DSCG1 - Attack and Defense of AI Enabled Systems - Hahna Latonick

BLUF: This presentation outlines the critical vulnerabilities in AI-enabled uncrewed systems and proposes a multi-layered defense strategy combining adversarial training, secure infrastructure, and continuous authorization frameworks to mitigate emerging threats 0:56.

Key Takeaways:
• The integration of AI into military uncrewed systems is rapidly shifting warfare dynamics, with recent drone swarm incidents in the US, Ukraine, and the Middle East demonstrating both the promise and peril of this technology 1:09.
• Adversarial attacks expand the attack surface of these systems, including sensor spoofing to mislead navigation, input deception to corrupt decision-making, data poisoning, and model extraction 3:57.
• Defensive strategies must include adversarial training to improve model robustness, explainable AI to identify biases, and strict physical security to prevent tampering with edge devices 6:25.
• Effective defense also requires AI-driven cybersecurity for anomaly detection, predictive analytics, and counter-UXS technologies like jamming or directed energy weapons to neutralize malicious platforms 8:03.
• Current AI limitations include data dependency, struggles with edge cases, and integration challenges, necessitating a shift from static to ongoing authorization processes for government deployment 9:40.
• Organizations should leverage NIST AI Risk Management Frameworks (AI 100-1 and 600-1) and the Dark Wolf AI Capabilities Authorization Approach to navigate complex regulatory and security requirements 11:10.

Closing Statement: As AI becomes integral to national security, defending these systems requires a paradigm shift toward continuous monitoring, robust adversarial defenses, and adaptive regulatory compliance.

Sources:

• 0:56 Introduction to the inflection point of AI in warfare and global security implications.
• 1:09 Real-world examples of drone swarms and asymmetric deterrence in modern conflicts.
• 3:57 Detailed explanation of adversarial attacks including sensor spoofing and data poisoning.
• 6:25 Practical defensive strategies such as adversarial training and physical security measures.
• 8:03 Use of AI-driven cybersecurity and counter-UXS technologies for threat mitigation.
• 9:40 Limitations of AI systems and the need for dynamic authorization processes.
• 11:10 Overview of NIST frameworks and the AI Capabilities Authorization Approach for compliance.