EIFL: Efficiently Protecting Global Model Privacy and Integrity Against an Untrusted Server in Federated Learning
EIFL lets federated-learning clients keep a global model private and verify it against an untrusted server.
Federated-learning servers can tamper with the aggregated global model or, in commercial settings, should not see it. EIFL combines two-stage aggregation with symmetric encryption for output privacy and an inner-product integrity check whose auxiliary random vectors need not stay hidden from the server. Binding that auxiliary data to the check removes a failure mode in which leaked shared secrets let the server bypass verification. The protocol tolerates client dropout by resending and is reported to reduce computation and communication versus earlier schemes.
- Two-stage aggregation and symmetric encryption hide the global model
- Inner-product checks verify integrity without secret auxiliary data
- Clients agree on random vectors instead of pre-shared secrets
- Dropout during verification is handled by resending
- Reported compute and communication costs beat prior schemes
From the original article
1 min read at arxiv.orgFederated learning (FL) typically adopts a server-client architecture, where the server aggregates clients' local models (i. e. , the input) and returns the aggregated global model (i.
By Zehui Liao · 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.