Amazon Nova 2 Lite pipeline automatically redacts PII in images
Uses Nova 2 Lite, SAM 3, and Textract to handle tricky edge cases like reflections and partial faces.
Amazon announced a new AI-powered pipeline for automatically redacting Personally Identifiable Information (PII) in images, using Amazon Nova 2 Lite as the intelligent orchestrator. The solution addresses the challenge of PII appearing in unexpected places in images—such as faces reflected in car surfaces, partially visible street signs, or documents on desks—which often defeat simple masking tools. Nova 2 Lite, a fast and cost-effective multimodal foundation model available in Amazon Bedrock, analyzes the entire image context to identify what constitutes sensitive information, then directs specialized downstream tools for precise redaction.
To achieve pixel-level accuracy, the pipeline calls Meta's Segment Anything Model (SAM 3) deployed on Amazon SageMaker AI to generate exact segmentation masks for visual PII like faces and fingerprints. For textual PII (names, addresses, ID numbers), Amazon Textract extracts all text and coordinates, which Nova then evaluates in context. This multi-step approach ensures compliance with GDPR and PCI DSS while preserving the image's overall utility. The solution handles diverse PII types including facial images, biometric data, license plates, and MAC addresses.
- Amazon Nova 2 Lite coordinates the entire pipeline, reasoning about PII in context using its multimodal vision capabilities.
- Integrates Meta's SAM 3 for pixel-level segmentation and Amazon Textract for OCR, enabling precise redaction of visual and textual PII.
- Handles challenging edge cases like faces in reflections, partial documents, and license plates in arbitrary orientations.
Why It Matters
Automates complex image redaction for GDPR/PCI DSS compliance, reducing manual effort and legal risk in data sharing.