A comprehensive collection of resources for PII detection, anonymization, privacy-preserving techniques, and GDPR compliance in Large Language Model applications.
Adaptive PII Mitigation Framework for Large Language Models (2025)
- Context-aware PII detection with regulation-specific remediation strategies (GDPR, CCPA, PIPEDA)
- Achieves F1 score of 0.95 for passport numbers, outperforming Microsoft Presidio and Amazon Comprehend
- arXiv Paper
PIIvot: A Lightweight NLP Anonymization Framework for Question-Anchored Tutoring Dialogues (2024)
- Novel approach reframing PII detection as potential-PII labeling with LLM-generated contextual replacements
- Preserves data integrity in educational settings while ensuring privacy
- arXiv Paper
The Text Anonymization Benchmark (TAB): A Dedicated Corpus and Evaluation Framework (2022)
- Seminal benchmark comprising 1,268 ECHR court cases with comprehensive PII annotations
- Introduces novel evaluation metrics for privacy protection and utility preservation
- Paper | Dataset
Can Large Language Models Really Recognize Your Name? (2024)
- Demonstrates LLM recall drops 20-40% for ambiguous names vs recognizable ones
- Creates AmBench benchmark showing LLMs confuse human names with non-human entities
- arXiv Paper
Unlocking the Potential of Large Language Models for Clinical Text Anonymization (2024)
- Proposes six evaluation metrics for generative anonymization with LLMs in clinical settings
- Establishes LLM-based models as reliable alternatives to traditional approaches
- Paper
Bootstrapping Text Anonymization Models with Distant Supervision (2022)
- Uses knowledge graphs and distant supervision to automatically generate training data
- Eliminates manual annotation while following privacy-first k-anonymity principles
- Paper
Large Language Models are Advanced Anonymizers (2024)
- Demonstrates LLMs' superior anonymization capabilities compared to commercial tools
- Introduces adversarial anonymization framework with privacy-utility trade-offs
- arXiv Paper
Fine-tuning LLMs with user-level differential privacy (2024)
- Google Research's technical implementation of user-level differential privacy for LLM fine-tuning
- Novel optimization techniques including Example-Level and User-Level Sampling
- Google Research Blog
VaultGemma: The world's most capable differentially private LLM (2024)
- Technical deep-dive into training LLMs with differential privacy from scratch
- Includes scaling laws and compute-privacy-utility trade-offs
- Google Research Blog
Differential Privacy in Natural Language Processing: The Story So Far (2022)
- Comprehensive survey of DP vulnerabilities and methodologies in NLP
- Essential reading for understanding DP applications to text data
- arXiv Paper
Does Differential Privacy Impact Bias in Pretrained NLP Models? (2024)
- First study examining how DP training affects bias in LLMs
- Shows DP can increase model bias against protected groups
- arXiv Paper
FedNLP: Benchmarking Federated Learning Methods for Natural Language Processing Tasks (2022)
- First comprehensive benchmarking framework for federated learning in NLP
- Universal interface between Transformer models and FL methods
- Paper
Privacy-Preserving Instructions for Aligning Large Language Models (2024)
- First work analyzing privacy risks in LLM alignment annotation process
- Studies over 12,000 real-world user instructions for sensitive content
- arXiv Paper
Do Membership Inference Attacks Work on Large Language Models? (2024)
- Large-scale evaluation showing MIAs on LLMs barely outperform random guessing
- Analyzes 160M to 12B parameter models with comprehensive attack evaluation
- arXiv Paper
User Inference Attacks on Large Language Models (2023)
- Introduces efficient user inference attack scaling to LLMs
- Demonstrates user-level privacy risks in fine-tuning scenarios
- arXiv Paper
Microsoft Presidio
- Most comprehensive open-source framework for detecting, redacting, and anonymizing sensitive data
- Context-aware detection using spaCy, 50+ predefined entity types, multi-language support
- GitHub | Documentation
DataFog Python
- Lightning-fast PII detection with 190x performance advantage (2.4ms for 10KB text)
- Multiple engines: Regex, spaCy, GLiNER with smart cascading
- GitHub
PII Masker (HydroXai)
- Advanced AI-powered detection using fine-tuned DeBERTa-v3 with 1024 token support
- High-precision detection with ~4% performance improvement over base models
- GitHub
PII Codex
- Research-focused package for PII detection, categorization, and severity assessment
- Built on Microsoft Presidio with NIST/DHS categorizations and risk scoring
- GitHub
PySyft
