1、Andrew Lohn,Anna Knack,Ant Burke and Krystal Jackson 0 Autonomous Cyber Defence A roadmap from lab to ops Andrew Lohn,Anna Knack,Ant Burke and Krystal Jackson June 2023 Andrew Lohn,Anna Knack,Ant Burke and Krystal Jackson 1 About CETaS.3 About CSET Georgetown.3 Acknowledgements.3 Executive Summary.4

2、 Recommendations.6 Introduction.8 What is autonomous cyber defence?.9 Research aims and methodology.11 1.Developments in Reinforcement Learning.13 1.1.What is reinforcement learning?.13 1.2.Reinforcement learning for cybersecurity.15 2.Technical Challenges.19 2.1.Complexity and combinatorial explosi

3、ons.19 2.2.Neural network architectures.20 2.3.Computational requirements.22 2.4.Defining rewards.23 2.5.Security concerns:offensive agents.23 2.6.Securing the securers.24 2.7.Transferability.25 3.Policy Challenges.26 3.1.Human-machine teaming.27 3.2.Testing.29 3.3.Skills gaps,shortages,and the futu

4、re of work.30 3.4.Data access.31 3.5.Strategic horizon funding.32 3.6.Liability and criminal responsibility.33 3.7.Supply chain security and export control.33 3.8.Social good and equality of access.34 4.Conclusions and Recommendations.35 4.1.Nurture the field.35 4.1.1.Invest in scaling gyms and agen

5、ts.35 4.1.2.Build and provide testing and training ranges.36 4.1.3.Coordinate data sharing.36 4.1.4.Develop,attract,and retain talent.36 Andrew Lohn,Anna Knack,Ant Burke and Krystal Jackson 2 4.1.5.Host competitions.36 4.2.Guide the field.37 4.2.1.Invest in understanding the risks and benefits of au

6、tonomous cyber defence.37 4.2.2.Determine whether defender agents require attacker agents.37 4.2.3.Determine thresholds for authorisation of autonomous cyber defence agents.37 4.2.4.Determine priorities for autonomous cyber defence agents.37 Appendix A:Methodology.39 A.1.Research approach.39 A.1.1.L