README.md

@@ -13,8 +13,7 @@ Additionally, several design decisions were made to maximize the usability of th
 - Make the number of dependencies minimal.
 - Make the software easily hackable.
 - Set the usage of multiple GPUs as the default.
-- Make the resulting library easily integrated with the pre-existing software, especially Artificial Intelligence (
-  AI) related ones.
+- Make the resulting library easily integrated with the pre-existing software, especially Artificial Intelligence (AI) related ones.
 
 # Key Features
 
@@ -30,7 +29,7 @@ from liberate import fhe
 from liberate.fhe import presets
 
 # Generate CKKS engine with preset parameters
-grade = "silver"  # logN=14
+grade = "silver"  # logN=15
 params = presets.params[grade]
 
 engine = fhe.ckks_engine(**params, verbose=True)
@@ -66,7 +65,8 @@ the [official documentation](https://docs.desilo.ai/liberate-fhe/getting-started
 ### Clone this repository
 
 ```shell
-git clone https://github.com/Desilo/liberate.git
+git clone https://github.com/Desilo/liberate-fhe.git
+cd liberate-fhe
 ```
 
 ### Install dependencies
@@ -108,7 +108,7 @@ instructions, examples, and documentation.
   title={{Liberate.FHE: A New FHE Library for Bridging the Gap between Theory and Practice with a Focus on Performance and Accuracy}},
   author={DESILO},
   year={2023},
-  note={\url{https://github.com/Desilo/liberate}},
+  note={\url{https://github.com/Desilo/liberate-fhe}},
 }
 ```
 

clean.py

@@ -0,0 +1,29 @@
+import pathlib
+import shutil
+
+
+def clean_project():
+    """
+        Remove common build directories and *.so files in a Python project.
+    """
+    # Directories to remove
+    dirs_to_remove = [
+        "__pycache__", ".pytest_cache",
+        "build", "*.egg-info",  # "dist",
+    ]
+    # File patterns to remove
+    files_to_remove = ["*.so"]
+
+    for path in pathlib.Path(".").rglob("*"):
+        # Remove specified directories
+        if path.is_dir() and any(path.match(d) for d in dirs_to_remove):
+            shutil.rmtree(path, ignore_errors=True)
+            print(f"Removed directory: {path}")
+        # Remove specified files
+        elif path.is_file() and any(path.match(f) for f in files_to_remove):
+            path.unlink()
+            print(f"Removed file: {path}")
+
+
+if __name__ == "__main__":
+    clean_project()

pyproject.toml

@@ -13,13 +13,16 @@ maintainers = [
 readme = [
     "README.md"
 ]
+
 homepage = "https://desilo.ai/"
 documentation = "https://docs.desilo.ai/liberate-fhe/"
-repository = "https://github.com/desilo/liberate"
+repository = "https://github.com/desilo/liberate-fhe/"
+
 license = "BSD-3-Clause-Clear"
+
 keywords = [
-    "DESILO", "python", "cryptohraphy", "privacy", "encryption", "cuda",
-    "homoomrphic encryption", "homomorphic encryption library", "fhe", "gpu accelerated", "liberate"
+    "DESILO", "python", "cryptoghraphy", "privacy", "encryption", "cuda",
+    "homomorphic encryption", "homomorphic encryption library", "fhe", "gpu accelerated", "liberate"
 ]
 classifiers = [
     "Environment :: GPU :: NVIDIA CUDA",
@@ -45,13 +48,13 @@ include = [
 ]
 
 [tool.poetry.dependencies]
-python = ">=3.10,<3.12"
+python = ">=3.10,<3.13"
 numpy = "^1.23.5"
 mpmath = "^1.3.0"
 scipy = "^1.10.1"
 matplotlib = "^3.7.1"
 joblib = "^1.2.0"
-torch = "^2.0.0"
+torch = "==2.2.1"
 tqdm = "^4.66.1"
 
 [tool.poetry.group.dev.dependencies]
@@ -73,3 +76,6 @@ line-length = 79
 [tool.isort]
 profile = "black"
 line_length = 79
+
+[tool.poetry.scripts]
+clean = "clean:clean_project"

setup.py

@@ -5,29 +5,29 @@
     CUDAExtension(
         name="randint_cuda",
         sources=[
-            "liberate/csprng/randint.cpp",
-            "liberate/csprng/randint_cuda_kernel.cu",
+            "src/liberate/csprng/randint.cpp",
+            "src/liberate/csprng/randint_cuda_kernel.cu",
         ],
     ),
     CUDAExtension(
         name="randround_cuda",
         sources=[
-            "liberate/csprng/randround.cpp",
-            "liberate/csprng/randround_cuda_kernel.cu",
+            "src/liberate/csprng/randround.cpp",
+            "src/liberate/csprng/randround_cuda_kernel.cu",
         ],
     ),
     CUDAExtension(
         name="discrete_gaussian_cuda",
         sources=[
-            "liberate/csprng/discrete_gaussian.cpp",
-            "liberate/csprng/discrete_gaussian_cuda_kernel.cu",
+            "src/liberate/csprng/discrete_gaussian.cpp",
