Proposal of Deep Learning Systems Testing Framework Using Generative Large Language Models

Abstract

Ensuring the robustness and reliability of deep learning systems is essential as they are used in critical applications like autonomous driving, medical diagnostics, and financial forecasting. Traditional testing methods often miss rare or adversarial scenarios that can cause failures. This proposal leverages generative large language models (LLMs), such as GPT-4, to generate adversarial inputs for testing deep learning systems. LLMs can produce diverse and contextually relevant inputs, enhancing the detection of vulnerabilities and improving the robustness of these models. The research aims to develop an LLM-based adversarial input generation framework, evaluate its effectiveness, compare it with traditional methods, and assess improvements in system robustness. This approach promises to advance AI robustness testing and ensure deep learning systems are more resilient in real-world applications.