Affine subspace constraints improve federated gradient inversion
Original title: Gradient Inversion through Linear and Affine Subspace Constraints
The authors report better reconstruction than chosen baselines, including when batch size is moderately overestimated.
A paper extends gradient inversion against federated learning by treating correctly scaled fully connected layer gradients as an affine combination of inputs, not only a linear one. The attack uses known linear-subspace or affine-subspace information at each FC layer to reconstruct client data from shared gradients. The authors report that subspace methods can substantially improve reconstruction quality over chosen baselines in some settings, and that moderate overestimation of batch size can preserve or improve quality when the true size is unknown. They also test reconstruction when a differential-privacy mechanism perturbs gradients.
Unconfirmed: Reconstruction gains are reported only versus the authors' chosen baselines and in certain settings.
- Prior attacks often optimize dummy inputs to match client gradients.
- Affine constraints apply to aggregate gradients from an FC layer with correct scaling.
- Moderate batch-size overestimation can still yield usable reconstructions.
- Differential privacy noise is examined but no specific privacy budget is stated.
Impact: Reconstruction of client training data from shared gradients
From the original article
1 min read at arxiv.orgFederated learning enables multiple clients to collaboratively train a machine-learning model without directly sharing their local datasets. However, as prior works have shown, the gradients exchanged during training may reveal sensitive information about client data.
By Konstantinos Bouros · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.