Skip to main content

omnibias.verify

Certified neural-network verification. Propagates rigorous enclosures (intervals / multivariate Taylor models from omnibias.core.verified) through a trained network, with branch-and-bound over the input box, to produce theorem-grade certificates.

import omnibias.verify as ov
Sound, not complete

Every enclosure provably contains the true output set (sound by construction). The procedure is not complete — branch-and-bound may return UNKNOWN rather than over-claim.

Network model

SymbolRole
Network, Layer, VerifiedLayerthe verifiable network model
LinearLayer, ReLULayer, GELULayer, TanhLayer, SigmoidLayer, MaxPoolLayerlayer types
IntervalMatrixinterval-valued matrix
SUPPORTED_ACTIVATIONS, KERNELScapability tables
verified_layers, verified_layer_bundle, VerifiedLayerBundleweight ingestion bundles

Weight ingestion from the differentiable backends lives in omnibias.verify.torch / omnibias.verify.jax (import explicitly).

Propagation

interval_propagate · taylor_propagate · interval_jacobian · output_range · taylor_output_bounds · linear_image · affine_layer · reachable_box · scalar_readout_range.

Enclosures

relu_enclosure · relu_taylor · gelu_enclosure · max_enclosure · gauss_cdf_iv · compose_sigma · compose_gelu · jet_remainder_iv.

Certificates

SymbolCertifies
certify_robustnessRobustnessCertificatelocal robustness of a classifier
lipschitz_bounda certified Lipschitz constant
monotonicityMonotonicityCertificatemonotonicity in chosen inputs
certify_no_dead_unitno permanently dead unit
certify_agreement_margin, agreement_margin_ivagreement margins
certify_pinn_aposterioriCertifiedPDEPipelineResultan a-posteriori PINN residual bound

Result containers: BoundResult, RangeResult, CertifiedBound.

Example

from omnibias.verify import certify_robustness
cert = certify_robustness(network, x0, epsilon=0.03)
print(cert) # holds / UNKNOWN, with the enclosing bounds

See also