Asymmetry in Coevolving Adversarial Systems

Asymmetries in adversarial systems arise from differences in the "situations" of attackers and defenders, for instance corresponding to differences in information access or cost/benefit tradeoffs. While numerous studies have shown that asymmetry is important, less has been done to rigorously characterize its impact or specify methods by which it can be leveraged by defenders. This paper presents a formal framework for analyzing the origins and roles of asymmetric advantage in coevolving adversarial systems, and uses this framework to develop quantitative tools for understanding and exploiting asymmetry. The proposed framework explains why asymmetry has such profound impact on the behavior of coevolving systems, and reveals a key feature of these systems: they can be reverse-engineered using only limited measurements. The analysis yields several new results, including 1.) a demonstration that machine learning systems increasingly deployed to harden vulnerabilities in essential systems are themselves highly vulnerable, and 2.) a methodology for designing "predictability-oriented" defenses that shifts advantages of asymmetry toward defenders. An empirical case study involving detection of money-laundering activity within the global interbank transactions system illustrates the utility of the proposed analytic framework in a high-consequence setting.