DataShield 🔒

Privacy-preserving data sanitization for AI training.

DataShield detects and removes sensitive information (PII, secrets, medical, financial data) from datasets before fine-tuning or RAG. It implements anonymization, differential privacy, and data minimization techniques.

---

Features

Feature	Description
PII Detection	Detects emails, phones, SSNs, passports, national IDs, and more
Secret Scanning	Finds API keys, tokens, passwords, certificates, and credentials
Sensitive Classification	Classifies medical, financial, legal, and personal data by keywords
Presidio Detection	Optional ML-powered PII detection via Microsoft Presidio
Anonymization	Replaces sensitive values with anonymized tokens (deterministic)
Redaction	Replaces sensitive content with [REDACTED]
Data Minimization	Removes unnecessary fields while preserving required ones
Transformation	Applies category-specific transforms (hashing, masking, etc.)
Differential Privacy	Adds calibrated Laplace/Gaussian noise with configurable epsilon
k-Anonymity	Ensures each record is indistinguishable from k-1 others
Epsilon Calculator	Estimates optimal privacy budget based on dataset characteristics
GDPR Compliance	Checks Art. 5, 9, 17, 25, 32, 35 compliance
HIPAA Compliance	Checks Privacy Rule, Security Rule, Minimum Necessary
MCPGuard Policies	Generates MCPGuard-compatible YAML security policies
mcp-taxonomy Adapter	Normalizes findings to the canonical MCP security taxonomy
MCPscop Integration	Forwards findings to MCPscop dashboard via webhook
HTML/JSON/Console Reports	Rich reports for auditing and sharing
Multi-format Input	Supports JSON, JSONL, and CSV datasets

Installation

pip install datashield-ai

With Presidio support (enhanced ML-based PII detection):

pip install datashield-ai[presidio]

With taxonomy integration:

pip install datashield-ai[taxonomy]

Quick Start

# Scan a dataset for sensitive data
datashield scan dataset.json

# Sanitize a dataset (detect + anonymize)
datashield sanitize dataset.json sanitized.json

# Anonymize with differential privacy
datashield anonymize dataset.json anonymized.json --epsilon 0.5 --k 5

# Verify compliance
datashield verify sanitized.json

# Generate MCPGuard policy
datashield policies dataset.json -o mcpguard_policy.yaml

# Generate HTML report
datashield report dataset.json -o report.html

Usage

Scan a dataset

# Basic scan (auto-detects format from extension)
datashield scan data.json
datashield scan data.jsonl
datashield scan data.csv

# Scan with confidence threshold
datashield scan data.json --threshold 0.6

# Exclude certain fields
datashield scan data.json --exclude metadata,internal_id

# Forward findings to MCPscop dashboard
datashield scan data.json --mcpscop

# Output as JSON
datashield scan data.json --format json -o scan_report.json

Sanitize a dataset

# Detect and anonymize sensitive data
datashield sanitize data.json sanitized.json

# Full pipeline with all techniques
datashield sanitize data.json sanitized.json \
  --anonymize true --redact true --minimize true --transform true

# Generate scan report alongside sanitized data
datashield sanitize data.json sanitized.json -r report.html

Anonymize with privacy guarantees

# Differential privacy + k-anonymity
datashield anonymize data.json anonymized.json --epsilon 1.0 --k 5

Verify compliance

# Check GDPR and HIPAA compliance
datashield verify sanitized.json

# Compare with original
datashield verify sanitized.json --original original.json

Generate MCPGuard policies

# Generate YAML security policy from dataset findings
datashield policies dataset.json -o mcpguard_policy.yaml

# Specify custom MCPGuard target
datashield policies dataset.json --target http://my-mcp-server:8000

Generate reports

# HTML report with visualizations
datashield report data.json -o report.html

# JSON report for programmatic use
datashield report data.json -o report.json --format json

Configuration

DataShield supports configuration via environment variables (prefix DATASHIELD_) or a .env file:

# .env
DATASHIELD_THRESHOLD=0.5
DATASHIELD_EXCLUDE_FIELDS=metadata,debug
DATASHIELD_DEFAULT_EPSILON=0.5
DATASHIELD_DEFAULT_K=10
DATASHIELD_MCPSCOP_URL=http://mcpscop:8080
DATASHIELD_MCPSCOP_API_KEY=your-key

Integration with Ecosystem

Project	Integration
MCPGuard	Generates MCPGuard-compatible YAML data policies via datashield policies
MCPscop	Forwards normalized findings via --mcpscop flag or MCPscopClient API
mcp-taxonomy	Normalizes findings to canonical taxonomy via datashield_finding_to_taxonomy()
palisade-scanner	Same detector pattern and architecture

