Core Concepts
Exact reconstruction of batches in federated learning is possible, challenging prior assumptions.
Abstract
Federated learning framework for collaborative machine learning.
Gradient inversion attacks can reconstruct client data from shared gradients.
Proposed algorithm, SPEAR, reconstructs batches with b > 1 exactly.
Leveraging low-rank structure and ReLU-induced sparsity for reconstruction.
Efficient GPU implementation for large network widths and depths.
Comparison to prior work and contributions highlighted.
Method overview, setting, low-rank decomposition, ReLU-induced sparsity, sampling, filtering, and validation.
Empirical evaluation on MNIST, CIFAR-10, and TINYIMAGENET datasets.
Ablation studies on batch size, network architecture, and training steps.
Related work on gradient inversion attacks and defenses.
Limitations, broader impact, and conclusion discussed.
Stats
"Exact reconstruction is possible for batch sizes of b = 1."
"Reconstructs batches of b ≲ 25 elements exactly."
"Highly parallelized GPU implementation for large network widths and depths."
Quotes
"Exact reconstruction of batches is possible in the honest-but-curious setting."
"SPEAR succeeds with high probability for batch sizes b > 1."