Creates aggregator with adaptive zeroing and distributed DP.
bits: int = 20,
zeroing: bool = True,
rotation_type: str = 'hd'
Zeroes out extremely large values for robustness to data corruption on
clients, and performs distributed DP (compression, discrete noising, and
SecAgg) with adaptive clipping for differentially private learning. For
details of the two main distributed DP algorithms see
or https://arxiv.org/pdf/2110.04995.pdf The adaptive clipping uses the
geometric method described in https://arxiv.org/abs/1905.03871
A float specifying the noise multiplier (with respect to
the initial L2 cipping) for the distributed DP mechanism for model
updates. A value of 1.0 or higher may be needed for meaningful privacy.
An integer specifying the expected number of
clients per round. Must be positive.
An integer specifying the bit-width for the aggregation. Note that
this is for the noisy, quantized aggregate at the server and thus should
account for the
expected_clients_per_round. Must be in the inclusive
range of [1, 22]. This is set to 20 bits by default, and it dictates the
computational and communication efficiency of Secure Aggregation. Setting
it to less than 20 bits should work fine for most cases. For instance, for
an expected number of securely aggregated client updates of 100, 12 bits
should be enough, and for an expected number of securely aggregated client
updates of 1000, 16 bits should be enough.
A bool indicating whether to enable adaptive zeroing for data
corruption mitigation. Defaults to
A string indicating what rotation to use for distributed DP.
Valid options are 'hd' (Hadamard transform) and 'dft' (discrete Fourier
transform). Defaults to