Generative Image Reconstruction From Gradients.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision
• Cybersecurity

Background:

• Federated learning (FL) enables collaborative model training without sharing raw data, preserving client privacy.
• Previous research demonstrated that gradients shared in FL can be exploited to infer private training data, including image reconstruction.
• Existing gradient inversion methods face limitations in image resolution, batch size, and often require prior knowledge of the client's dataset.

Purpose of the Study:

• To propose a novel method, Generative Image Reconstruction from Gradients (GIRG), for recovering training images from shared gradients in FL.
• To address the limitations of existing methods by enabling high-resolution image reconstruction with large batch sizes and without prior data knowledge.
• To evaluate the privacy risks associated with gradient sharing in current FL practices.

Main Methods:

• GIRG employs a conditional generative model to reconstruct training images and their associated labels directly from shared gradients.
• The method optimizes the weights of the generative model, rather than input vectors, to produce accurate 'dummy' images.
• GIRG does not require any prior information about the client's training data for image reconstruction.

Main Results:

• GIRG successfully reconstructs high-resolution images from gradients, even with large batch sizes.
• The method demonstrates the capability to recover images from aggregated gradients originating from multiple FL participants.
• Empirical results confirm the effectiveness of GIRG in recovering detailed image information from shared gradients.

Conclusions:

• The proposed GIRG method poses a significant privacy risk to FL by enabling robust image reconstruction from shared gradients.
• Current FL practices that rely solely on gradient sharing are vulnerable to sophisticated inversion attacks.
• There is an urgent need for enhanced privacy-preserving mechanisms to mitigate gradient inversion risks in collaborative training.