Odds#

class scio.scores.Odds(*, act_norm=None, mode='raw', epsilon=?, noise_samples=256, z_aggregation='sum', rng_seed=0)[source]#

Bases: BaseScoreClassif

Odds for classification.

Parameters:

  • epsilon (float) – \(L^{\infty}\) scale of uniformly sampled random perturbations.

  • noise_samples (int) – Number of random samples to compute log-preferences. Defaults to 2**8.

  • z_aggregation (ZAggrLike) – See ZAggr. Defaults to "sum".

  • batch_size (int) –

    Currently Not Implemented!

    Necessary for computational reasons. Data and env-dependent. Use 0 for unlimited. Defaults to 2**13.

  • rng_seed (int, optional) – If provided, manual seed for torch.Generator, used during calibration (in which case the random state is reset at every fit() call). Defaults to 0.

  • mode – See BaseScoreClassif.

  • act_norm – See BaseScore.

Warning

Current implementation is unbatched. This means it does not scale well for large number of random samples.

References

[RKH19]

Kevin Roth, Yannic Kilcher, and Thomas Hofmann. The Odds are Odd: a statistical test for detecting adversarial examples. In Proceedings of the 36th International Conference on Machine Learning, volume 97, 5498–5507. 2019. URL: https://proceedings.mlr.press/v97/roth19a.html.

Hint

Below this point, the documentation is meant for development purposes only. Manual use of any listed member is highly discouraged. For usage, see Inferring with Confidence.

Useful methods defined here

logits_pref(samples, reference_idx)

Compute logits difference to reference index.

uniform(*size, to)

Sample uniformly random Tensor in \([-1, 1]^{size}\).
static logits_pref(samples, reference_idx)[source]#

Compute logits difference to reference index.

In [RKH19] notations, corresponds to \(f_{y, z}\) for all \(z\) given samples \(f\) and reference_idx \(y\).

Parameters:

  • samples (Tensor) – Logits. Shape (n_samples, n_classes).

  • reference_idx (int) – Index of the reference to subtract.

uniform(*size, to)[source]#

Sample uniformly random Tensor in \([-1, 1]^{size}\).

The output device and dtype are those of to. The same applies to its shape if size if empty.

Parameters:

  • *size (int) – If provided, unpacked target size.

  • to (Tensor) – Tensor defining the target dtype and device. Also defines the target size if size is empty.