Programmatic interface to SHEEP
This library is a programmatic interface in python to generate a circuit for the bigger and more useful SHEEP library.
The library has a few data types :
- variables - A single bit (Could also be used as a normal scalar)
- enc_vec - One dimensional bit vector (Could be used a one dimensional vector of any data type)
- enc_mat - Two dimensional bit matrix (Could be used a one dimensional vector of any data type)
- enc_tensor3 - Three dimensional bit tensor.
To create a circuit, the basic class to inherit is mini_mod in mathsheep.interactions. To add more components, you can use self.add(component) inside the create function as shown below.:
class oneb_adder(mini_mod): def __init__(self, name, inputs, outputs, nb=None, randomize_temps=1, carry=True): mini_mod.__init__(self, name, inputs, outputs) self.create(...) def create(self, ...): self.add(..)
Two types of components can be added.
Assignments (from matSHEEP.interactions)
There are a few predefined mini_mods. They can be found in
- oneb_adder - Add two bits
- nb_adder - Adders x and y with incoming carrt where input is [cin x y]
- nb_adder_xy - Adds x and y with input = (x, y)
- compare_cp - Compares ciphertext with plaintext with input = (c,p)
- reduce_add - Counts the number of ones in a bit vector.
- linear_layer_1d - Inner Product of a weight vector with encrypted bit vector followed by a sign function.
- linear_layer - Inner Product of a weight matrix with an encrypted bit vector followed by a sign function.
- conv_layer - A convolution Layer. (Look at examples)
- vec_mono_op_cond - Takes a plaintext cond vector, a plaintext tuple ass_types containing only alias and negate as values and an encrypted bit vector input. It outputs an encrypted bit vector where the ith position has the ass_types[cond[idx]] operation applied on input[idx].
- Similar operation for matrix and tensor.
You can also visualize the circuits you create. test.sheep is a circuit file.:
import sys import matSHEEP.create_graph as cg complete_node = cg.get_circuit_graph('./test.sheep') ng = cg.networkx_graph(complete_node) ng.draw()
And you can get
gFor more high level operations and results using layers of Neural Networks visit this markdown