TFHE-rs v0.11.0

Summary

TFHE-rs v0.11.0 introduces significant improvements and new features, including enhanced Zero Knowledge proof performance, a new FheAsciiString type for secure string operations, and a 28% CPU performance boost for shift/rotate operations on 64-bit values. On the GPU side, we've improved GPU performance by 30% for 64-bit addition and expanded GPU support to arrays of ciphertexts. In addition, TFHE-rs v0.11 introduces noise assertions for debugging, new default parameters, and several fixes and optimizations to enhance overall usability and reliability.

What's Changed

Breaking Changes

Warning

Incompatible data structures:
Data compatibility is not guaranteed for versions older than 0.10.
The following structures and their seeded variants are incompatible:

• All GgswCiphertext and its Fourier and NTT variants
• GgswCiphertextList, LweBootstrapKey and its Fourier and NTT variants
• LweKeyswitchKey, LweMultiBitBootstrapKey and its Fourier variant
• LwePackingKeyswitchKey, LwePrivateFunctionalPackingKeyswitchKey
• LwePrivateFunctionalPackingKeyswitchKeyList
• All the shortint, integer and High Level API keys using those lower level primitives, including ServerKey types and their compressed variants.

Parameter and function changes:

• The DispersionParameter trait in core_crypto now returns new types for all its functions instead of raw f64.
• Zero Knowledge proofs can now only be used with dedicated parameters.
• ZkVerificationOutCome has been renamed to ZkVerificationOutcome
• CompactPkePublicParams has been deprecated; use CompactPkeCrs instead.
• Arch specific features (like x86_64-unix) have been dropped in favor of auto-detection.
• Default parameters are now TUniform parameters.
• ActivatedRandomGenerator has been renamed to DefaultRandomGenerator.

Please update your code accordingly.

New features

CPU

• Add strings module, usable in the HL API via the FheAsciiString type
• Add Zero Knowledge v2 with vastly improved performance
• LweSecretKey and GlweSecretKey now support safe serialization
• Add safe_serialized_size to get the seriliazed size of entities supporting safe serialization
• Add a batched PBS loading the key only once and processing all ciphertexts at once
• Add the safe_deserialize method for CompactPkeCrs in the Javascript API
• Add C API primitives to get/clear the last error from TFHE-rs as a const char * and enable/disable automatic printing of errors to stderr
• Add debug noise-asserts feature to verify shortint noise levels are valid with respect to a user-provided maximum noise level

GPU

• Add absolute value
• Add signed division
• Add array types
• Add comparison of arrays and array search

Improvements

CPU

• Encrypted shift and rotate have improved performance (+28% for 64 bits values)
• Conformance checks for Zero Knowledge proofs now also verify the validity of elliptic curve points

GPU

• Modify double to u64 cast in some kernels to use intrinsics.
• Optimize integer multiplication when one of the input ciphertexts holds a boolean value.
• Improve single carry propagation performance.
• Add a function to check if a cuda device is available on the machine.

Fixes

CPU

• High Level API CompressedCiphertextListBuilder now uses the current ServerKey device when compressing data instead of panicking when mixing CPU and GPU usage
• Fix compact public key encryption of a single ciphertext in core_crypto, shortint, integer and the High Level API were not affected by the bug
• Fix serialized_size_limit to include header sizes of entities
• Fix gadget decomposition over one level to be balanced
• Fix Zero Knowledge proof compatibility between 32 bits (WASM) and 64 bits plaftorm
• Fix a corner-case for carry propagation for signed sub with overflow which could only be triggered by manually selecting propagation algorithm
• Fix a corner-case for comparisons in case of an empty (no blocks) integer ciphertext
• Fix full_propagate_parallelized to also reset the noise of blocks not part of the carry propagation
• Fix encoding during compression which was pessimizing the failure probability

GPU

• Fixes to compression.
• Fix sample extract for the many LUT use case.
• Fix many LUT with multiple GPUs.
• Fix memory error in cg classical PBS.
• Fix gadget decomposition over one level to be balanced.
• General fixes and improvements to the multi-bit PBS.
• Fix signed overflowing subtraction for one block case.
• Fix equal and scalar ne.
• Fix single GPU execution on a device other than 0 for all functions.
• Fix ct degree / noise level tracking in several operations.