Implementing Cloud Dlp For Data Protection

Implementing Cloud DLP for Data Protection

When to Use

• When compliance frameworks (GDPR, HIPAA, PCI DSS) require automated sensitive data discovery and protection
• When building data governance programs that classify and label data across cloud storage
• When implementing data loss prevention controls for cloud-based data pipelines
• When auditing cloud environments for unprotected sensitive data (PII, PHI, financial data)
• When integrating DLP scanning into CI/CD pipelines to prevent sensitive data from reaching production

Do not use for endpoint DLP (use Microsoft Purview or Symantec DLP agents), for email DLP (use Microsoft 365 DLP or Google Workspace DLP), or for network-level data exfiltration prevention (use VPC endpoint policies and network firewalls).

Common Misconfigurations & Verification

• min_likelihood set too strict: GCP DLP defaulting to VERY_LIKELY (or Macie thresholds set high) drops real PII. LIKELY is the practical floor for SSN/credit-card detection — anything stricter under-reports.
• Inspection configured but de-identification never applied: an inspectJob/Macie scan only finds data. Mask/tokenize/redact transforms run in a separate deidentifyContent/InfoTypeTransformations call. Verify the pipeline actually invokes the transform, not just inspection.
• DLP job results discarded: without an actions.saveFindings.outputConfig (GCP) or a classification-export-configuration (Macie), findings vanish after the job. Confirm gcloud dlp jobs describe JOB_ID shows a save action.
• Macie scope excludes the real formats: scoping to specific OBJECT_EXTENSION values that omit the actual files means a "clean" report that scanned nothing.
• KMS-wrapped crypto key not granted: FPE/tokenization fails if the DLP service agent lacks cloudkms.cryptoKeyVersions.useToDecrypt on the wrapping key.

aws macie2 list-classification-jobs --query 'items[].[name,jobStatus]'
gcloud dlp jobs list --filter="state=DONE" --format="table(name,inspectDetails.result.processedBytes)"
gcloud dlp jobs describe JOB_ID --format="value(inspectDetails.result.infoTypeStats)"

Prerequisites

• Amazon Macie enabled with appropriate S3 bucket permissions
• Google Cloud DLP API enabled (gcloud services enable dlp.googleapis.com)
• Azure Information Protection or Microsoft Purview configured
• IAM permissions for DLP service administration and data access
• Knowledge of data sensitivity categories relevant to the organization (PII, PHI, PCI, proprietary)

Workflow

Step 1: Deploy Amazon Macie for S3 Data Discovery

Enable Macie and configure automated sensitive data discovery jobs for S3 buckets.

# Enable Amazon Macie
aws macie2 enable-macie

# List all S3 buckets Macie can scan
aws macie2 describe-buckets \
  --query 'buckets[*].[bucketName,classifiableSizeInBytes,unclassifiableObjectCount.total]' \
  --output table

# Create a classification job for specific buckets
aws macie2 create-classification-job \
  --job-type SCHEDULED \
  --name "weekly-pii-scan" \
  --schedule-frequency-details '{"weekly":{"dayOfWeek":"MONDAY"}}' \
  --s3-job-definition '{
    "bucketDefinitions": [{
      "accountId": "ACCOUNT_ID",
      "buckets": ["customer-data-bucket", "analytics-data-lake", "backup-bucket"]
    }],
    "scoping": {
      "includes": {
        "and": [{
          "simpleScopeTerm": {
            "key": "OBJECT_EXTENSION",
            "values": ["csv", "json", "parquet", "txt", "xlsx"],
            "comparator": "EQ"
          }
        }]
      }
    }
  }' \
  --managed-data-identifier-ids '["SSN","CREDIT_CARD_NUMBER","EMAIL_ADDRESS","AWS_CREDENTIALS","PHONE_NUMBER"]'

# Create custom data identifier for internal employee IDs
aws macie2 create-custom-data-identifier \
  --name "EmployeeID" \
  --regex "EMP-[0-9]{6}" \
  --description "Internal employee ID format"

# Check job status and results
aws macie2 list-classification-jobs \
  --query 'items[*].[name,jobStatus,statistics.approximateNumberOfObjectsToProcess]' \
  --output table

Step 2: Configure Google Cloud DLP API for Data Inspection

Use Google Cloud DLP to inspect and de-identify sensitive data across GCP resources.

