Robust aggregation of encrypted client contributions

Develop robust aggregation for encrypted client contributions in HE-OFT to provide protection against actively malicious participants who upload crafted head displacements.

Background

The cryptographic guarantees in HE-OFT primarily address semi-honest parties. A malicious client can upload an arbitrary encrypted head displacement, potentially biasing the shared classifier head, while the server cannot inspect the contribution because it remains encrypted.

The paper identifies two compatible defenses: enforcing a norm bound on each encrypted upload through a zero-knowledge proof, or performing robust aggregation directly under encryption. It leaves these defenses undeveloped and later characterizes robust aggregation over encrypted contributions as an unresolved challenge.

References

We identify two promising directions for future work. The first is robustness to actively malicious participants, which the present guarantee does not cover. Thus, robust aggregation over encrypted contributions remains an open challenge.

— HE-OFT: Privacy-Preserving One-Shot Federated Fine-Tuning under Homomorphic Encryption  (2610.08255 - Kanpak et al., 6 Oct 2026) in Section VIII, subsection “Malicious participants” and concluding paragraph of Section VIII