Breaking Neural Reasoning Architectures With Metamorphic Relation-Based Adversarial Examples.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning

Background:

• Neural reasoning architectures aim to replicate human-like reading, reasoning, and inference.
• Differentiable Neural Computers (DNCs) are advanced models designed for logical reasoning beyond surface-level text matching.
• Text-based question answering (QA) is a key benchmark for evaluating AI reasoning capabilities.

Purpose of the Study:

• To investigate the logical reasoning abilities of DNCs in text-based QA tasks.
• To introduce a novel adversarial attack method to probe DNC vulnerabilities.
• To assess the robustness of current neural reasoning architectures against sophisticated attacks.

Main Methods:

• Development of a conceptually simple yet effective adversarial attack leveraging metamorphic relations.
• Application of the adversarial attack to state-of-the-art DNC models.
• Empirical exploration of defense strategies against the proposed adversarial attack.

Main Results:

• The adversarial attack drastically reduced DNC accuracy from 100% to a worst-case 1.5%.
• The study identified significant weaknesses and susceptibilities in modern neural reasoning architectures.
• The proposed adversarial framework proved effective in improving model adversarial robustness.

Conclusions:

• Modern neural reasoning architectures, including DNCs, exhibit critical vulnerabilities to adversarial attacks.
• Metamorphic relation-based adversarial attacks are a potent tool for evaluating AI reasoning.
• The developed adversarial framework offers a scalable method for enhancing the robustness of AI models.