Data Privacy in AI/ML: From Unlearning to Encrypted Learning and Beyond
Latest 7 papers on data privacy: Aug. 1, 2026
Data privacy is no longer just a buzzword; it’s a critical cornerstone of trust and compliance in the rapidly evolving landscape of AI and Machine Learning. As models become more powerful and data more ubiquitous, ensuring sensitive information remains protected is paramount. Recent research showcases exciting breakthroughs, tackling privacy challenges across diverse domains, from robust information extraction to the very fabric of how we train and deploy AI systems.
The Big Idea(s) & Core Innovations
The overarching theme across these innovations is the pursuit of stronger privacy guarantees without sacrificing model utility or efficiency. One significant leap comes from Sony Research India with their paper, “Obliviate: Efficient Unlearning in Recommender Systems”. They tackle the ‘right to be forgotten’ by introducing a two-stage unlearning framework that efficiently removes the influence of deleted user data. Their key insight? Deletion effects in recommenders are largely low-dimensional, enabling efficient unlearning via a Low-Rank Unlearning Adapter (LUA) and a Locality-Aware Calibration (LAC) stage, achieving up to 9% utility gains and being 3x faster than full retraining.
Shifting gears to information extraction, researchers from the Ruhr West University of Applied Sciences introduce a novel method in “Enhancing Generative Information Extraction with Two-step Validation: A Product Attribute Use Case”. Their two-step validation method, where Large Language Models (LLMs) correct Pretrained Language Model (PLM) predictions rather than directly extracting entities, significantly boosts performance, especially for ‘weakly expressed’ entities. This reformulation allows smaller, open-source LLMs to achieve performance comparable to larger models, enabling local deployment and addressing data privacy concerns for sensitive product information.
Meanwhile, safeguarding visual data is the focus of a comprehensive survey by ELLIS Institute Finland and the University of Oulu. Their paper, “Face De-Identification: A Domain-Centric Survey from Capture to Processing”, provides the first unified taxonomy of face de-identification techniques across physical, sensor, and digital domains. A key insight is the increasing traction of multi-domain approaches, as digital methods, while dominant, benefit from stronger trust-boundary guarantees offered by physical and sensor-level interventions.
For more fundamental privacy guarantees, University of Maryland Baltimore County explores a sophisticated approach in “A Maximum Entropy Implementation of Differential Privacy Under Linear Invariants”. They address the complex challenge of applying differential privacy (DP) while maintaining specific aggregate constraints (like population totals in Census data). Their innovation lies in generating correlated Gaussian or Laplace noise that satisfies linear aggregation constraints with high probability, providing formal DP guarantees in the invariant data universe. This is a partial solution to an open question about the null space of correlation matrices.
Finally, the Koç University and Bilkent University present CURE in “CURE: Privacy-Preserving Split Learning Done Right”, a groundbreaking privacy-preserving split learning system. Instead of encrypting all data and models, CURE leverages homomorphic encryption (CKKS scheme) to encrypt only the server-side model parameters. This elegant solution protects label confidentiality, achieving up to 210x speedup compared to state-of-the-art privacy-preserving alternatives while maintaining model accuracy.
Under the Hood: Models, Datasets, & Benchmarks
These advancements are built upon and contribute to a rich ecosystem of tools and resources:
- Obliviate was validated on widely used recommender system benchmarks like MovieLens-1M, Amazon Fashion, and Yelp 2018 datasets, applicable to models like MF-BPR and LightGCN. While no code is yet publicly available, the paper demonstrates robust performance across these standards.
- The two-step validation for generative IE leverages PLMs fine-tuned on datasets like Amazon Product Descriptions and Kaggle’s E-commerce Text Classification, with a demo application and code available at https://github.com/doyouwantsometea/pie_paper and https://github.com/hrw-neurolab/transferhub_pie_demo.
- The Face De-Identification survey is a comprehensive resource, identifying common datasets and metrics, and pointing to an ‘Awesome-FaceDe-ID’ GitHub repository (https://github.com/CV-AC/Awesome-FaceDe-ID) for further exploration.
- For differential privacy with linear invariants, the methodology utilizes advanced mathematical constructs like multivariate normal distributions and algorithms like POCS (Projection Onto Convex Sets) and projected gradient methods, relevant for Census-like datasets.
- CURE developed an open-source API implementation for HE operations in deep learning networks, tested with ResNet building blocks, and made publicly available at https://github.com/CRYPTO-KU/CURE-Privacy-Preserving-Split-Learning.
Separately, a scoping review from Temple University, “Natural Language Processing in Health Professions Education: A Scoping Review”, highlights a critical need for addressing data privacy concerns in health education AI, pointing to LoRA techniques as a way to enable resource-conscious, localized LLM deployment.
In the realm of Federated Learning, Shanghai Jiao Tong University and collaborators introduce “FedCod: An Efficient Communication Protocol for Cross-Silo Federated Learning with Coding”. This protocol uses adaptive coding redundancy and client-to-client communication to reduce communication time by up to 62% in geo-distributed settings. It’s an application-layer approach, integrated into frameworks like FedLab (https://github.com/SJTU-IPADS/FedLab), and tested across global AWS and Azure regions, effectively making FL more robust to network heterogeneity.
Impact & The Road Ahead
These advancements represent a significant stride towards building more trustworthy and compliant AI systems. Obliviate brings the ‘right to be forgotten’ closer to practical reality for recommender systems, crucial for GDPR compliance. The two-step validation for generative IE paves the way for secure, local deployment of advanced NLP, particularly vital for SMEs handling sensitive data. CURE’s efficient privacy-preserving split learning could revolutionize collaborative AI in sensitive sectors like healthcare and genomics, enabling model training without ever exposing raw labels.
The domain-centric view of face de-identification underscores the need for multi-modal privacy strategies, moving beyond just digital manipulation to interventions at the point of capture. Meanwhile, the maximum entropy DP approach offers a rigorous method for releasing privacy-preserving statistics while preserving crucial aggregate invariants, a cornerstone for public trust in governmental data releases.
The future will likely see further convergence of these techniques: unlearning mechanisms integrated into federated learning, homomorphic encryption becoming more pervasive in data pipelines, and a continuous push towards stronger, verifiable privacy guarantees across the entire AI lifecycle. As AI becomes embedded in every facet of our lives, these innovations are essential for building a future where powerful AI coexists harmoniously with robust data privacy.
Share this content:
Discover more from SciPapermill
Subscribe to get the latest posts sent to your email.
Post Comment