Clarify the Utility–Cost Tradeoff of Privacy-Preserving Techniques in Heterogeneous Federated Learning

Determine the real utility–cost tradeoff of privacy-preserving techniques for federated learning—including differential privacy, homomorphic encryption, secure aggregation, multi-party computation, and trusted execution environments—when deployed under heterogeneous conditions that reflect practical statistical, device, and communication variability, in order to assess their practical viability and impact on model utility and system overhead.

Background

The paper focuses on measuring poisoning-based security risks in federated learning and introduces TFLlib to evaluate attacks under realistic deployment constraints. Privacy-specific risks and defenses are deliberately deferred, noting that production FL stacks rarely integrate many proposed privacy-preserving techniques.

The authors highlight that, under real-world heterogeneity (statistical, device, and communication), the practical costs and utility impacts of these privacy mechanisms are not well understood, motivating a dedicated evaluation to characterize their tradeoffs before adoption in deployed systems.

References

many privacy-preserving techniques proposed in the literature are still rarely integrated into production FL stacks, and their real utility-cost tradeoff under heterogeneous deployments remains unclear.

— Unveiling the Security Risks of Federated Learning in the Wild: From Research to Practice  (2603.20615 - Chen et al., 21 Mar 2026) in Section 7, Future Work

This trade-off between training quality and data privacy is an open problem.

— Calibrated Decision Models for Autonomous Penetration-Testing Harnesses: JEV and Laya as System One Decision Layers for LLM-Driven Pentest Agents  (2609.28940 - Barbosa, 24 Sep 2026) in Section 20.2, "Federated training across engagements"

We therefore do not expect optimal utility, effectiveness, or overhead for various operations, and improving these trade-offs is left as future work.

— Trusted Model Environment for Private Semantic Computations  (2609.30032 - Duddu et al., 24 Sep 2026) in Section 5, Evaluation

Furthermore, integrating privacy-preserving techniques such as differential privacy or secure multi-party computation into the trust score computation remains an open challenge.

— Similarity-Aware Personalized Federated Learning in Heterogeneous Environments  (2609.02241 - A et al., 2 Sep 2026) in Section 'Limitations and Future Work'

evaluation under DP-FTRL and privacy auditing remains future work.

— RegionFed: Federated Learning for Personalized Query Understanding in Heterogeneous Retail Environments  (2609.05403 - Nguyen et al., 4 Sep 2026) in Appendix, Section 'Limitations, Broader Impact, and Ethics', subsection 'Limitations'

How adaptive aggregation, heterogeneous budgets, and communication-efficient methods interact at cross-facility scale remains open.

— From Pilots to Production: Lessons in Cross-Institutional Federated Training and Artificial Intelligence for Science  (2609.39803 - Kotevska et al., 30 Sep 2026) in Section 4, “What Remains Unsolved, and a Shared Agenda”

A quantitative study of this privacy--utility trade-off is left for future work.

— Tram-FL: Reducing Communication and Computation Costs through Sequential Model Circulation in Decentralized Federated Learning  (2610.07859 - Maejima et al., 6 Oct 2026) in Section 3.1, System Model