Defense Against Explanation Manipulation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Model Interpretability

Background:

• Explainable machine learning (XAI) is crucial for trust in AI applications.
• Current explanation methods are vulnerable to manipulation, where explanations can be altered without changing model predictions.
• Existing solutions focus on stable methods or model configuration changes.

Purpose of the Study:

• To introduce a new training scheme, Adversarial Training on EXplanations (ATEX), to enhance internal explanation stability.
• To improve the robustness of machine learning models against explanation manipulation.
• To investigate the relationship between explanation stability and adversarial attack vulnerability.

Main Methods:

• Developed Adversarial Training on EXplanations (ATEX), a novel training scheme.
• ATEX imposes constraints on model predictions rather than direct explanation values, avoiding second-order derivatives.
• Evaluated ATEX's effectiveness in improving explanation stability and model robustness.

Main Results:

• ATEX significantly improves the internal explanation stability of machine learning models.
• The proposed method enhances model robustness against targeted explanation manipulation.
• ATEX demonstrates benefits such as explanation smoothing and improved adversarial training efficacy.

Conclusions:

• Adversarial Training on EXplanations (ATEX) offers a robust solution for enhancing explanation stability in machine learning.
• The training scheme improves model transparency and security, making AI more trustworthy.
• Explanation stability is linked to adversarial robustness, with ATEX providing dual benefits.