declearn.main.config.PrivacyConfig

Dataclass wrapping parameters to set up local differential privacy.

The parameters wrapped by this class specify the DP-SGD algorithm [1], providing with a budget, an accountant method, a sensitivity clipping threshold, and RNG-related parameters for the noise-addition module.

Accountants supported by Opacus 1.2.0 include:

• rdp : Renyi-DP accountant, see [1]
• gdp : Gaussian-DP, see [2]
• prv : Privacy loss Random Variables privacy accountant, see [3]

Note: for more details, refer to the Opacus source code and the doctrings of each accountant. See https://github.com/pytorch/opacus/tree/main/opacus/accountants

Attributes:

Name	Type	Description
budget	float, float	Target total privacy budget per client, expressed in terms of (epsilon-delta)-DP over the full training schedule.
accountant	str	Accounting mechanism used to estimate epsilon by Opacus.
sclip_norm	float	Clipping threshold of sample-wise gradients' euclidean norm. This parameter binds the sensitivity of sample-wise gradients.
use_csprng	bool	Whether to use cryptographically-secure pseudo-random numbers (CSPRNG) rather than the default numpy generator. This is significantly slower than using the default numpy RNG.
seed	int or None	Optional seed to the noise-addition module's RNG. Unused if safe_mode=True.

References

• [1] Abadi et al, 2016. Deep Learning with Differential Privacy. https://arxiv.org/abs/1607.00133
• [2] Dong et al, 2019. Gaussian Differential Privacy. https://arxiv.org/abs/1905.02383
• [3] Gopi et al, 2021. Numerical Composition of Differential Privacy. https://arxiv.org/abs/2106.02848

Source code in declearn/main/config/_dataclasses.py

@dataclasses.dataclass
class PrivacyConfig:
    """Dataclass wrapping parameters to set up local differential privacy.

    The parameters wrapped by this class specify the DP-SGD algorithm [1],
    providing with a budget, an accountant method, a sensitivity clipping
    threshold, and RNG-related parameters for the noise-addition module.

    Accountants supported by Opacus 1.2.0 include:

    * rdp : Renyi-DP accountant, see [1]
    * gdp : Gaussian-DP, see [2]
    * prv : Privacy loss Random Variables privacy accountant, see [3]

    Note: for more details, refer to the Opacus source code and the
    doctrings of each accountant. See
    https://github.com/pytorch/opacus/tree/main/opacus/accountants

    Attributes
    ----------
    budget: (float, float)
        Target total privacy budget per client, expressed in terms of
        (epsilon-delta)-DP over the full training schedule.
    accountant: str
        Accounting mechanism used to estimate epsilon by Opacus.
    sclip_norm: float
        Clipping threshold of sample-wise gradients' euclidean norm.
        This parameter binds the sensitivity of sample-wise gradients.
    use_csprng: bool
        Whether to use cryptographically-secure pseudo-random numbers
        (CSPRNG) rather than the default numpy generator.
        This is significantly slower than using the default numpy RNG.
    seed: int or None
        Optional seed to the noise-addition module's RNG.
        Unused if `safe_mode=True`.

    References
    ----------
    - [1]
        Abadi et al, 2016.
        Deep Learning with Differential Privacy.
        https://arxiv.org/abs/1607.00133
    - [2]
        Dong et al, 2019.
        Gaussian Differential Privacy.
        https://arxiv.org/abs/1905.02383
    - [3]
        Gopi et al, 2021.
        Numerical Composition of Differential Privacy.
        https://arxiv.org/abs/2106.02848
    """

    budget: Tuple[float, float]
    sclip_norm: float
    accountant: str = "rdp"
    use_csprng: bool = False
    seed: Optional[int] = None

    def __post_init__(self):
        """Type- and value-check (some of) the wrapped parameters."""
        # Verify budget validity.
        if isinstance(self.budget, list):
            self.budget = tuple(self.budget)
        if not (
            isinstance(self.budget, tuple)
            and len(self.budget) == 2
            and isinstance(self.budget[0], (float, int))
            and self.budget[0] > 0
            and isinstance(self.budget[1], (float, int))
            and self.budget[1] >= 0
        ):
            raise TypeError("'budget' should be a tuple of positive floats.")
        # Verify max_norm validity.
        if not (
            isinstance(self.sclip_norm, (float, int)) and self.sclip_norm > 0
        ):
            raise TypeError("'sclip_norm' should be a positive float.")
        # Verify accountant validity.
        accountants = ("rdp", "gdp", "prv")
        if self.accountant not in accountants:
            raise TypeError(f"'accountant' should be one of {accountants}")

__post_init__()

Type- and value-check (some of) the wrapped parameters.

Source code in declearn/main/config/_dataclasses.py

def __post_init__(self):
    """Type- and value-check (some of) the wrapped parameters."""
    # Verify budget validity.
    if isinstance(self.budget, list):
        self.budget = tuple(self.budget)
    if not (
        isinstance(self.budget, tuple)
        and len(self.budget) == 2
        and isinstance(self.budget[0], (float, int))
        and self.budget[0] > 0
        and isinstance(self.budget[1], (float, int))
        and self.budget[1] >= 0
    ):
        raise TypeError("'budget' should be a tuple of positive floats.")
    # Verify max_norm validity.
    if not (
        isinstance(self.sclip_norm, (float, int)) and self.sclip_norm > 0
    ):
        raise TypeError("'sclip_norm' should be a positive float.")
    # Verify accountant validity.
    accountants = ("rdp", "gdp", "prv")
    if self.accountant not in accountants:
        raise TypeError(f"'accountant' should be one of {accountants}")