BatchNorm buffers can fake deletion in split-learning unlearning
Original title: Phantom Forgetting: How BatchNorm State Fakes Deletion in Edge Split Machine Unlearning
CIFAR-10 forget accuracy fell 61.50 points from BN recalibration with every learnable weight fixed.
A paper asks whether Batch Normalization running statistics can imitate forgetting in edge-server split learning when no learnable weight changes. A state-only diagnostic separates weight changes from BN-buffer changes on the client, server, and both sides. Across all ten CIFAR-10 forget classes, centralized BN recalibration on forget data cut mean forget accuracy by 61.50 percentage points. Edge-only, server-only, and joint recalibration produced mean drops of 37.27, 41.71, and 51.90 points. The authors say forget accuracy should not certify deletion without an edge-aware state-integrity audit.
- Server-side state gave the largest selective gap versus retain degradation.
- Joint recalibration caused the largest overall prediction change.
- A state-induced forgetting gap discounts simultaneous retain-set loss.
- The effect is localized to edge, server, and joint BN placements.
From the original article
1 min read at arxiv.orgSplit learning is a natural fit for edge intelligence: a resource-constrained edge device runs the first layers of a network and offloads the remainder to a server, so that raw data never leave the device. When such a system must unlearn revoked or sensitive data for example to honor a right-to-erasure request, the deletion is commonly assessed through changes in forget-set accuracy.
By Anand Agrawal · 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.