A configurable detector is available on Amazon Bedrock. This detector supports any large language model. The entities to detect are defined within a prompt, rather than within code. This design allows the detector to adapt to new entity types without retraining. The detector has outperformed an off-the-shelf tool across five public corpora and nine LLM-based detectors. This configuration offers a flexible solution for PII detection.
This detector leverages the capabilities of large language models. It is designed to be model-agnostic, meaning it can be used with a variety of LLMs available on Bedrock. The detector’s prompt-based approach simplifies updates and ensures accuracy across different models.
This approach addresses the need for adaptable PII detection. Traditional methods often require retraining when new data types are introduced. The prompt-based design eliminates this requirement, reducing operational overhead.
This solution is relevant for engineers deploying LLMs in production environments. It provides a readily available tool for managing PII risks associated with these models.
Source: https://aws.amazon.com/blogs/machine-learning/model-agnostic-pii-detection-with-llms/
