Verified Relative Output Margins for Neural Network Twins

Anahita Baninajjar, Kamran Hosseini, Ahmed Rezine, Amir Aminifar

Keywords: Relative Output Margin, Formal Verification, Deep Neural Networks

TL;DR: This framework compares two neural networks by quantifying relative output margins and providing provably-correct bounds on decision quality across an input region.

Abstract: Given two neural network classifiers with the same input and output domains, our goal is to compare the two networks in relation to each other over an entire input region (e.g., within a vicinity of an input sample). Towards this, we introduce and quantify the Relative Output Margin (ROM) with which decisions are made. A larger output margin for a network w.r.t. another indicates that this network consistently makes a correct decision every time the other network does, and it does so in the entire input region. More importantly, as opposed to best-effort testing schemes, our framework is able to establish provably-correct (formally verified) bounds on ROM gains/losses over an entire input region. The proposed framework is relevant in the context of several application domains, e.g., for comparing a trained network and its corresponding compact (e.g., pruned, quantized, distilled) network. We evaluate our framework using the MNIST, CIFAR10, and two real-world medical datasets, to show its relevance.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

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: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

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.

Submission Number: 9665