Skip to main content
Back

TU Munich's EnsembleCert Certifies 26.5% More Label Poisoning Attacks With 100x Fewer Models

A new certification framework proves neural networks stay correct under label poisoning, delivering tighter guarantees with 100x fewer ensemble partitions.

Pro

  • EnsembleCert is the first white-box certification framework for partition-aggregation ensembles against label poisoning: paper.
  • ScaLabelCert delivers the first polynomial-time exact certificate for sufficiently wide neural networks against label-flipping attacks.
  • On CIFAR-10 it certifies +26.5% more label flips in median while using 100 times fewer partitions than SS-DPA.
  • Uses the neural tangent kernel equivalence to reduce LabelCert's NP-hard MILP to a sorting or knapsack problem.
  • A single wide network certified by ScaLabelCert can outperform its partition-aggregation ensemble on CIFAR-10.
  • Code and configs are available on GitHub.

Poisoning the labels of a training set is one of the cheapest ways to sabotage a supervised model. Flip a few classes here and there, and predictions at inference time can quietly flip too. A team from TU Munich has published EnsembleCert and ScaLabelCert, a pair of methods that give mathematically guaranteed answers to a specific question: how many labels would an attacker have to corrupt before the model changes its mind about a given test example?

On CIFAR-10, the approach certifies up to 26.5% more label flips in median across the test set than the existing black-box method, while using 100 times fewer partitions. That challenges the prevailing assumption that heavy partitioning is required for strong certified robustness. Code is on GitHub.

Cracking open the black box

Most current defenses against label poisoning work through ensembling. Existing certification frameworks lean on techniques such as smoothing or partition-aggregation, but treat the corresponding base classifiers as black boxes, yielding overly conservative guarantees. The dominant approach, Deep Partition Aggregation (DPA), splits the training data into many disjoint partitions, trains one base classifier per partition, and takes a majority vote at inference.

The certificate then rests on a worst-case assumption: a single label flip in a partition can flip that partition's vote. That is usually pessimistic, because a well-trained base classifier often needs many flips before it changes its prediction, but the black-box view has no way of knowing that. Teams have compensated by using enormous numbers of partitions, which means training a prohibitively large number of small models.

What EnsembleCert changes