syft.frameworks.torch.mpc.securenn

This is an implementation of the SecureNN paper https://eprint.iacr.org/2018/442.pdf

Note that there is a difference here in that our shares can be negative numbers while they are always positive in the paper

Module Contents

syft.frameworks.torch.mpc.securenn.p = 67
syft.frameworks.torch.mpc.securenn.Q_BITS = 62
syft.frameworks.torch.mpc.securenn.no_wrap
syft.frameworks.torch.mpc.securenn.decompose(tensor)

decompose a tensor into its binary representation.

syft.frameworks.torch.mpc.securenn.flip(x, dim)

Reverse the order of the elements in a tensor

syft.frameworks.torch.mpc.securenn._random_common_bit(*workers)

Return a bit chosen by a worker and sent to all workers, in the form of a MultiPointerTensor

syft.frameworks.torch.mpc.securenn._random_common_value(max_value, *workers)

Return n in [0, max_value-1] chosen by a worker and sent to all workers, in the form of a MultiPointerTensor

syft.frameworks.torch.mpc.securenn._shares_of_zero(size, field, crypto_provider, *workers)

Return n in [0, max_value-1] chosen by a worker and sent to all workers, in the form of a MultiPointerTensor

syft.frameworks.torch.mpc.securenn.select_share(alpha_sh, x_sh, y_sh)

Performs select share protocol If the bit alpha_sh is 0, x_sh is returned If the bit alpha_sh is 1, y_sh is returned

Parameters
Returns

z_sh = (1 - alpha_sh) * x_sh + alpha_sh * y_sh

syft.frameworks.torch.mpc.securenn.private_compare(x_bit_sh, r, beta)

Perform privately x > r

Parameters

  • x (AdditiveSharedTensor) – the private tensor

  • r (MultiPointerTensor) – the threshold commonly held by alice and bob

  • beta (MultiPointerTensor) – a boolean commonly held by alice and bob to hide the result of computation for the crypto provider

Returns

β′ = β ⊕ (x > r).

syft.frameworks.torch.mpc.securenn.msb(a_sh)

Compute the most significant bit in a_sh, this is an implementation of the SecureNN paper https://eprint.iacr.org/2018/442.pdf

Parameters

a_sh (AdditiveSharingTensor) – the tensor of study

Returns

the most significant bit

syft.frameworks.torch.mpc.securenn.share_convert(a_sh)

Convert shares of a in field L to shares of a in field L - 1

Parameters

a_sh (AdditiveSharingTensor) – the additive sharing tensor who owns the shares in field L to convert

Returns

An additive sharing tensor with shares in field L-1

syft.frameworks.torch.mpc.securenn.relu_deriv(a_sh)

Compute the derivative of Relu

Parameters

a_sh (AdditiveSharingTensor) – the private tensor on which the op applies

Returns

0 if Dec(a_sh) < 0 1 if Dec(a_sh) > 0 encrypted in an AdditiveSharingTensor

syft.frameworks.torch.mpc.securenn.relu(a_sh)

Compute Relu

Parameters

a_sh (AdditiveSharingTensor) – the private tensor on which the op applies

Returns

Dec(a_sh) > 0 encrypted in an AdditiveSharingTensor

syft.frameworks.torch.mpc.securenn.division(x_sh, y_sh, bit_len_max=Q_BITS // 2)

Performs division of encrypted numbers

Parameters

  • y_sh (x_sh,) – the private tensors on which the op applies

  • bit_len_max – the number of bits needed to represent the highest value in the tensors we may want to avoid giving this value so Q_BITS is the default value

Returns

element-wise integer division of x_sh by y_sh

syft.frameworks.torch.mpc.securenn.maxpool(x_sh)

Compute MaxPool: returns fresh shares of the max value in the input tensor and the index of this value in the flattened tensor

Parameters

x_sh (AdditiveSharingTensor) – the private tensor on which the op applies

Returns

maximum value as an AdditiveSharingTensor index of this value in the flattened tensor as an AdditiveSharingTensor

syft.frameworks.torch.mpc.securenn.maxpool_deriv(x_sh)

Compute derivative of MaxPool

Parameters

x_sh (AdditiveSharingTensor) – the private tensor on which the op applies

Returns

an AdditiveSharingTensor of the same shape as x_sh full of zeros except for a 1 at the position of the max value

syft.frameworks.torch.mpc.securenn.maxpool2d(a_sh, kernel_size: int = 1, stride: int = 1, padding: int = 0)

Applies a 2D max pooling over an input signal composed of several input planes. This interface is similar to torch.nn.MaxPool2D. :param kernel_size: the size of the window to take a max over :param stride: the stride of the window :param padding: implicit zero padding to be added on both sides