GAIR Paper 111 | Who Moved My Large Model? IJCAI 2026 Latest Survey Reveals the Large Model "Implicit Identity" Anti-Counterfeiting Battle

The latest IJCAI 2026 survey focuses on the large model security field, introducing the concept of "implicit identity" anti-counterfeiting. This technology aims to address new attack methods such as malicious distillation, wrapping, and core data theft, providing endogenous security protection for high-value large models.

As large model training costs climb to the hundreds of millions of dollars level, the value of model assets is becoming increasingly prominent. The large-scale suspected malicious distillation incident of Claude disclosed by Anthropic indicates that traditional external defenses such as account bans and access control are difficult to cope with covert "freeloading" behaviors.

After introducing the implicit identity mechanism, large models can embed invisible markers in output content, thereby tracing data sources and identifying illegal copying. This will promote the industry to establish a more comprehensive model copyright protection system, reducing security concerns for enterprises deploying large models.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.