AutoCoMBOT: Autonomy in Cyberspace through robot learning and Man-BOt Teaming
Abstract
Preparing for the high-tech warfare of the future is a grand challenge. The emergence of autonomous cyber forces and agents and the complex ecosystem of Multi-Domain Operations create layers of sophistication and new opportunities for adversarial acts. Autonomous intelligent agents, a.k.a., bots, can be cyber-physical or just in the form of a program, but are almost always envisioned to outnumber the human forces. Future bots will be capable of gathering sensory data and other inputs on patterns and events and actively making decisions and playing deception games, while human agents will be working together with the bots to make critical decisions. To date, our understanding of the possible warfare scenarios involving bots, robust intelligence, as well as needed attack and defense strategies for human-bot activity is very limited. This MURI proposal, called AutoComBOT Ð Autonomy in Cyberspace through rObot learning and Man-BOt Teaming, introduces a novel multi-pronged approach to address the standing challenges for future warfare involving multitude of cyber bots. Important constraints in bot warfare scenarios include limited access to a reliable central controller, speed of action, synchronization, and coordination among the distributed cyber-physical, cyber, and human entities. Our interdisciplinary team of US and Australian PIs will develop a new comprehensive foundational framework for safe and robust artificial intelligence intrinsic to autonomous agents, distributed bots, self-adaptivity, introspection, human-AI teaming, as well as automated methods for human-AI games, deception and recovery in dynamic settings. The focus of the AutoComBOT research is on three modular but inter-linked thrusts: 1) Robust Learning: We formulate key challenging scenarios in robust learning with distributed agents as instances of graph optimization problems; the formulation is leveraged to derive new bounds, metrics, attacks, and defenses. 2) Introspection/Anti-fragility Adaptation: This is formulated as active learning scenarios with offline preprocessing and training, online introspection/antifragility adaptation, and dynamic adversary deception. 3) New Team Science Concepts for Cyber Bots: This thrust focuses on effective team assembly, shared cognition for human-AI interaction, quantifying performances over time and tasks, robust team training by reinforcement learning, robustness to low-probability high-impact events, and self-adaptive mixed initiatives control and deception. The AutoComBOT results will significantly contribute to the efforts of the DoD/DST to prepare for future bot-involved warfare and ensure national security of the US, Australia, and our allies. The team will work closely with ARL and other DoD/DST labs in the US and Australia to realize this mission. To ensure a wide-scale adoption, we provide SW tools and APIs to make the algorithms robust and customizable with standard interfaces for ML packages. The thrusts will be evaluated and benchmarked on public data for multiple dynamic signal modalities, e.g., image, video, audio, text, and RF. AutoComBOT has an outstanding multidisciplinary team of diverse PIs; each PI brings in a strong set of relevant work/recent publications as well as novel ideas to the project
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 01, 2023
- Source ID
- W911NF2110322
Entities
People
- Farinaz Koushanfar
Organizations
- Army Contracting Command
- United States Army
- University of California, San Diego