US Agencies Accuse China AI Firms of Model Distillation

United States cybersecurity and intelligence agencies have accused China-based artificial intelligence (AI) companies of engaging in a "systematic extraction" of proprietary functionalities and capabilities from American frontier AI models. This alleged activity, described as occurring at an "industrial-scale," forms the "core" of their AI development strategy, according to a joint advisory issued by the Cybersecurity and Infrastructure Security Agency (CISA), the Federal Bureau of Investigation (FBI), and the National Security Agency (NSA) on May 15, 2024. The advisory warns that these Chinese firms are employing "distillation attacks" to replicate the performance of advanced U.S. AI models without incurring the substantial research and development costs. This process involves training smaller, more efficient models on the outputs of larger, more sophisticated models, effectively extracting their knowledge and capabilities. The targeted models explicitly mentioned in the advisory include Anthropic's Claude, OpenAI's GPT series, Google's Gemini, and xAI's Grok. The agencies assert that this practice not only infringes on intellectual property but also poses significant national security risks. By obtaining advanced AI capabilities through illicit means, Chinese entities could accelerate their own AI development, potentially closing the gap with leading U.S. technology firms and enabling the deployment of AI for malicious purposes, such as sophisticated cyberattacks or disinformation campaigns. The advisory highlights that these distilled models may exhibit similar performance characteristics to their parent models but could also inherit underlying biases or vulnerabilities. Furthermore, the scale of this operation suggests a coordinated effort by multiple Chinese entities, indicating a deliberate strategy to leverage foreign intellectual property for domestic technological advancement. The U.S. agencies are urging U.S. organizations to implement robust security measures to detect and prevent such intellectual property theft and to be vigilant against the proliferation of AI models derived from these attacks. The advisory also touches upon the potential for these distilled models to be used in critical infrastructure or sensitive applications, raising concerns about their reliability and security. The agencies are encouraging the reporting of any suspicious activities or intelligence related to these AI distillation attacks to aid in further investigation and mitigation efforts. This development underscores the escalating competition in the global AI landscape and the increasing focus on protecting intellectual property and national security interests in the realm of advanced artificial intelligence.