- Comprehensive library for secure, private Deep Learning using federated learning and differential privacy
- Multi-Party Computation (MPC), homomorphic encryption, remote data science
- GitHub
TensorFlow Privacy
- Google's library for training ML models with differential privacy
- DP-SGD optimizers, privacy accounting, membership inference attack defenses
- GitHub
OpenDP Library & SmartNoise
- Microsoft-Harvard partnership providing comprehensive differential privacy toolkit
- SmartNoise SQL for DP queries, synthetic data generation, Rust core with Python bindings
- GitHub Core | SDK
Opacus (PyTorch Differential Privacy)
- Meta AI's library for training PyTorch models with differential privacy
- Drop-in replacement for PyTorch optimizers with minimal code changes
- GitHub
IBM Differential Privacy Library (diffprivlib)
- General-purpose differential privacy library with scikit-learn compatible API
- Multiple DP mechanisms with easy integration for existing workflows
- GitHub
Faker
- Most popular library for generating fake data with 150+ providers
- Localization support for 50+ countries, extensible provider system
- GitHub
SDV (Synthetic Data Vault)
- Comprehensive synthetic data generation ecosystem for single-table, multi-table, time-series
- Privacy controls, evaluation metrics, production-ready implementations
- GitHub
Gretel Synthetics
- Neural network-based synthetic data generation with differential privacy support
- High utility synthetic data with privacy guarantees for production use
- GitHub
iiiorg/piiranha-v1-detect-personal-information
- State-of-the-art multilingual PII detection (99.44% accuracy) across 17 PII types and 6 languages
- Fine-tuned microsoft/mdeberta-v3-base with exceptional precision (98.48%) and recall (98.27%)
- HuggingFace Model
beki/flair-pii-distilbert
- 5-class NER model trained on protocol trace data (F1-Score: 0.9522)
- Specialized for detecting PER, LOC, ORG, DATE_TIME, NRP entities in network traces
- HuggingFace Model
betterdataai/PII_DETECTION_MODEL
- Lightweight multilingual PII detection covering 29 classes across 7 languages
- Built on Qwen2-0.5B for low latency with long document support
- HuggingFace Model
Text Anonymization Benchmark (TAB)
- Gold standard benchmark with European Court Documents and multiple annotator labels
- Comprehensive framework for evaluating anonymization techniques
- HuggingFace Dataset
beki/privy
- Synthetic PII dataset with 60+ PII types in protocol trace formats (JSON, SQL, HTML, XML)
- Generated using Privy tool from OpenAPI specifications for comprehensive training
- HuggingFace Dataset
allenai/WildChat-4.8M
- Large conversation dataset with PII detection and anonymization features
- Includes redaction indicators and anonymized content for conversational AI training
- HuggingFace Dataset
beki/pii-anonymizer
- Interactive Streamlit app using Presidio framework with custom PII models
- Real-time PII detection and anonymization demonstration
- HuggingFace Space
haakohu/deep_privacy2
- Advanced computer vision anonymization for faces and full-body anonymization
- Comprehensive human anonymization using multiple detection networks
- HuggingFace Space
Microsoft Presidio Framework Tutorial
- Comprehensive guide for building PII detection systems with Python API documentation
- Custom recognizer creation, multi-language support, context-aware enhancement
- Documentation
Building a Customized PII Anonymizer with Microsoft Presidio
- Step-by-step tutorial for implementing custom multi-language PII anonymizer
- Practical code examples with Norwegian National ID detection case study
- Towards Data Science
Detecting and Masking PII with Piiranha
- Tutorial for using state-of-the-art Piiranha model with 99.44% accuracy
- Python implementation examples covering 17 PII entity types
- Medium Tutorial
Data Anonymization Techniques for Secure LLM Utilization
- Comprehensive guide covering k-anonymity, differential privacy, synthetic data generation
- Specific techniques for LLM deployments with GDPR compliance considerations
- Protecto AI Guide
Using Local LLMs to Anonymize Data for Remote LLMs
- Innovative approach using smaller local LLMs (llama 3.1 8B) for pre-anonymization
- Detailed prompt engineering for entity replacement and de-anonymization workflows
- Medium Guide
Google Cloud PII De-identification Pipeline
- Enterprise-grade implementation using Google Cloud DLP for automated data transformation
- Dataflow pipelines, format-preserving encryption, streaming data processing
- Google Cloud Architecture
Azure AI PII Detection Service
- Official documentation for Azure's cloud-based PII detection with API references
- Conversation PII detection, multi-language support, batch processing capabilities
