AI Privacy Resources
The primary sources this site works from: EU and California regulation, CJEU rulings, regulator guidance, privacy research, and tooling.
Regulation and Case Law
- Regulation (EU) 2024/1689, the EU AI Act
The consolidated text. Article 50 carries the transparency duties; Articles 51, 53 and 55 carry the general-purpose AI model provider obligations.
European Union #regulation - GDPR Article 22: automated individual decision-makingEuropean Union #regulation
- GDPR Article 35: data protection impact assessmentEuropean Union #regulation
- GDPR Article 17: right to erasureEuropean Union #regulation
- GDPR Article 4: definitions, including pseudonymisationEuropean Union #regulation
- CCPA updates, risk assessments, cybersecurity audits and ADMT regulations
The finalized California package that attaches pre-use notice, opt-out, access and appeal duties to automated decisionmaking technology.
California Privacy Protection Agency #regulation - EDPS v SRB, Case C-413/23 P (press release)
Whether pseudonymised data is personal data is assessed from the position of the recipient, not in the abstract.
Court of Justice of the European Union, 2025 #case-law - Schufa, Case C-634/21
A score that determinatively shapes a third party's decision is itself the automated decision under Article 22.
Court of Justice of the European Union, 2023 #case-law - Schrems II, Case C-311/18Court of Justice of the European Union, 2020 #case-law
Regulator Guidance
- EDPB Opinion 28/2024 on AI models and personal data
A trained model is not automatically anonymous, and legitimate interest is available only through the three-step assessment.
European Data Protection Board #guidance - EDPB Guidelines 01/2025 on Pseudonymisation
The pseudonymisation-domain framing. Consultation version, not a final text.
European Data Protection Board #guidance - EDPB Recommendations 01/2020 on supplementary measures
The transfer impact assessment and supplementary-measures framework for transfers relying on Article 46.
European Data Protection Board #guidance - EDPB Guidelines on Automated Decision-Making and ProfilingEuropean Data Protection Board #guidance
- DPIA Guidelines (WP 248 rev.01)
The nine-criteria test that operationalises the Article 35(3) DPIA triggers.
Article 29 Working Party #guidance - Legitimate interest: web scraping focus sheet
The mitigations a member-state regulator expects when training data is collected by scraping.
CNIL, 2025 #guidance - Rights related to automated decision-making including profilingInformation Commissioner's Office #guidance
- Standard Contractual ClausesEuropean Commission #guidance
Foundational Privacy Research
- Membership Inference Attacks against Machine Learning Models
The shadow-model attack that made training-set membership a measurable property of a deployed model.
Shokri et al., 2017 #research - Membership Inference Attacks From First Principles (LiRA)
Why average attack accuracy is the wrong yardstick, and what true-positive rate at low false-positive rate shows instead.
Carlini et al., 2021 #research - Extracting Training Data from Large Language ModelsCarlini et al., 2021 #research
- Privacy Risks of General-Purpose Language ModelsPan et al., 2020 #research
- Synthetic Data: Anonymisation Groundhog Day
The reference study for why synthetic output is not anonymous by default.
Stadler, Oprisanu, Troncoso #research - Deep Leakage from GradientsZhu, Liu, Han #research
- Communication-Efficient Learning of Deep Networks from Decentralized DataMcMahan et al., 2016 #research
- Practical Secure Aggregation for Federated Learning on User-Held DataBonawitz et al. #research
- Models in the dark: rectification and erasure under GDPR in ML supply chains
Why erasure and rectification have no clean technical answer once data is in the weights, and how the problem propagates downstream.
#research
Standards and Frameworks
- NIST SP 800-188: De-Identifying Government Datasets
The federal de-identification guidance: risk-based assessment, Disclosure Review Boards, and re-identification studies after transformation.
NIST, 2023 #framework - Adversarial Machine Learning: A Taxonomy and Terminology (AI 100-2e2025)
Places membership inference and the rest of the privacy-attack class inside a reference taxonomy.
NIST, 2025 #framework - NIST AI Risk Management FrameworkNIST #framework
- UK-US Blog Series on Privacy-Preserving Federated LearningNIST #framework
Privacy Engineering Tools
- Microsoft Presidio: PII detection and de-identification
Detection for unstructured text and images. Probabilistic by design, and the documentation says so.
Microsoft #tool - ARX Data Anonymization Tool
k-anonymity, l-diversity, t-closeness and differential privacy on tabular data, with re-identification risk and utility metrics.
#tool - AmnesiaATHENA Research Center and OpenAIRE #tool
- GreenmaskGreenmaskIO #tool
- SDV: Synthetic Data VaultDataCebo #tool
- MOSTLY AI Synthetic Data SDK
Apache-2.0, runs locally, and exposes differential privacy as a training option.
MOSTLY AI #tool