Skip to main content

Supported PII & Sensitive Data Types

⬅️ Back to Features Catalog

The LLM-Shield-Proxy employs a 3-Tier Cascade Engine to detect and redact sensitive data. Below is an exhaustive list of the data types natively detected by Tiers 1 and 2, along with representative samples of semantic types covered by Tier 3 NLP models.

🛡️ Tier 1: Microsecond Regex (High-Fidelity Structured Data)​

Tier 1 utilizes zero-allocation, pre-compiled google-re2 DFA regular expressions to detect highly structured, standardized formats in linear O(N) time.

Exhaustive List of Tier 1 Native Detectors:

  1. CREDIT_CARD: Major credit card formats (Visa, MasterCard, Amex, Discover) validated via length and spacing.
  2. SSN: US Social Security Numbers (e.g., XXX-XX-XXXX).
  3. EMAIL: Standard email addresses.
  4. PHONE: International and domestic phone numbers, including extensions.
  5. IP_ADDRESS: IPv4 network addresses.
  6. AWS_API_KEY: Amazon Web Services Access Keys (e.g., AKIA..., ASIA..., sk-...).
  7. GITHUB_PAT: GitHub Personal Access Tokens (e.g., ghp_...).
  8. SSH_PRIVATE_KEY: RSA/DSA/ECDSA/Ed25519 private keys (detects PEM headers).
  9. JWT_TOKEN: JSON Web Tokens.
  10. MRN: Medical Record Numbers (Standardized formats).

[!TIP] You can easily add your own Tier 1 detectors using the Bring Your Own Regex (BYOR) feature via policies.yaml.

🧠 Tier 2: Shannon Entropy (Unstructured Secrets & Cryptography)​

Because developers often leak proprietary API keys or database passwords that don't match a standard Regex format, Tier 2 acts as a mathematical dragnet.

What gets detected:

  • Cryptographic Keys & Salts: High-entropy strings exceeding \tau_H \ge 4.5 bits/symbol.
  • Obfuscated / Smuggled Data: Base64 or Hex-encoded payloads injected into the prompt.
  • Proprietary Tokens: Any alphanumeric string of sufficient length (16+ characters) that is mathematically random enough to be classified as a secret.

🤖 Tier 3: NLP NER Models (Semantic & Contextual Data)​

Tier 3 utilizes Named Entity Recognition (NER) models (like Presidio or local ONNX runtimes) to understand the semantic meaning of low-entropy text.

While Tiers 1 and 2 detect structured data perfectly, Tier 3 is required for contextual data (where the words look like normal English but hold sensitive meaning).

Standard PII (General NER)​

  • PERSON: Full names, patient names, employee names.
  • ORGANIZATION: Company names, hospitals, government agencies.
  • LOCATION: Physical addresses, cities, states, zip codes.
  • DATE_TIME: Specific dates of birth, admission dates, meeting times.

Domain-Specific Models​

If you deploy domain-specific NER models via the Proxy's ONNX runtime, the following domain-specific entities can be redacted for compliance:

🏥 HIPAA & Healthcare (Clinical NER)​

  • PHI_DIAGNOSIS: Medical conditions, ICD-10 codes.
  • PHI_MEDICATION: Prescriptions, dosages.
  • PHI_TREATMENT: Surgical procedures, therapy notes.
  • HEALTH_PLAN_BENEFICIARY: Insurance group numbers, Medicaid IDs.

🏦 SOC 2, PCI-DSS & Financial (Financial NER)​

  • BANK_ACCOUNT: Checking/Savings account numbers.
  • ROUTING_NUMBER: ABA Routing numbers.
  • SWIFT_BIC: International bank identifiers.
  • FINANCIAL_TRANSACTION: Specific ledger amounts or invoice numbers linked to an entity.

Next Steps: Learn how the Proxy substitutes this detected data with synthetic equivalents in the Format-Preserving Synthetic Masking guide.