fhelium.engine.ckks_decryptor

CKKS decryption and bounded coefficient reconstruction.

CkksDecryptor

class View source

CkksDecryptor(*, config: CkksConfig, device: torch.device, rns_layout: RnsLayout, rns_runtime: RnsRuntime, montgomery_parameters: MontgomeryParameters, plaintext_codec: CkksPlaintextCodec, engine_id: str)

Decrypt ciphertext phases to bounded approximate coefficients.

The component evaluates two- or three-component secret-key phases in RNS and reconstructs only a bounded trailing-Q class into binary64 coefficients for decoding. That reconstruction is not exact full-$Q_{\ell }$ CRT reconstruction. Key inventory and engine validation remain engine-owned.

decrypt

method

def decrypt(ciphertext: Ciphertext, secret_key: SecretKey) -> Plaintext: ...

Decrypt into bounded binary64 coefficient reconstruction.

Secret-key evaluation first forms

$$u(X)=\sum _{j=0}^{d-1}{c}_j(X)s(X{)}^j(\mathrm{mod}{B}_{\ell }),d\in \{2,3\},$$

as coefficient-domain canonical standard RNS. The direct decrypt path then reconstructs the centered class from the trailing one/two Q primes into finite torch.float64 [*batch, coefficient] on the engine device. This approximate_coefficients result is valid only for decoding; it is not exact full-$Q_{\ell }$ CRT and cannot return to RNS. Level and actual scale are preserved. Inputs are unchanged and output storage is independent.

decrypt_message

method

def decrypt_message(ciphertext: Ciphertext, secret_key: SecretKey, *, is_real: bool=False): ...

Decrypt and decode with the ciphertext's actual scale.

This composes decrypt's bounded approximate-coefficient reconstruction with $m_{\mathrm{approx}}={\mathcal{E}}_g(p)/\mathrm{\Delta }(ciphertext)$. Output is CPU [*batch, slot] with final extent $S=N/2$, complex unless is_real=True. It is approximate and does not provide an exact decrypt-to-encrypt representation round trip. Inputs are unchanged.