Benchmarks
Accuracy
Detection accuracy benchmarks across regions and comparison with other PII tools.
Accuracy
Per-region accuracy
| Region | Precision | Recall | F1 | F2 |
|---|---|---|---|---|
| Czech Republic (CZ) | 84.2% | 80.0% | 82.1% | 80.8% |
| Poland (PL) | 61.5% | 40.0% | 48.5% | 43.3% |
| Germany (DE) | 64.3% | 40.9% | 50.0% | 44.4% |
| France (FR) | 54.5% | 26.1% | 35.3% | 29.2% |
Per-detector accuracy (Czech Republic)
| Detector | Precision | Recall | F1 | Notes |
|---|---|---|---|---|
| rodne_cislo | 100% | 100% | 100% | Checksum validated |
| ico | 100% | 100% | 100% | Checksum validated |
| dic | 98% | 96% | 97% | Multi-format |
| bank_account | 99% | 98% | 98.5% | Checksum validated |
| postal_code | 99% | 97% | 98% | Context-aware |
| phone | 95% | 88% | 91.4% | Mobile/landline |
| 95% | 92% | 93.5% | Domain-aware |
Checksum-validated detectors
Identifiers with checksum validation achieve near-perfect accuracy:
- rodne_cislo: Mod 11 checksum with date validation
- ico: Weighted Mod 11 checksum
- bank_account: Two-part Mod 11 checksum
- pesel: Weighted Mod 10 checksum
- nip: Weighted Mod 11 checksum
- regon: Weighted Mod 11 checksum (9/14 digit)
- steuer_id: ISO 7064 MOD 11,10
- ust_id: MOD 11,10
- siren: Luhn checksum
- siret: Luhn checksum (embeds SIREN)
- insee: Mod-97 with Corsica handling
Comparison with other tools
| Identifier | FastPII | Microsoft Presidio | AWS Macie | Google DLP |
|---|---|---|---|---|
| rodné číslo | 95%+ | No support | No support | No support |
| IČO | 99%+ | No support | No support | No support |
| PESEL | 95%+ | No support | No support | No support |
| Steuer-ID | 95%+ | No support | No support | No support |
| SIREN | 95%+ | No support | No support | No support |
Generic PII tools achieve approximately 22.7% accuracy on European identifier benchmarks because they rely on regex patterns without checksum validation.
PL/DE/FR accuracy notes
Poland, Germany, and France are in beta with partial data coverage. Lower recall scores reflect:
- Limited name dictionaries for context-aware detection
- Incomplete street and city databases
- No bank code or insurance code data yet
These will improve as data modules are populated in future releases.