See through Gradients: Image Batch Recovery via GradInversion

TLDR

Training deep neural networks relies on gradient estimation from data batches, and although averaging gradients over batches has been assumed to preserve privacy, prior work only demonstrated data recovery under restrictive conditions, leading to the belief that larger batch averaging is safe. This study introduces GradInversion, enabling recovery of all input images from large batches (8–48 images) in deep networks such as ResNet‑50 on ImageNet. GradInversion formulates an optimization that transforms random noise into natural images matching the gradients while enforcing image fidelity, includes a target‑class recovery algorithm, and employs a group consistency framework where multiple agents collaboratively refine the reconstruction. The results show that gradients contain sufficient information to recover every individual image with high fidelity, even for complex datasets, deep networks, and large batch sizes.

Abstract

Training deep neural networks requires gradient estimation from data batches to update parameters. Gradients per parameter are averaged over a set of data and this has been presumed to be safe for privacy-preserving training in joint, collaborative, and federated learning applications. Prior work only showed the possibility of recovering input data given gradients under very restrictive conditions – a single input point, or a network with no non-linearities, or a small 32 × 32 px input batch. Therefore, averaging gradients over larger batches was thought to be safe. In this work, we introduce GradInversion, using which input images from a larger batch (8 – 48 images) can also be recovered for large networks such as ResNets (50 layers), on complex datasets such as ImageNet (1000 classes, 224 × 224 px). We formulate an optimization task that converts random noise into natural images, matching gradients while regularizing image fidelity. We also propose an algorithm for target class label recovery given gradients. We further propose a group consistency regularization framework, where multiple agents starting from different random seeds work together to find an enhanced reconstruction of the original data batch. We show that gradients encode a surprisingly large amount of information, such that all the individual images can be recovered with high fidelity via GradInversion, even for complex datasets, deep networks, and large batch sizes.