Identifying Pre-training Data in LLMs: A Neuron Activation-Based Detection Framework
This addresses dataset contamination and bias issues in LLMs, providing a tool for detecting pre-training data, though it is incremental as it builds on the existing PDD task with a new method.
The paper tackles the problem of identifying whether specific data was included in a large language model's pre-training corpus, which is important for legal and ethical concerns, and introduces NA-PDD, a method based on neuron activation patterns that significantly outperforms existing methods across benchmarks and models.
The performance of large language models (LLMs) is closely tied to their training data, which can include copyrighted material or private information, raising legal and ethical concerns. Additionally, LLMs face criticism for dataset contamination and internalizing biases. To address these issues, the Pre-Training Data Detection (PDD) task was proposed to identify if specific data was included in an LLM's pre-training corpus. However, existing PDD methods often rely on superficial features like prediction confidence and loss, resulting in mediocre performance. To improve this, we introduce NA-PDD, a novel algorithm analyzing differential neuron activation patterns between training and non-training data in LLMs. This is based on the observation that these data types activate different neurons during LLM inference. We also introduce CCNewsPDD, a temporally unbiased benchmark employing rigorous data transformations to ensure consistent time distributions between training and non-training data. Our experiments demonstrate that NA-PDD significantly outperforms existing methods across three benchmarks and multiple LLMs.