How to read this table
- Assumption is the hardness problem the scheme relies on. Lattice problems (LWE, RLWE, NTRU) are believed to resist quantum attacks; factoring and discrete-log schemes are not.
- Plaintext is what one ciphertext slot holds and which operations are native. CKKS gives approximate reals; BGV/BFV give exact integers; TFHE/FHEW give bits or small integers through lookup tables; Gentry–Lee gives whole matrices.
- SIMD says whether one ciphertext packs many values. This matters most for the large matrix products in LLMs.
- Bootstrapping says whether and how cheaply ciphertext noise can be refreshed, which bounds circuit depth.
- Fit for LLM inference is a one-line judgement of strengths and limits against CKKS.
- Group 9 lists non-HE primitives (secret sharing, garbled circuits, function secret sharing, functional encryption). They need interaction or leak more, but are the main competitors in practice.
- Every scheme links to a visual lesson that runs the same tiny LLM neuron through a real toy version of the scheme, step by step.
- Click More for the original paper title, authors, venue and notable follow-up papers.
| Scheme | Year | Assumption | Plaintext / arithmetic | SIMD | Bootstrapping | Fit for LLM inference | Original paper | Libraries | Status |
|---|