![[2602.15756] A Note on Non-Composability of Layerwise Approximate Verification for Neural Inference](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.15756] A Note on Non-Composability of Layerwise Approximate Verification for Neural Inference
Summary
This paper discusses the limitations of layerwise approximate verification in neural inference, presenting a counterexample that challenges the assumption that verifying each layer ensures the correctness of the final output.
Why It Matters
Understanding the non-composability of layerwise verification is crucial for enhancing the reliability of machine learning models, especially in security-sensitive applications. This insight can inform future research and development in AI safety and verification methodologies.
Key Takeaways
- Layerwise verification does not guarantee overall output correctness.
- A counterexample demonstrates potential vulnerabilities in neural networks.
- Adversarial errors can significantly affect final inference results.
Computer Science > Cryptography and Security arXiv:2602.15756 (cs) [Submitted on 17 Feb 2026] Title:A Note on Non-Composability of Layerwise Approximate Verification for Neural Inference Authors:Or Zamir View a PDF of the paper titled A Note on Non-Composability of Layerwise Approximate Verification for Neural Inference, by Or Zamir View PDF HTML (experimental) Abstract:A natural and informal approach to verifiable (or zero-knowledge) ML inference over floating-point data is: ``prove that each layer was computed correctly up to tolerance $\delta$; therefore the final output is a reasonable inference result''. This short note gives a simple counterexample showing that this inference is false in general: for any neural network, we can construct a functionally equivalent network for which adversarially chosen approximation-magnitude errors in individual layer computations suffice to steer the final output arbitrarily (within a prescribed bounded range). Subjects: Cryptography and Security (cs.CR); Machine Learning (cs.LG) Cite as: arXiv:2602.15756 [cs.CR] (or arXiv:2602.15756v1 [cs.CR] for this version) https://doi.org/10.48550/arXiv.2602.15756 Focus to learn more arXiv-issued DOI via DataCite (pending registration) Submission history From: Or Zamir [view email] [v1] Tue, 17 Feb 2026 17:41:59 UTC (6 KB) Full-text links: Access Paper: View a PDF of the paper titled A Note on Non-Composability of Layerwise Approximate Verification for Neural Inference, by Or ZamirView PDFHT...