# Inspect a Cloud Storage bucket for sensitive data
gcloud dlp inspect-content \
  --content-type=TEXT_PLAIN \
  --min-likelihood=LIKELY \
  --info-types=PHONE_NUMBER,EMAIL_ADDRESS,CREDIT_CARD_NUMBER,US_SOCIAL_SECURITY_NUMBER \
  --storage-type=CLOUD_STORAGE \
  --gcs-uri="gs://sensitive-data-bucket/data/*.csv"

# Create an inspection job for BigQuery
cat > dlp-job.json << 'EOF'
{
  "inspectJob": {
    "storageConfig": {
      "bigQueryOptions": {
        "tableReference": {
          "projectId": "PROJECT_ID",
          "datasetId": "customer_data",
          "tableId": "transactions"
        },
        "sampleMethod": "RANDOM_START",
        "rowsLimit": 10000
      }
    },
    "inspectConfig": {
      "infoTypes": [
        {"name": "CREDIT_CARD_NUMBER"},
        {"name": "US_SOCIAL_SECURITY_NUMBER"},
        {"name": "EMAIL_ADDRESS"},
        {"name": "PHONE_NUMBER"},
        {"name": "PERSON_NAME"}
      ],
      "minLikelihood": "LIKELY",
      "limits": {"maxFindingsPerRequest": 1000}
    },
    "actions": [{
      "saveFindings": {
        "outputConfig": {
          "table": {
            "projectId": "PROJECT_ID",
            "datasetId": "dlp_results",
            "tableId": "findings"
          }
        }
      }
    }]
  }
}
EOF

gcloud dlp jobs create --project=PROJECT_ID --body-from-file=dlp-job.json

Step 3: Implement Data De-identification with Cloud DLP

Configure de-identification transforms to mask, tokenize, or redact sensitive data.

# deidentify_pipeline.py - De-identify sensitive data using Google Cloud DLP
from google.cloud import dlp_v2

def deidentify_data(project_id, text):
    """De-identify PII in text using Cloud DLP."""
    client = dlp_v2.DlpServiceClient()

    inspect_config = {
        "info_types": [
            {"name": "EMAIL_ADDRESS"},
            {"name": "PHONE_NUMBER"},
            {"name": "CREDIT_CARD_NUMBER"},
            {"name": "US_SOCIAL_SECURITY_NUMBER"},
        ],
        "min_likelihood": dlp_v2.Likelihood.LIKELY,
    }

    deidentify_config = {
        "info_type_transformations": {
            "transformations": [
                {
                    "info_types": [{"name": "EMAIL_ADDRESS"}],
                    "primitive_transformation": {
                        "character_mask_config": {
                            "masking_character": "*",
                            "number_to_mask": 0,
                            "characters_to_ignore": [
                                {"common_characters_to_ignore": "PUNCTUATION"}
                            ],
                        }
                    },
                },
                {
                    "info_types": [{"name": "CREDIT_CARD_NUMBER"}],
                    "primitive_transformation": {
                        "crypto_replace_ffx_fpe_config": {
                            "crypto_key": {
                                "kms_wrapped": {
                                    "wrapped_key": "WRAPPED_KEY_BASE64",
                                    "crypto_key_name": "projects/PROJECT/locations/global/keyRings/dlp/cryptoKeys/tokenization",
                                }
                            },
                            "common_alphabet": "NUMERIC",
                        }
                    },
                },
                {
                    "info_types": [{"name": "US_SOCIAL_SECURITY_NUMBER"}],
                    "primitive_transformation": {
                        "redact_config": {}
                    },
                },
            ]
        }
    }

    item = {"value": text}
    parent = f"projects/{project_id}/locations/global"

    response = client.deidentify_content(
        request={
            "parent": parent,
            "deidentify_config": deidentify_config,
            "inspect_config": inspect_config,
            "item": item,
        }
    )
    return response.item.value

Step 4: Configure Azure Information Protection

Set up sensitivity labels and DLP policies in Microsoft Purview for Azure resources.

