High-Throughput Universally Composable Threshold FHE Decryption (eprint.iacr.org)

A new paper presents a practical, high-throughput protocol for threshold fully homomorphic encryption (FHE) decryption that eliminates the longtime reliance on "noise flooding" and achieves simulation-based (Universal Composability, UC) security. Threshold FHE lets multiple parties jointly decrypt FHE ciphertexts so that encrypted computation can be used in low-communication multi-party computation (MPC) pipelines—a key building block for privacy-preserving ML and other collaborative workflows. Prior solutions traded off efficiency or composability by masking ciphertext noise with flooding, which inflates parameters and runtime; this work instead removes noise deterministically via an MPC rounding procedure, avoiding the flooding overhead entirely. Technically, the protocol amortizes the cost of secure comparisons with an offline/online design: expensive preprocessing prepares special gates for non-linear comparisons offline, leaving a lightweight online phase with low communication and computation. The construction is proven UC-secure and can be instantiated for different adversary models (security-with-abort under a dishonest majority, or guaranteed output delivery with an honest majority). Empirically, the authors report ≈20,000× higher online throughput and up to 37× lower latency versus state-of-the-art. For the AI/ML community, this makes threshold FHE-based MPC far more practical—especially for workloads heavy in non-linear comparisons (e.g., secure inference, private model aggregation), enabling more scalable, composable privacy-preserving ML deployments.