- Microsoft Learn
AWS PII Detection with Comprehend and Macie
- HIPAA-compliant PII/PHI detection using AWS services with S3 Object Lambda integration
- Healthcare data protection patterns and compliance best practices
- AWS Industries Blog
Real-World LLM Project Anonymization
- AI-powered document processing system achieving 95% processing time reduction
- Modular anonymization architecture with AWS deployment and business impact metrics
- The Software House Case Study
Medical Text Anonymization with Local LLMs
- Comprehensive anonymization tool for medical text using locally deployed LLMs
- HIPAA vs GDPR compliance with quantitative evaluation and precision metrics
- NEJM AI Paper
European Data Protection Board (EDPB) Opinion 28/2024 on AI Models
- Comprehensive 35-page guidance on when AI models can be considered anonymous
- Three-step anonymity test and mitigating measures for GDPR compliance
- Official Document
UK ICO AI and Data Protection Guidance
- Updated guidance restructured around GDPR principles for AI systems
- AI Risk Toolkit for practical privacy risk assessment during development
- ICO Guidance
CNIL AI Development Recommendations (2025)
- Latest French recommendations on AI system development and individual rights facilitation
- PANAME project for privacy auditing of AI models with sector-specific guidance
- CNIL Guidance
ISO/IEC 29100:2024 - Privacy Framework
- Updated international privacy framework with common terminology and principles
- Technology-neutral approach applicable to all AI systems processing PII
- ISO Standard
ISO/IEC 27559:2022 - Privacy Enhancing Data De-identification
- Framework for identifying and mitigating re-identification risks in de-identified data
- Critical for AI model training data anonymization and attack prevention
- ISO Standard
NIST AI Risk Management Framework (AI RMF 1.0)
- Comprehensive framework with four core functions: GOVERN, MAP, MEASURE, MANAGE
- Seven trustworthy AI characteristics including "privacy enhanced" considerations
- NIST Framework
Google Secure AI Framework (SAIF)
- Six core elements for securing AI systems with integrated privacy protections
- Security-by-default approach for ML and generative AI applications
- Google SAIF
Privacy by Design Seven Foundational Principles
- Essential framework for embedding privacy into AI system architecture from design phase
- Proactive approach with privacy as default setting throughout development lifecycle
DPIA Templates for AI Systems
- EDPB, ICO, and CNIL templates specifically adapted for AI system risk assessment
- Step-by-step processes for identifying and mitigating privacy risks in LLM development
Privacy Meter - ML Privacy Auditing
- Open-source library for quantitative privacy risk assessment through inference attacks
- Membership inference, range membership inference, and attribute inference testing
- GitHub
OWASP AI Security and Privacy Guide
- Practical privacy engineering guidance throughout the AI/ML system lifecycle
- Data minimization principles and privacy-enhancing technology implementations
- OWASP Project
GDPR Article 25 Implementation for AI
- Data protection by design and by default requirements with technical measures
- Lifecycle approach to privacy protection in AI system development
Legitimate Interest Assessment for AI Training
- Three-step framework for using legitimate interest as legal basis for LLM training
- AI-specific considerations and mitigating measures per EDPB guidelines
For organizations new to privacy-preserving LLM development, we recommend this implementation order:
- Start with PII Detection: Deploy Microsoft Presidio or DataFog for immediate PII identification
- Implement Privacy by Design: Follow NIST AI RMF and EDPB guidelines for system architecture
- Conduct DPIAs: Use regulatory templates to assess privacy risks in your AI development
- Explore Differential Privacy: Begin with TensorFlow Privacy or Opacus for training privacy
- Consider Synthetic Data: Evaluate SDV or Gretel Synthetics for privacy-safe training data
This awesome list focuses on practical, implementable resources alongside theoretical foundations. We prioritize recent developments (2022-2024) while including seminal works that remain relevant. Resources are selected based on:
- Practical utility for developers and researchers
- Technical rigor and peer review quality
- Active maintenance and community support
- GDPR compliance relevance and regulatory alignment
- Open source availability when possible
This resource collection is maintained for educational and research purposes. Always consult with legal experts for specific GDPR compliance requirements in your jurisdiction and use case.
