quantum learning systems to adversarial perturbations and developing defense strategies. One key focus is on universal adversarial perturbations, which can deceive a quantum classifier by making small alterations to input samples. This area of study is crucial as universal perturbations have the potential to simplify malicious attacks and pose a serious threat to quantum machine learning models. A recent paper by Yun-Zhong Qiu titled "Universal adversarial perturbations for multiple classification tasks with quantum classifiers" highlights the exploration of universal perturbations in heterogeneous classification tasks. The research reveals that even highly accurate quantum classifiers on distinct tasks can be deceived by a single carefully crafted universal perturbation. This finding is demonstrated using advanced quantum continual learning models with elastic weight consolidation methods to prevent catastrophic forgetting. Real-life datasets, including hand-written digits and medical MRI images, were utilized to showcase the effectiveness of generating universal perturbations for diverse classification tasks. These results offer valuable insights into creating simple and efficient methods for generating universal perturbations in various scenarios, providing crucial guidance for future advancements in quantum learning technologies. This research sheds light on the potential risks posed by adversarial attacks in the realm of quantum machine learning and underscores the importance of developing robust defense mechanisms to safeguard against such threats.