Private Training

Private training in machine learning aims to develop and deploy models without directly exposing sensitive training data. Current research heavily focuses on mitigating privacy risks in federated learning and other distributed training paradigms, exploring techniques like differential privacy, random data augmentation, and secure multi-party computation, often applied to vision transformers and other deep neural networks. This field is crucial for enabling the use of sensitive data in machine learning applications across healthcare, finance, and other sectors while addressing significant ethical and regulatory concerns around data privacy.

Papers

August 4, 2022

Privacy-Preserving Action Recognition via Motion Difference Quantization
Sudhakar Kumawat, Hajime Nagahara
Action Recognition Privacy Preserving Private Training

July 24, 2022

Learnable Privacy-Preserving Anonymization for Pedestrian Images
Junwu Zhang, Mang Ye, Yao Yang
Consistent Anonymization Effect Full Body Private Training Face Anonymization Pedestrian Image

June 10, 2022

Deep Leakage from Model in Federated Learning
Zihao Zhao, Mengen Luo, Wenbo Ding
Full Model Model Weight Private Training Model Gradient Gradient Leakage Attack Deep Leakage

May 12, 2022

Fair NLP Models with Differentially Private Text Encoders
Gaurav Maheshwari, Pascal Denis, Mikaela Keller, Aurélien Bellet
Differential Privacy Adversarial Training Text Representation Sensitive Attribute Private Training NLP Fairness

May 10, 2022

Secure and Private Source Coding with Private Key and Decoder Side Information
Onur Günlü, Rafael F. Schaefer, Holger Boche, H. Vincent Poor
Secure Approach Private Training Secret Key Source Reconstruction

April 25, 2022

Analysing the Influence of Attack Configurations on the Reconstruction of Medical Images in Federated Learning
Mads Emil Dahlgaard, Morten Wehlast Jørgensen, Niels Asp Fuglsang, Hiba Nassar
Medical Image Full State Reconstruction External Influence Target Distribution Private Training Optimal Attack Initialization Scheme

March 29, 2022

Auditing Privacy Defenses in Federated Learning via Generative Gradient Leakage
Zhuohang Li, Jiaxin Zhang, Luyang Liu, Jian Liu
Generative Adversarial Network Gradient Information Empirical Privacy Private Training Gradient Leakage

March 3, 2022

Label-Only Model Inversion Attacks via Boundary Repulsion
Mostafa Kahla, Si Chen, Hoang Anh Just, Ruoxi Jia
Model Inversion Attack Private Training Label Only Model Inversion Attack Blackbox Attack Black Box Model Inversion Attack

January 27, 2022

Domain-Invariant Representation Learning from EEG with Private Encoders
David Bethge, Philipp Hallgarten, Tobias Grosse-Puppendahl, Mohamed Kari, Ralf Mikut, Albrecht Schmidt, Ozan Özdenizci
Emotion Recognition Domain Invariant Representation Private Training Privacy Preserving Representation EEG Based Deep

January 19, 2022

Privacy-Aware Human Mobility Prediction via Adversarial Networks
Yuting Zhan, Alex Kyllo, Afra Mashhadi, Hamed Haddadi
Adversarial Learning Mobility Data Private Training Utility Privacy Trade Human Mobility Prediction Trajectory Recovery Mobility Datasets

November 23, 2021

Jointly Learning from Decentralized (Federated) and Centralized Data to Mitigate Distribution Shift
Sean Augenstein, Andrew Hard, Kurt Partridge, Rajiv Mathews
LeArning Abstract Distribution Shift Federated Prompt Cooperation Decentralized FL Device Training DIstribution Estimation Private Training Training Data Distribution Centralized Data

November 14, 2021

Eluding Secure Aggregation in Federated Learning via Model Inconsistency
Dario Pasquini, Danilo Francati, Giuseppe Ateniese
Secure Aggregation Private Training Consistent Model Attribution Attack