Precise Verification of Transformers through ReLU-Catalyzed Abstraction Refinement

2026-05-13

AI Summary

Transformers are widely deployed in safety-critical applications, yet the dot-product operations in their self-attention mechanisms hinder formal verification by yielding low precision and high false-positive rates. This work addresses this challenge by introducing an innovative approach that leverages ReLU functions to construct tight nonlinear bounds for dot products. Building on this insight, it adapts well-established ReLU convex relaxation techniques—previously used in feedforward networks—to the verification of Transformers, yielding two efficient frameworks: one rule-driven and the other optimization-driven. The proposed method substantially improves verification accuracy while maintaining acceptable computational overhead. Experimental results on sentiment analysis benchmarks demonstrate that the approach outperforms state-of-the-art verification methods across most evaluation metrics.

Abstract

Formal verification of transformers has become increasingly important due to their widespread deployment in safety-critical applications. Compared to classic neural networks, the inferences of transformers involve highly complex computations, such as dot products in self-attention layers, rendering their verification extremely difficult. Existing approaches explored over-approximation methods by constructing convex constraints to bound the output ranges of transformers, which can achieve high efficiency. However, they may sacrifice verification precision, and consequently introduce significant approximation error that leads to frequent occurrences of false alarms. In this paper, we propose a transformer verification approach that can achieve improved precision. At the core of our approach is a novel usage of ReLU, by which we represent a precise but non-linear bound for dot products such that we can further exploit the rich body of literature for convex relaxation of ReLU to derive precise bounds. We extend two classic approaches to the context of transformers, a rule-based one and an optimization-based one, resulting in two new frameworks for efficient and precise verification. We evaluate our approaches on different model architectures and robustness properties derived from two datasets about sentiment analysis, and compare with the state-of-the-art baseline approach. Compared to the baseline, our approach can achieve significant precision improvement for most of the verification tasks with acceptable compromise of efficiency, which demonstrates the effectiveness of our approach.

Problem

• transformer verification
• formal verification
• approximation error
• self-attention
• false alarms

Innovation

• ReLU-catalyzed abstraction
• transformer verification
• dot product bounding
• convex relaxation
• formal verification