GPU-Fuzz: Finding Memory Errors in Deep Learning Frameworks

2026-02-11

AI Summary

This work addresses the critical challenge of GPU memory errors in deep learning frameworks, which can lead to system crashes or security vulnerabilities and demand efficient detection mechanisms. The authors propose a novel approach that integrates formal constraint modeling with fuzz testing: operator parameters are encoded as logical constraints, and a solver is employed to generate test cases that precisely trigger boundary behaviors, thereby systematically exposing memory defects in GPU kernels. This study represents the first integration of formal methods with fuzzing in this context, substantially enhancing both detection efficiency and coverage. Empirical evaluation across PyTorch, TensorFlow, and PaddlePaddle demonstrates the method’s effectiveness and practicality, uncovering 13 previously unknown memory bugs.

Technology Category

• Computer Vision: Adversarial Attacks & Robustness
• Constraint Satisfaction and Optimization: Satisfiability Modulo Theories
• Machine Learning: Deep Neural Architectures and Foundation Models

Application Category

• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Security and Privacy: Large-scale security measurements
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

GPU memory errors are a critical threat to deep learning (DL) frameworks, leading to crashes or even security issues. We introduce GPU-Fuzz, a fuzzer locating these issues efficiently by modeling operator parameters as formal constraints. GPU-Fuzz utilizes a constraint solver to generate test cases that systematically probe error-prone boundary conditions in GPU kernels. Applied to PyTorch, TensorFlow, and PaddlePaddle, we uncovered 13 unknown bugs, demonstrating the effectiveness of GPU-Fuzz in finding memory errors.

Problem

• GPU memory errors
• deep learning frameworks
• memory safety
• bug detection

Innovation

• GPU-Fuzz
• memory error detection
• constraint-based fuzzing
• deep learning frameworks
• GPU kernel testing