Advances in Privacy-Preserving Machine Learning and Biometric Security

The field of machine learning and biometric security is moving towards a more privacy-preserving and secure direction. Researchers are exploring innovative methods to protect sensitive information while maintaining model utility. One of the key trends is the integration of differential privacy into various machine learning architectures, including federated learning and graph neural networks. This approach enables the protection of sensitive data while allowing for collaborative model training and data analysis. Another area of focus is the development of more secure and efficient biometric systems, including face recognition and fingerprint identification. These systems aim to provide high accuracy while minimizing the risk of data leakage and inference attacks. Overall, the field is witnessing significant advancements in privacy-preserving machine learning and biometric security, with a focus on developing practical and efficient solutions for real-world applications. Noteworthy papers include:

  • FedNCA, which introduces a novel federated learning system tailored for medical image segmentation tasks, enabling training on low-cost edge devices while minimizing communication costs.
  • CARIBOU, which proposes a convergent privacy framework with contractive GNN layers for multi-hop aggregations, ensuring the contractiveness required for theoretical guarantees while preserving model utility.
  • LH2Face, which presents a novel loss function for hard high-quality face recognition, achieving superior performance on challenging datasets.
  • PPFL-RDSN, which develops a privacy-preserving federated learning-based framework for encrypted lossy image reconstruction, demonstrating comparable performance to state-of-the-art centralized methods while reducing computational burdens.

Sources

Equitable Federated Learning with NCA

A Framework for Multi-source Privacy Preserving Epidemic Analysis

Closing the Performance Gap in Biometric Cryptosystems: A Deeper Analysis on Unlinkable Fuzzy Vaults

Convergent Privacy Framework with Contractive GNN Layers for Multi-hop Aggregations

LH2Face: Loss function for Hard High-quality Face

PPFL-RDSN: Privacy-Preserving Federated Learning-based Residual Dense Spatial Networks for Encrypted Lossy Image Reconstruction

Privacy-preserving Preselection for Face Identification Based on Packing

L-VAE: Variational Auto-Encoder with Learnable Beta for Disentangled Representation

Embedding-Based Federated Data Sharing via Differentially Private Conditional VAEs

Built with on top of