# An Optimal Control View of Adversarial Machine Learning

**Authors:** Xiaojin Zhu  
**arXiv:** [1811.04422](https://arxiv.org/abs/1811.04422)  
**Date:** 2018-11-11  
**Summary page:** https://www.summarizepaper.com/en/paper/1811.04422v1/

## Key points

- The paper "An Optimal Control View of Adversarial Machine Learning" by Xiaojin Zhu presents a novel perspective on adversarial machine learning through optimal control theory.
- The framework views the machine learner as a dynamical system and considers adversarial actions as inputs with control costs determined by the adversary's objectives.
- Various forms of adversarial attacks in machine learning are covered, including test-item attacks, training-data poisoning, and adversarial reward shaping.
- Researchers are encouraged to leverage advancements in control theory and reinforcement learning to enhance strategies for defending against malicious actors manipulating AI systems.
- Integrating control-theoretic principles into the study of adversarial machine learning is crucial for bolstering security and reliability in AI systems.

## Summary

In the paper "An Optimal Control View of Adversarial Machine Learning" by Xiaojin Zhu, the author presents a novel perspective on adversarial machine learning through the lens of optimal control theory. This framework views the machine learner as a dynamical system and considers adversarial actions as inputs with control costs determined by the adversary's objectives to cause harm while remaining undetected. The approach encompasses various forms of adversarial attacks in machine learning such as test-item attacks, training-data poisoning, and adversarial reward shaping. By adopting an optimal control viewpoint, researchers in adversarial machine learning are encouraged to leverage advancements in control theory and reinforcement learning to enhance their strategies for defending against malicious actors seeking to manipulate AI systems. This unique perspective not only sheds light on the dynamics of adversarial interactions but also provides a solid foundation for developing more robust and resilient machine learning algorithms capable of mitigating the impact of adversarial threats. Overall, Zhu's work underscores the importance of integrating control-theoretic principles into the study of adversarial machine learning to bolster security and reliability in AI systems.