Homomorphic Encryption

HE is compelling when a service must compute over encrypted inputs, but practical ML workloads are bounded by operator support, latency, ciphertext size, and model design.

Research Handoff

Problem	Who benefits if solved	Why it is difficult	Starting directions
Practical HE model architectures	Model providers, ML engineers, and buyers comparing private inference options	HE constraints interact with layers, activations, precision, batching, and accuracy	Benchmark small model families; publish unsupported operations; compare redesign cost
Debugging encrypted computation	ML engineers, incident responders, and HE library maintainers	Production failures can hide behind ciphertext, encoding, scaling, and approximation errors	Build plaintext shadow traces; publish test vectors; flag parameter drift and unsupported paths
HE inference benchmarks that matter	Buyers, benchmark maintainers, and vendors	One latency number rarely includes keys, transfer, batching, accuracy loss, and cloud cost	Report end-to-end measurements; compare HE, TEE, and plaintext baselines; include ciphertext size
Explaining HE failure early	Architects, sales engineers, and product teams	Teams often discover incompatibility only after model adaptation work has begun	Write model preflight analyzers; map layers to HE support; estimate latency bands before implementation

Practical HE Model Architectures

Field	Card
Problem	What model architectures are actually practical for HE inference?
The itch	Teams hear that encrypted inference is possible, then discover their model uses layers, activations, or precision that are painful under HE.
Why it matters	Product teams need to know whether to redesign the model, switch PETs, or abandon private inference.
Current workaround	Use toy models, vendor demos, or late-stage prototypes.
Why the workaround is insufficient	It hides accuracy loss, batching constraints, and deployment cost.
What good progress would look like	A model catalog with supported operators, latency, accuracy, ciphertext size, and parameter choices for common tasks.
Difficulty	Medium
Good for	ML researcher, systems builder, cryptographer, benchmark maintainer
Related PETs	HE, private inference, TEEs
Possible first contribution	Benchmark three small architectures for tabular or image inference under one HE library and publish unsupported operations.

Debugging Encrypted Computation

Field	Card
Problem	How can developers debug encrypted computation?
The itch	When encrypted inference returns the wrong answer, developers cannot inspect intermediate plaintext values in the deployed path.
Why it matters	Debuggability affects trust, incident response, and adoption by normal ML engineering teams.
Current workaround	Compare against plaintext simulations or ask cryptography specialists to inspect parameters.
Why the workaround is insufficient	It misses production-only failures such as encoding mistakes, scale drift, key handling bugs, and data distribution shifts.
What good progress would look like	Debug tooling that links encrypted traces to safe plaintext simulations, parameter warnings, and reproducible test vectors.
Difficulty	Hard
Good for	Systems builder, cryptographer, ML engineer
Related PETs	HE
Possible first contribution	Build a test harness that runs a plaintext shadow computation and reports where HE approximation error diverges.

HE Inference Benchmarks That Matter

Field	Card
Problem	How can HE inference be benchmarked across latency, cost, and accuracy?
The itch	Benchmarks often report one impressive latency number without deployment context.
Why it matters	Buyers need to compare HE with TEEs, client-side inference, or standard hosted inference.
Current workaround	Vendor-specific demos and single-model papers.
Why the workaround is insufficient	It rarely captures batching, key setup, ciphertext transfer, accuracy loss, and cloud cost.
What good progress would look like	A benchmark format that reports end-to-end latency, throughput, ciphertext size, accuracy, parameter choices, and cost.
Difficulty	Good first research problem
Good for	Benchmark maintainer, systems builder, ML researcher
Related PETs	HE, TEEs, private inference
Possible first contribution	Create a reproducible benchmark for one tabular classifier comparing plaintext, TEE, and HE inference.

Explaining HE Failure Early

Field	Card
Problem	How can tools tell teams early that HE is the wrong PET for a workload?
The itch	Engineers may spend weeks adapting a model before realizing the latency or operator set will not work.
Why it matters	Honest rejection criteria save time and reduce overclaiming.
Current workaround	Ask experts to review the model manually.
Why the workaround is insufficient	Expert review does not scale and may happen after sales or architecture commitments.
What good progress would look like	A preflight analyzer that flags unsupported layers, expensive operations, precision risks, and expected latency bands.
Difficulty	Medium
Good for	Systems builder, cryptographer, ML engineer
Related PETs	HE, private inference
Possible first contribution	Write a model-inspection script for one framework that maps layers to HE support and warnings.