AGREE: A Simple Aggregator of Detectors’ Decisions

Federica Granese, Marco Romanelli, Marine Picot, Francisco Messina, Pablo Piantanida

Keywords: AI Safety, Algorithms Evaluation, Deep Learning, Adversarial Examples

TL;DR: We propose a simple yet effective method to aggregate the decisions based on the soft-probability outputs of multiple trained detectors, possibly provided by a third party.

Abstract:

A simple yet effective method to aggregate the decisions based on the soft-probability outputs of multiple trained detectors, possibly provided by a third party, is introduced. We formally derive a mathematically sound theoretical framework, which is straightforward as it does not require further training of the given detectors, and modular, allowing existing (and future) detectors to be merged into a single one. As an application, we evaluate our framework by tackling the recently proposed problem of simultaneous adversarial examples detection, i.e. when the attacks at the evaluation time can be simultaneously crafted according to a variety of algorithms and objective loss functions. While each single detector tends to underperform or fail in the aforementioned attack scenario,
our framework successfully aggregates the knowledge of the available detectors to guarantee a more reliable decision.
We validate our AGgregatoR of dEtectors' dEcisions (Agree) on popular datasets (e.g., CIFAR10 and SVHN) and we show that it consistently outperforms the state-of-the-art when simultaneous adversarial attacks are present at evaluation time.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)