Jiatong Yi, Yanyang Li
SimMIA is a new framework for membership inference attacks on large language models that excels in black-box settings using only generated text, achieving state-of-the-art results.
Researchers have developed a new method called SimMIA to test whether specific data was used to train large language models, even when only the generated text is available and without needing access to the model's internal workings. This is important because it helps ensure transparency and accountability in how these models are trained. The team also introduced a benchmark called WikiMIA-25 to evaluate the effectiveness of such methods on modern language models. SimMIA has shown to perform exceptionally well, even compared to methods that have access to more detailed model information.
