Friday 23 October 2026, 09:05 AM

Resolving the decade-long MKFHE scalability bottleneck with constant ciphertext sizes

The September 2026 MKFHE breakthrough resolves a decade-long scalability bottleneck, enabling constant ciphertext sizes for dynamic MPC and private AI.

For years, secure multi-party computation has been pitched as a foundational tool for data privacy. The premise is straightforward: multiple organizations pool their data to train a machine learning model, but nobody actually sees the raw information. It sounds great in a whitepaper. In practice, Multi-Key Fully Homomorphic Encryption (MKFHE) has been stuck in a scalability nightmare.

We finally have the math to fix it.

Escaping the linear expansion trap

Historically, MKFHE suffered from an O(N) expansion issue. If you added more users to a computation, the ciphertext sizes and computing costs grew linearly. That made large-scale applications mathematically unfeasible for any real-world engineering team I have ever worked with.

In late Q3 2026, researchers Abtin Afshar and Rishab Goyal from UW-Madison published two papers that resolve this bottleneck. They provided the first concrete evidence that MKFHE can achieve constant ciphertext, public key, and secret key sizes under standard Learning with Errors (LWE) assumptions.

They figured out a way to compile Ring-GSW and BFV-like schemes to maintain constant ciphertext sizes regardless of the user count. They use a non-interactive preprocessing phase that dynamically aggregates public keys. This converts independently encrypted inputs into a short single-key ciphertext under a virtual key.

The performance numbers are what caught my attention. Using a leveled BFV-based scheme and a Go-language implementation, they benchmarked a 32-user computation. The multiplications executed up to two orders of magnitude faster than prior state-of-the-art schemes like CDKS and KKLSS. We are looking at MKFHE operating at speeds nearly identical to single-key FHE.

Dropping the trusted setup

Architectural friction points in cryptography often revolve around the Common Reference String (CRS). Relying on a trusted setup always introduces a weak link for real-world deployment. You have to trust whoever generated the keys did not keep a backdoor.

At the CRYPTO 2026 conference, Seonhong Min, Jeongeun Park, and Yongsoo Song presented a method to bypass this entirely. They successfully extended Ring-GSW to a multi-key variant in the plain model. This gives us a two-round MPC setup and provides concrete conversions for BGV, BFV, and CKKS schemes without relying on trusted setups at all.

Making it work on standard internet connections

Moving data across a network is often where cryptographic systems choke. These new MKFHE constructions introduce a highly efficient single-round distributed decryption protocol where the size of a user's partial decryption share is exactly equal to the plaintext length.

Cryptographers call this rate-1 simulatability. For those of us building products, it means optimal communication complexity. This drastically reduces the network bandwidth required for multi-party decryption, making it viable over standard internet connections instead of requiring specialized fiber links between data centers.

A new model for decentralized compute

With constant ciphertext sizes and reduced network overhead, the way we handle collective data is going to look fundamentally different by the 2030s. Right now, data is heavily siloed because sharing it means losing control of it. Over the next decade, this math allows us to decouple data utility from data ownership.

We are moving toward a model where ad-hoc, decentralized AI training becomes the default, rather than a specialized privacy feature. Hospitals can pool encrypted patient records to train predictive diagnostic models without ever exposing sensitive health data to a central server or each other. Global financial consortiums can detect fraud rings across different banks without sharing proprietary ledger information. Entities can finally pool encrypted data without agreeing on a single encryption key.

Google is already laying the groundwork for this shift. They recently updated their open-source Homomorphic Encryption Intermediate Representation (HEIR) compiler to support larger models. Highlighted in the August 2026 FHE.org Digest, this update shows that industry tooling is rapidly maturing to absorb these exact FHE optimizations for large-scale applications.

The engineering reality

The theoretical maturity of these breakthroughs is exceptionally high. The proof-of-concept software proves the math works and the speeds are there. But getting this into production is a different story.

Widespread commercial adoption will require translating these complex cyclotomic ring operations into memory-safe production code. We need accessible SDKs so developers can build on top of these protocols without needing a background in advanced cryptography. Even with constant ciphertext sizes, the CPU overhead of homomorphic encryption remains heavy. Until we see specialized FHE hardware accelerators hit the data center racks, managing computational latency will be the primary engineering constraint for any team trying to implement this at scale.