Skip to content
conftrace_
2026 ICML ICML 2026

Position: The Privacy-Auditability Paradox in Federated Learning: Why We Need Controllable Secure Aggregation

Federated Learning (FL) has emerged as the de facto standard for privacy-preserving machine learning, largely due to Secure Aggregation (SecAgg) protocols rendering individual user contributions cryptographically invisible to the server; we contend that this pursuit of unconditional blind aggregation has produced a systemic vulnerability, the Privacy-Auditability Paradox. By making individual updates indistinguishable in the aggregate, default SecAgg creates a Sanitization Gap in which input-level poisoning is undetectable, and a Regulatory Dead Zone in which compliance with the EU AI Act’s robustness and traceability mandates are structurally undermined. In this position paper, we argue that the community must transition from blind secure aggregation to Controllable Secure Aggregation (CSA), in which privacy is not a binary state of secrecy but a programmable spectrum of cryptographic access. CSA replaces binary secrecy with fine-grained, policy-based governance built on a complementary stack of functional encryption, zero-knowledge proofs, and threshold cryptography, yielding Verified Blindness: the server remains blind by default but admits a consensus-bound, cryptographically regulated break-glass channel for auditing specific inputs. Adopting CSA is not merely a technical upgrade but a necessary condition for transforming FL from an unregulated academic prototype into a robust, compliant, and trustworthy privacy infrastructure.

More from ICML 2026

All ICML 2026 papers →