Logo image
Back
Exposing Privacy Vulnerabilities in Federated Learning: A GAN-Based Model Inversion Attack
Conference proceeding   Peer reviewed

Exposing Privacy Vulnerabilities in Federated Learning: A GAN-Based Model Inversion Attack

Md Morshedul Islam, Suraj Neupane, Md Jobair Hossain Faruk, Hossain Shahriar and Alfredo Cuzzocrea
IEEE International Conference on Big Data, (2025), pp.4227-4236
IEEE International Conference on Big Data (BigData) (Macau, China, 12/08/2025–12/11/2025)
03/2026
DOI: https://doi.org/10.1109/BigData66926.2025.11401939

Metrics

1 Record Views

Abstract

Adaptation models aggregation algorithm Deep Convolutional GAN Federated learning model inversion attack MRI dataset Generative adversarial networks (GAN) Big Data Cybersecurity Machine Learning
Federated learning (FL) is a decentralized machine learning (ML) approach that helps handle big data and enhances privacy by training models locally on devices and sharing only model updates with the central server instead of raw data. While FL reduces the risk of direct data exposure, it is still susceptible to advanced privacy attacks, many of which were originally studied in centralized ML settings. Recently, there has been growing interest in adapting these attacks to the federated learning context. However, this adaptation is non-trivial due to fundamental differences in system architecture, data distribution, and threat models between centralized and federated learning. One such notable threat is the model inversion attack, where an adversary attempts to reconstruct representative samples of a target class. In domains such as healthcare, where strict privacy safeguards are required, these attacks introduce substantial risks. This work proposes a model inversion attack based on Generative Adversarial Networks (GANs), tailored for FL. Our method operates under realistic assumptions about the knowledge and capabilities of the attacker. The attack is both model-agnostic and aggregation-agnostic, requiring no dependence on the model architecture or aggregation strategy used in FL. We evaluated the attack using the large-scale Brain MRI ND-5 dataset and demonstrated that our GAN-based approach can effectively reconstruct class-specific MRI images, thus undermining the privacy guarantees of the FL systems. Our method is generalizable and applicable to various federated learning scenarios.

Details

Logo image