# Connect to Microsoft Purview compliance
Connect-IPPSSession

# Create sensitivity labels
New-Label -DisplayName "Confidential - PII" \
  -Name "Confidential-PII" \
  -Tooltip "Contains personally identifiable information" \
  -ContentType "File, Email"

New-Label -DisplayName "Highly Confidential - Financial" \
  -Name "HighlyConfidential-Financial" \
  -Tooltip "Contains financial data subject to PCI DSS" \
  -ContentType "File, Email"

# Create auto-labeling policy for Azure Storage
New-AutoSensitivityLabelPolicy -Name "Auto-Label-PII" \
  -ExchangeLocation All \
  -SharePointLocation All \
  -OneDriveLocation All \
  -Mode Enable

New-AutoSensitivityLabelRule -Policy "Auto-Label-PII" \
  -Name "Detect-SSN" \
  -ContentContainsSensitiveInformation @{
    Name = "U.S. Social Security Number (SSN)";
    MinCount = 1;
    MinConfidence = 85
  } \
  -ApplySensitivityLabel "Confidential-PII"

# Azure: Configure DLP policy for Storage accounts
az security assessment create \
  --name "storage-sensitive-data" \
  --assessed-resource-type "Microsoft.Storage/storageAccounts"

# Enable Microsoft Defender for Storage with sensitive data threat detection
az security pricing create --name StorageAccounts --tier standard \
  --subplan DefenderForStorageV2 \
  --extensions '[{"name":"SensitiveDataDiscovery","isEnabled":"True"}]'

Step 5: Integrate DLP into Data Pipelines

Add DLP scanning to ETL and data pipeline workflows to prevent sensitive data leakage.

# pipeline_dlp_gate.py - DLP gate for data pipelines
import boto3
import json

macie_client = boto3.client('macie2')
s3_client = boto3.client('s3')

def scan_pipeline_output(bucket, prefix):
    """Scan pipeline output data for sensitive content before promotion."""
    job_response = macie_client.create_classification_job(
        jobType='ONE_TIME',
        name=f'pipeline-scan-{prefix}',
        s3JobDefinition={
            'bucketDefinitions': [{
                'accountId': boto3.client('sts').get_caller_identity()['Account'],
                'buckets': [bucket]
            }],
            'scoping': {
                'includes': {
                    'and': [{
                        'simpleScopeTerm': {
                            'key': 'OBJECT_KEY',
                            'comparator': 'STARTS_WITH',
                            'values': [prefix]
                        }
                    }]
                }
            }
        },
        managedDataIdentifierSelector='ALL'
    )

    return job_response['jobId']

def check_scan_results(job_id):
    """Check if DLP scan found sensitive data."""
    response = macie_client.list_findings(
        findingCriteria={
            'criterion': {
                'classificationDetails.jobId': {'eq': [job_id]},
                'severity.description': {'eq': ['High', 'Critical']}
            }
        }
    )
    return len(response.get('findingIds', [])) > 0

def gate_decision(bucket, prefix):
    """DLP gate: block pipeline if sensitive data found."""
    job_id = scan_pipeline_output(bucket, prefix)
    has_sensitive_data = check_scan_results(job_id)

    if has_sensitive_data:
        return {
            'decision': 'BLOCK',
            'reason': 'Sensitive data detected in pipeline output',
            'action': 'Apply de-identification before promoting to production'
        }
    return {'decision': 'ALLOW', 'reason': 'No sensitive data detected'}

Step 6: Monitor DLP Findings and Generate Reports

Aggregate DLP findings across cloud providers and generate compliance reports.

