Correct way to use mpc.np_concatenate()

[abraj]

I am getting the error IndexError: tuple index out of range for the following simple code snippet:

secint = mpc.SecInt(32)

arr1 = np.array([secint(1), secint(2), secint(3)])
arr2 = np.array([secint(4), secint(5), secint(6)])

concat_arr = mpc.np_concatenate((arr1, arr2))  # IndexError: tuple index out of range
print(concat_arr)

Am I not using mpc.np_concatenate() correctly?

[lschoe]

Indeed, you need to do things a bit differently.

To start with, do not use the "internal/low-level" function mpc.np_concatenate() in your code at user level. Instead, use the standard Numpy call, like this:

concat_arr = np.concatenate((arr1, arr2))

The MPyC runtime will figure out (under the hood) what to do if any secure arrays are involved. However, arrays arr1and arr2 as you've defined them aren't secure arrays: these are just arrays that happen to have secint elements.

So, what you probably want to write is this:

arr1 = secint.array(np.array([1, 2, 3]))
arr2 = secint.array(np.array([4, 5, 6]))

concat_arr = np.concatenate((arr1, arr2))

If you print this concat_arr you get something like <mpyc.sectypes.ArraySecInt32 object at 0x000001A67053ABC0>.
To see the value in the clear do something like print(await mpc.ouput(concat_arr)).

So, the contents of array concat_arr is secret-shared, but its shape is public, and you can look it up as for any Numpy array using concat_arr.shape which gives (6,) as result.

Take a look at demos like np_id3gini.py and np_lpsolver.py and so on to see the use of secure Numpy arrays "in action".

[abraj]

Got it! Thanks.

Btw, are you suggesting that all mpc.np_* functions are "internal/low-level", and we should avoid using them, if possible.
What about array attributes like z.T? (Is it okay to use them?)
I mean, should np.transpose(z) be used as opposed to z.T or mpc.np_transpose(z) ?

[lschoe]

Sure, that works. You can try it out yourself, but you can also see it used for example in the demo for computing the Moore-Penrose pseudoinverse:

mpyc/demos/pseudoinverse.py

Lines 63 to 78 in 15f747c

	@mpc.coroutine
	async def pseudo_inverse(A):
	"""Return n x m pseudoinverse of given m x n matrix A. """
	m, n = A.shape
	if m > n:
	return pseudo_inverse(A.T).T
	await mpc.returnType((type(A), (n, m)))
	U = await random_square_matrix(type(A), m) # U is a public preconditioner
	A_AT = A @ A.T
	X = U @ (A_AT @ A_AT) @ U.T
	X = reflexive_generalized_inverse(X)[0]
	X = U.T @ X @ U
	X = A.T @ (X @ A_AT)
	return X

Using np.transpose(z) is also fine, but mpc.np_transpose(z) is low-level/internal MPyC code.

[abraj]

Thanks.