CyberML // cybersecurity of AI systems
This NIST Trustworthy and Responsible AI report develops a taxonomy of concepts and defnes terminology in the feld of adversarial machine learning (AML). The taxonomy is built on surveying the AML literature and is arranged in a conceptual hierarchy that includes key types of ML methods and lifecycle stages of attack, attacker goals and objectives, and attacker capabilities and knowledge of the learning process. The report also provides corresponding methods for mitigating and managing the consequences of attacks and points out relevant open challenges to take into account in the lifecycle of AI systems. The terminology used in the report is consistent with the literature on AML and is complemented by a glossary that defnes key terms associated with the security of AI systems and is intended to assist non-expert readers. Taken together, the taxonomy and terminology are meant to inform other standards and future practice guides for assessing and managing the security of AI systems, by establishing a common language and understanding of the rapidly developing AML landscape.
Humans are capable of strategically deceptive behavior: behaving helpfully in most situations, but then behaving very differently in order to pursue alternative objectives when given the opportunity. If an AI system learned such a deceptive strategy, could we detect it and remove it using current state-of-the-art safety training techniques? To study this question, we construct proof-of-concept examples of deceptive behavior in large language models (LLMs). For example, we train models that write secure code when the prompt states that the year is 2023, but insert exploitable code when the stated year is 2024. We find that such backdoor behavior can be made persistent, so that it is not removed by standard safety training techniques, including supervised fine-tuning, reinforcement learning, and adversarial training (eliciting unsafe behavior and then training to remove it). The backdoor behavior is most persistent in the largest models and in models trained to produce chain-of-thought reasoning about deceiving the training process, with the persistence remaining even when the chain-of-thought is distilled away. Furthermore, rather than removing backdoors, we find that adversarial training can teach models to better recognize their backdoor triggers, effectively hiding the unsafe behavior. Our results suggest that, once a model exhibits deceptive behavior, standard techniques could fail to remove such deception and create a false impression of safety