# Macie: Get finding statistics
aws macie2 get-finding-statistics \
  --group-by "severity.description" \
  --finding-criteria '{"criterion":{"category":{"eq":["CLASSIFICATION"]}}}'

# Macie: List findings by sensitivity type
aws macie2 list-findings \
  --finding-criteria '{
    "criterion": {
      "classificationDetails.result.sensitiveData.category": {"eq": ["PERSONAL_INFORMATION"]},
      "severity.description": {"eq": ["High"]}
    }
  }' \
  --sort-criteria '{"attributeName": "updatedAt", "orderBy": "DESC"}'

# GCP DLP: List job results
gcloud dlp jobs list --project=PROJECT_ID --filter="state=DONE" \
  --format="table(name, createTime, inspectDetails.result.processedBytes, inspectDetails.result.totalEstimatedTransformations)"

# Export Macie findings to S3 for compliance reporting
aws macie2 create-findings-report \
  --finding-criteria '{"criterion":{"category":{"eq":["CLASSIFICATION"]}}}' \
  --sort-criteria '{"attributeName":"severity.score","orderBy":"DESC"}'

Key Concepts

Tools & Systems

• Amazon Macie: ML-powered sensitive data discovery and classification for S3 with automated finding generation
• Google Cloud DLP API: Programmable API for inspecting, classifying, de-identifying, and redacting sensitive data
• Microsoft Purview: Data governance platform with sensitivity labeling, auto-classification, and DLP policy enforcement
• Azure Information Protection: Data classification and labeling service integrated with Microsoft 365 and Azure storage
• Nightfall AI: Third-party cloud DLP tool supporting scanning across SaaS applications and cloud infrastructure

Common Scenarios

Scenario: Discovering PII in an Unprotected S3 Data Lake

Context: A compliance audit reveals that the analytics team's S3 data lake contains customer PII (names, emails, SSNs) in CSV files without encryption or access controls. The organization must classify all data and implement DLP controls.

Approach:

1. Enable Macie and create a one-time classification job against the data lake bucket
2. Review Macie findings to identify which objects contain PII and what types
3. Create custom data identifiers for organization-specific formats (employee IDs, account numbers)
4. Implement a weekly scheduled Macie job for ongoing discovery
5. Build a data pipeline gate that scans new data before promotion to the data lake
6. Apply de-identification transforms (masking SSNs, tokenizing emails) for analytics use cases
7. Configure S3 bucket policies to restrict access to classified data to authorized roles only

Pitfalls: Macie charges per GB scanned. Large data lakes can generate significant costs. Use scoping rules to focus on high-risk object types (CSV, JSON, Parquet) and exclude known-safe formats (compressed archives, binary files). De-identification must preserve data utility for analytics while removing re-identification risk.

Output Format

Cloud DLP Compliance Report
==============================
Organization: Acme Corp
Scan Period: 2026-02-01 to 2026-02-23
Environments: AWS (12 buckets), GCP (3 datasets), Azure (5 storage accounts)

DATA DISCOVERY SUMMARY:
  Total objects/records scanned:    2,847,000
  Objects with sensitive data:        45,200 (1.6%)
  Unique sensitivity categories:      8

SENSITIVE DATA FINDINGS:
  PII (names, emails, phone):       23,400 objects
  Financial (credit cards, bank):     8,700 objects
  Health (PHI, medical records):      3,200 objects
  Credentials (API keys, tokens):     1,400 objects
  Government ID (SSN, passport):      5,800 objects
  Custom (employee ID, account):      2,700 objects

FINDINGS BY SEVERITY:
  Critical:    1,400 (exposed credentials)
  High:       14,200 (unprotected PII/PHI)
  Medium:     18,600 (standard PII)
  Low:        11,000 (non-sensitive patterns)

PROTECTION STATUS:
  Data with encryption at rest:       78%
  Data with access controls:          65%
  Data with sensitivity labels:       12%
  Pipeline data with DLP gates:       30%

REMEDIATION ACTIONS:
  Objects quarantined:                1,400
  De-identification applied:          8,200
  Access controls tightened:         14,200
  Sensitivity labels applied:        45,200