Edge-Based Re-identification: Restoring Placeholders safely

When external models respond, placeholders must be swapped back before they reach users. We detail restore logic in redact.rs.

Restoring Redacted Tokens at the Local Edge

After the external model finishes reasoning on redacted text, it returns output containing placeholders like [EMAIL_0]. ATL-Trust looks up the local translation dictionary and restores the actual data at the secure edge, preserving client readability.

The re-identification engine loops through the local mapping dictionary, replacing placeholders with original values in the output payload before delivering it to the end user.

// Restore function in redact.rs
pub fn restore(mut text: String, mappings: &HashMap) -> String {
    for (placeholder, original) in mappings {
        text = text.replace(placeholder, original);
    }
    text
}

Re-identification Processing

The restore endpoint loads the mapping dictionary from memory, replacing placeholders in the model's response before sending it to the client device. This translation happens entirely in local RAM.

By performing re-identification on-device or within the company's VPC, we keep translation keys isolated, preventing sensitive data exposure on public networks.

• Re-identification happens within the secure local network.
• Prevents leak of translation keys to external networks.
• Runs in O(N) where N is the number of active redactions.