Coalition Situational Understanding Via Adaptive, Trusted and Resilient Distributed Artificial Intelligence Analytics

Abstract Artificial intelligence (AI) and machine learning (ML) promises transformative effects on coalition multidomain and hybrid operations. AI/ML approaches that support situational understanding in the context of ad-hoc coalition operations at the tactical edge are of considerable current research interest. Coalition operations need distributed AI/ML that is robust to contested and complex multi-actor situations. Information with a high degree of complexity needs to be collected across a range of sensed modalities and processed at high tempo, aligned with human needs and capabilities. Research carried out in the joint US/UK Distributed Analytics and Information Science (DAIS) programme since 2016 is addressing coalition needs for adaptable, trusted and resilient AI/ML: adaptable AI refers to AI systems which can rapidly adapt in dynamic situations; trusted AI means that human users are able to rapidly calibrate their trust in AI systems; and resilient AI concerns AI systems which are resilient to adversary attacks and deception. This paper focuses on DAIS research centred on the rapid exploitation and integration of coalition AI/ML assets including both symbolic (logic-based) and subsymbolic (deep neural network-based) approaches. To provide a focus for the paper, we consider settings involving detecting patterns of interrelated events that form situations of interest where only sparse training data (for ML) is available. Rapid trust calibration is addressed via a combination of explainable AI - involving both symbolic and subsymbolic approaches to explainability - and effective management of uncertainty - considering both aleatoric and epistemic types of uncertainty. While not the primary focus of this paper, resilience is considered by showing that the integrated neuro-symbolic system performs robustly against targeted model-poisoning adversarial attacks, and also that the processing of multimodal sensed data by explainable AI/ML services makes the integrated system much harder to attack. For easier assimilation of the programme of work, we use a single integrated case study of a coordinated attack in an urban setting based on the NATO Anglova exercise
  • Alun Preece (Cardiff)
  • Dave Braines (IBM UK)
  • Federico Cerutti
  • Gavin Pearson (Dstl)
  • Lance Kaplan (ARL)
Date Oct-2021