+            "src/liberate/csprng/discrete_gaussian_cuda_kernel.cu",
         ],
     ),
     CUDAExtension(
         name="chacha20_cuda",
         sources=[
-            "liberate/csprng/chacha20.cpp",
-            "liberate/csprng/chacha20_cuda_kernel.cu",
+            "src/liberate/csprng/chacha20.cpp",
+            "src/liberate/csprng/chacha20_cuda_kernel.cu",
         ],
     ),
 ]
@@ -36,8 +36,8 @@
     CUDAExtension(
         name="ntt_cuda",
         sources=[
-            "liberate/ntt/ntt.cpp",
-            "liberate/ntt/ntt_cuda_kernel.cu",
+            "src/liberate/ntt/ntt.cpp",
+            "src/liberate/ntt/ntt_cuda_kernel.cu",
         ],
     )
 ]
@@ -50,7 +50,7 @@
         script_args=["build_ext"],
         options={
             "build": {
-                "build_lib": "liberate/csprng",
+                "build_lib": "src/liberate/csprng",
             }
         },
     )
@@ -62,7 +62,8 @@
         cmdclass={"build_ext": BuildExtension},
         options={
             "build": {
-                "build_lib": "liberate/ntt",
+                "build_lib": "src/liberate/ntt",
             }
         },
     )
+

src/liberate/csprng/csprng.py

@@ -222,23 +222,18 @@ def generate_nonce(self, seed):
         # nonce is 64bits.
         return self.generate_initial_bytes(8, seed=None)
 
-    def randbytes(self, shares=None, repeats=0, length=None, reshape=False):
+    def randbytes(self, shares=None, repeats=0, reshape=False):
         # Generates (shares_i + repeats) X length random bytes.
         if shares is None:
             shares = self.shares
 
-        if length is None:
-            L = self.L
-        else:
-            L = length
-
         # Set the target states.
         target_states = []
         for devi in range(self.num_devices):
             start_channel = self.shares[devi] - shares[devi]
             end_channel = self.shares[devi] + repeats
             device_states = self.channeled_states[devi][
-                start_channel:end_channel, :L, :
+                start_channel:end_channel, :, :
             ]
             target_states.append(device_states.view(-1, 16))
 
@@ -247,22 +242,17 @@ def randbytes(self, shares=None, repeats=0, length=None, reshape=False):
 
         # If not reshape, flatten.
         if reshape:
-            random_bytes = [rb.view(-1, L, 16) for rb in random_bytes]
+            random_bytes = [rb.view(-1, self.L, 16) for rb in random_bytes]
 
         return random_bytes
 
-    def randint(self, amax=3, shift=0, repeats=0, length=None):
+    def randint(self, amax=3, shift=0, repeats=0):
         # The default values are for generating the same uniform ternary
         # arrays in all GPUs.
 
         if not isinstance(amax, (list, tuple)):
             amax = [[amax] for share in self.shares]
 
-        if length is None:
-            L = self.L
-        else:
-            L = length
-
         # Calculate shares.
         # If repeats are greater than 0, those channels are
         # subtracted from shares.
@@ -278,7 +268,7 @@ def randint(self, amax=3, shift=0, repeats=0, length=None):
             start_channel = self.shares[devi] - shares[devi]
             end_channel = self.shares[devi] + repeats
             device_states = self.channeled_states[devi][
-                start_channel:end_channel, :L, :
+                start_channel:end_channel, :, :
             ]
             target_states.append(device_states)
 
@@ -289,24 +279,19 @@ def randint(self, amax=3, shift=0, repeats=0, length=None):
 
         return rand_int
 
-    def discrete_gaussian(self, non_repeats=0, repeats=1, length=None):
+    def discrete_gaussian(self, non_repeats=0, repeats=1):
         if not isinstance(non_repeats, (list, tuple)):
             shares = [non_repeats] * self.num_devices
         else:
             shares = non_repeats
 
-        if length is None:
-            L = self.L
-        else:
-            L = length
-
         # Set the target states.
         target_states = []
         for devi in range(self.num_devices):
             start_channel = self.shares[devi] - shares[devi]
             end_channel = self.shares[devi] + repeats
             device_states = self.channeled_states[devi][
-                start_channel:end_channel, :L, :
+                start_channel:end_channel, :, :
             ]
             target_states.append(device_states.view(-1, 16))
 