API

import asyncio
from datashield.scanner import Scanner
from datashield.detectors import PIIDetector, SecretScanner, PresidioDetector
from datashield.taxonomy import datashield_finding_to_taxonomy
from datashield.policies.mcpguard import MCPGuardPolicyGenerator

# Create scanner with detectors
scanner = Scanner(detectors=[PIIDetector(), SecretScanner()])

# Scan dataset
data = [{"email": "user@example.com", "api_key": "sk-1234"}]
report = asyncio.run(scanner.scan(data))

print(f"Risk score: {report.risk_score}")
print(f"Findings: {report.total_findings}")

# Normalize to mcp-taxonomy
for finding in report.findings:
    event = datashield_finding_to_taxonomy(finding)
    print(f"  → {event.attack_category.value}: {event.title}")

# Generate MCPGuard policy
gen = MCPGuardPolicyGenerator()
policy = gen.from_findings(report.findings)
print(gen.to_yaml(policy))

Project Structure

datashield/
├── src/datashield/
│   ├── cli.py              # Typer CLI interface (6 commands)
│   ├── scanner.py           # Core scanning engine + Pydantic models
│   ├── config.py            # pydantic-settings configuration
│   ├── taxonomy.py          # mcp-taxonomy adapter
│   ├── detectors/           # Detection modules
│   │   ├── pii_detector.py
│   │   ├── secret_scanner.py
│   │   ├── sensitive_classifier.py
│   │   ├── pattern_matcher.py
│   │   └── presidio_detector.py   # Optional ML-based PII detection
│   ├── sanitizers/          # Sanitization modules
│   │   ├── anonymizer.py
│   │   ├── redactor.py
│   │   ├── minimizer.py
│   │   └── transformer.py
│   ├── privacy/             # Privacy preservation
│   │   ├── differential.py
│   │   ├── k_anonymity.py
│   │   └── epsilon_calculator.py
│   ├── compliance/          # Compliance verification
│   │   ├── gdpr.py
│   │   ├── hipaa.py
│   │   └── verifier.py
│   ├── reporters/           # Output formats
│   │   ├── console.py       # Rich console output
│   │   ├── json.py
│   │   └── html.py          # Jinja2 HTML reports
│   ├── policies/            # Policy generation
│   │   └── mcpguard.py      # MCPGuard YAML policy generator
│   ├── integrations/        # Ecosystem integrations
│   │   └── mcpscop.py       # MCPscop webhook client
│   └── utils/
│       └── crypto.py        # Cryptographic utilities
└── tests/                   # 202+ tests (92% coverage)

Compliance

DataShield helps verify compliance with:

• GDPR (Art. 5, 9, 17, 25, 32, 35) — Data minimization, right to erasure, DPIA
• HIPAA (Privacy Rule, Security Rule, Breach Rule) — PHI de-identification, minimum necessary

Academic References

• arXiv:2605.25716 — Efficient and Privacy-Preserving Architecture for Cross-Institutional Collaborative RAG
• arXiv:2605.25791 — Efficient and Privacy-Preserving Distribution Statistics Analytics on Mobile Spatial Data
• arXiv:2605.25002 — Verifiable Secure Aggregation via Dual Servers with Linear Tags
• arXiv:2605.26019 — Retrieval-Augmented Detection of Potentially Abusive Clauses

Troubleshooting

"mcp_taxonomy is required" error

The taxonomy adapter requires mcp-taxonomy. Install it with:

pip install datashield-ai[taxonomy]

If you don't need taxonomy integration, this error means you're calling datashield_finding_to_taxonomy() directly — the CLI commands work without it.

Presidio not found

Presidio-based PII detection requires:

pip install datashield-ai[presidio]

This downloads models (~500 MB) on first use.

Large files causing memory errors

DataShield loads the entire file into memory. For files >500 MB, set:

export DATASHIELD_MAX_SIZE_MB=1000

Or pass --max-size if available. Streaming support is planned for a future release.

How to interpret risk scores

Score	Category	Meaning
70+	Critical	High-risk data (credentials, keys) present
40-69	High	PII/secrets detected, immediate action needed
20-39	Medium	Some sensitive fields found
5-19	Low	Minor privacy concerns
<5	Safe	No significant risks detected

Getting help

datashield --help          # Top-level help
datashield scan --help     # Help for a specific command
datashield --version       # Show version

License

MIT — see LICENSE

Author

Carlos-Projects — Carlos@AIAgentObservatory.org