Data Reconstruction Attack

Data reconstruction attacks exploit information leaked during machine learning processes, particularly in federated learning (FL) and vertical federated learning (VFL), to reconstruct the private training data of participating clients. Current research focuses on developing and analyzing these attacks, often leveraging techniques like gradient inversion, linear layer leakage, and diffusion models, across various model architectures including neural networks. Understanding and mitigating these attacks is crucial for ensuring the privacy of sensitive data used in collaborative machine learning, impacting the security and trustworthiness of numerous applications in finance, e-commerce, and beyond.

Papers

June 8, 2022

Gradient Obfuscation Gives a False Sense of Security in Federated Learning
Kai Yue, Richeng Jin, Chau-Wai Wong, Dror Baron, Huaiyu Dai
Security Related Gradient Compression Privacy Preserving Machine Learning False Sense Data Reconstruction Attack Gradient Obfuscation

February 6, 2022

BEAS: Blockchain Enabled Asynchronous & Secure Federated Machine Learning
Arup Mondal, Harpreet Virk, Debayan Gupta
Federated Learning Blockchain Based Platform Differential Privacy Guarantee Private Datasets Model Poisoning Attack Data Reconstruction Attack Asynchronous Architecture

December 6, 2021

When the Curious Abandon Honesty: Federated Learning Is Not Private
Franziska Boenisch, Adam Dziedzic, Roei Schuster, Ali Shahin Shamsabadi, Ilia Shumailov, Nicolas Papernot
Deep Convolutional Neural Network Data Leakage Curiosity Inducing Situation Data Reconstruction Attack Gradient Sharing Field Trap

Data Reconstruction Attack

Papers

Gradient Obfuscation Gives a False Sense of Security in Federated Learning

BEAS: Blockchain Enabled Asynchronous & Secure Federated Machine Learning

When the Curious Abandon Honesty: Federated Learning Is Not Private