Improving Robustness of Intent Detection Under Adversarial Attacks: A Geometric Constraint Perspective.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning

Background:

• Deep neural networks (DNNs) for natural language processing (NLP) are susceptible to adversarial examples.
• Existing defense methods often use linear classifiers and softmax, which do not promote well-separated feature representations, making models vulnerable.
• Intent detection in dialog systems, a key NLP task, lacks robust defense mechanisms against such attacks.

Purpose of the Study:

• To propose a simple yet efficient defense method to enhance the robustness of NLP systems against adversarial examples.
• To improve the feature representation learning in DNNs for better discrimination between classes.
• To address the limitations of current defense strategies in intent detection tasks.

Main Methods:

• Introduced an M-similarity metric to reduce the variance of intraclass features, aiming for better geometric conditions in the feature space.
• Derived optimal geometric constraints for category anchors based on overall misclassification probability (OMP) with theoretical guarantees.
• Formulated the geometric constraints as a manifold optimization problem on the Stiefel manifold to overcome traditional optimization challenges.

Main Results:

• The proposed method significantly improves the robustness of NLP models against adversarial examples.
• The approach maintains excellent performance on non-adversarial (normal) examples.
• Experimental results show superior performance compared to existing baseline defense methods.

Conclusions:

• The proposed geometric constraint defense method offers a promising solution for enhancing the resilience of NLP systems, particularly in intent detection.
• Optimizing feature space geometry through manifold optimization is an effective strategy against adversarial attacks.
• This work contributes to building more reliable and secure dialog systems.