The Bureau of Industry and Security recently announced new export control rules regarding anti-terrorism and regional stability, which will significantly affect the trade of high-end AI chips. As with any complex regulatory change, there is a risk of unintended consequences and surprise challenges in its implementation. The Bureau has therefore rightly encouraged comments and collaborative efforts between industry stakeholders and government regulators so that we can prevent adversaries from acquiring advanced AI chips.
As noted in a recent Center for Strategic and International Studies report, one important risk is chip smuggling. The new export restrictions apply only to U.S.-owned entities in China, not to Chinese-owned entities operating in the rest of the world. In recent years, Illegal chip exports have outpaced those during the Cold War, and people are being caught by Chinese customs while smuggling AI acceleration chips. In response, the U.S. government should develop robust international AI forensics capabilities.
In attempting to address this issue, we should look to history. The U.S. experience with nuclear smuggling during the Cold War provides an important lesson in the deterrence value of investing in forensics infrastructure. In the 1990s, the unprecedented theft of nuclear material from the former Soviet Union signaled a clear shift in the significance of the nuclear smuggling problem. In response to these incidents, nuclear forensic science was developed through the work of an informal group of concerned scientists, the Nuclear Forensics International Technical Working Group (ITWG). In the 30 years since nuclear forensic research emerged as a new branch of science, ITWG Nuclear Forensics Laboratories (INFL) have engaged in research on forensic investigations.
Policymakers should consider how similar approaches could be applied today, and how we can use forensic science to deter the theft of advanced chips and advanced AI models. Here, in broad strokes, is one potential approach to consider: the U.S. could establish AI forensics laboratories within the state's cybersecurity infrastructure. The capability for AI forensics could be augmented by the establishment of a national AI forensics laboratory network, which would have the ability to perform forensic analysis of AI-driven incidents.
Developing an AI forensics program could involve convening expertise from multiple government agencies and countries. Developing an AI forensics laboratory network could facilitate collaboration between these entities, and the network could consist of laboratories with complementary cyber forensics capabilities. In addition to their technical skills, each laboratory could develop AI security norms. Early research areas could include remote sensing of data centers to detect large training runs, fingerprinting AI inference chips from random sampling, and fingerprinting models based on outputs/API access. Law enforcement agencies could advise on how best to develop procedures for collecting evidence. Together, these international entities could collaborate to build a robust AI and chips forensic science.
Although much more should be said on the topic, establishing a laboratory network is one idea worth considering to provide a framework for collaboration. It could allow experts within various fields to contribute their knowledge towards achieving successful outcomes when dealing with criminal cases involving AI technologies.