Executive Summary
This survey systematically categorizes and analyzes threats and countermeasures in AI systems, covering both training and inference stages, centralized and distributed settings, and conventional and foundation AI models. It highlights key attack paradigms like data poisoning, backdoor attacks, and privacy inference, alongside various defense strategies. The research aims to provide AI practitioners with a thorough understanding of system vulnerabilities and current countermeasures.
Why It Matters
This document is crucial for defense and geopolitical analysts as it details the vulnerabilities and attack vectors in AI systems, which are increasingly integrated into military and national security applications, impacting strategic capabilities and intelligence operations.
Key Takeaways
- AI systems face significant security and privacy challenges across their lifecycle, from training to inference, in both centralized and distributed environments.
- Key attack paradigms include data poisoning, backdoor attacks, privacy inference, and model hijacking, targeting both conventional and foundation AI models.
- The survey provides a comprehensive overview of existing threats and countermeasures, emphasizing the need for robust solutions to develop resilient AI technologies.
Strategic Relevance
Understanding AI security vulnerabilities is strategically vital for national defense, as compromised AI systems can lead to intelligence breaches, operational failures, or enable adversaries to manipulate critical infrastructure and decision-making processes.