Improving Name and Identity Redaction using Heuristic Regexes

Scrubbing name entities is challenging because they are often mistaken for normal words. Learn how local heuristic regexes provide low-latency redaction.

Heuristic Scanning for Names in Unstructured Prompts

Unlike credit card formats, name entities lack fixed shapes. Using heavy Named Entity Recognition (NER) models for every prompt adds massive CPU delays. ATL-Trust uses localized heuristic patterns to flag name entities efficiently, preserving processing speed.

Our name heuristic matches capitalization patterns following introductory phrases like 'dear' or 'my name is'. This flags name entities without the need for resource-intensive transformer-based models.

// Regex for name captures in redact.rs
let re_name = Regex::new(r"(?i)\b(?:my name is|i am|name is|dear|sender:)\s+([A-Z][a-z]+(?:\s+[A-Z][a-z]+)?)\b").unwrap();
for cap in re_name.captures_iter(&scan_text) {
    let matched_name = cap.get(1).unwrap().as_str().to_string();
    // Replace with placeholder...
}

Name Scanner Implementation

The engine scans for introductory markers like 'my name is' or 'dear' followed by capitalized strings, replacing matches with placeholders to protect user identity. This runs within the local edge gateway.

By keeping the translation keys in secure local memory, we ensure that PII is never exposed to external networks, satisfying data-minimization policies under GDPR.

• Avoids heavy GPU requirements of transformer-based NER.
• Ensures local, low-latency name scrubbing within VPC limits.
• Reduces false positives by validating patterns before replacements.