@@ -327,6 +312,9 @@ def randround(self, coef):
         """Randomly round coef. Coef must be a double tensor.
         coef must reside in the fist GPU in the GPUs list"""
 
+        # The following slicing is OK, since we're using only the first
+        # contiguous stream of states.
+        # It will not make the target state strided.
         L = self.num_coefs // 16
         rand_bytes = chacha20_cuda.chacha20((self.states[0][:L],), self.inc)[
             0

src/liberate/fhe/context/ckks_context.py

@@ -157,7 +157,7 @@ def __init__(
             num_scales=None,
             num_special_primes=2,
             sigma=3.2,
-            uniform_tenary_secret=True,
+            uniform_ternary_secret=True,
             cache_folder=CACHE_FOLDER,
             security_bits=128,
             quantum="post_quantum",
@@ -203,11 +203,11 @@ def __init__(
         self.distribution = distribution
         # Sampling strategy.
         self.sigma = sigma
-        self.uniform_tenary_secret = uniform_tenary_secret
-        if self.uniform_tenary_secret:
-            self.secret_key_sampling_method = "uniform tenary"
+        self.uniform_ternary_secret = uniform_ternary_secret
+        if self.uniform_ternary_secret:
+            self.secret_key_sampling_method = "uniform ternary"
         else:
-            self.secret_key_sampling_method = "sparse tenary"
+            self.secret_key_sampling_method = "sparse ternary"
 
         # dtypes.
         self.torch_dtype = {30: torch.int32, 62: torch.int64}[

src/liberate/fhe/context/security_parameters.py

@@ -15,8 +15,8 @@
 # We separate them in respective dictionaries.
 #
 # Also, there are 3 different methods of sampling the messages according to respective distributions.
-# Those are uniform, error, and (-1, 1) (tenary).
-# We differentiate the message distribution by dictionary keys: 'uniform', 'error', and 'tenary'.
+# Those are uniform, error, and (-1, 1) (ternary).
+# We differentiate the message distribution by dictionary keys: 'uniform', 'error', and 'ternary'.
 
 # This is the pre-quantum security requirements.
 logq_preq = {}
@@ -60,7 +60,7 @@
     613,
     478,
 ]
-logq_preq["tenary"] = [
+logq_preq["ternary"] = [
     27,
     19,
     14,
@@ -123,7 +123,7 @@
     573,
     445,
 ]
-logq_postq["tenary"] = [
+logq_postq["ternary"] = [
     25,
     17,
     13,
@@ -158,7 +158,7 @@ def partitq(q):
 
 # Gather up.
 logq = {}
-distributions = ["uniform", "error", "tenary"]
+distributions = ["uniform", "error", "ternary"]
 logq["pre_quantum"] = {
     distributions[disti]: partitq(logq_preq[dist])
     for disti, dist in enumerate(distributions)
@@ -176,7 +176,7 @@ def minimum_cyclotomic_order(
         "pre_quantum",
         "post_quantum",
     ], "Wrong quantum security model!!!"
-    assert distribution in ["uniform", "error", "tenary"]
+    assert distribution in ["uniform", "error", "ternary"]
     assert security_bits in [128, 192, 256]
 
     x = logq[quantum][distribution][security_bits]
@@ -192,7 +192,7 @@ def maximum_qbits(
         "pre_quantum",
         "post_quantum",
     ], "Wrong quantum security model!!!"
-    assert distribution in ["uniform", "error", "tenary"]
+    assert distribution in ["uniform", "error", "ternary"]
     assert security_bits in [128, 192, 256]
 
     x = cyclotomic_n

src/liberate/fhe/presets/params.py

@@ -3,7 +3,6 @@
         "logN": 14,
         "num_special_primes": 1,
         "devices": [0],
-        "bias_guard": False,
         "scale_bits": 40,
         "num_scales": None,
     },

src/liberate/fhe/tests/test_generate_engine.py

@@ -26,7 +26,7 @@ def ckks_engine(
         "num_special_primes": 2,
         "buffer_bit_length": 62,
         "sigma": 3.2,
-        "uniform_tenary_secret": True,
+        "uniform_ternary_secret": True,
         "cache_folder": "cache/",
         "quantum": "post_quantum",
         "distribution": "uniform",

src/liberate/ntt/rns_partition.py

@@ -56,12 +56,6 @@ def __init__(
         self.num_scales = self.num_ordinary_primes - 1
 
         self.base_prime_idx = self.num_ordinary_primes - 1
-        self.special_prime_idx = list(
-            range(
-                self.num_ordinary_primes + 1,
-                self.num_ordinary_primes + 1 + self.num_special_primes,
-            )
-        )
 
         self.compute_destination_arrays()
         self.compute_rescaler_locations()