New Adversarial Image Detection Based on Sentiment Analysis.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning Security

Background:

• Deep neural networks (DNNs) are susceptible to adversarial examples, which are subtly modified inputs designed to cause misclassification.
• Existing adversarial attack models are advancing rapidly, often outpacing current detection methods.
• There is a critical need for robust and efficient detection techniques to ensure the security of DNNs.

Purpose of the Study:

• To develop a novel and effective adversarial example detector for deep neural networks.
• To improve the detection rate of the latest adversarial attacks on image datasets.
• To propose a new approach utilizing sentiment analysis for adversarial example detection.

Main Methods:

• Proposed a novel detector leveraging sentiment analysis to identify adversarial examples.
• Developed a modularized embedding layer to convert DNN hidden-layer feature maps into word vectors for sentiment analysis.
• Evaluated the detector's performance against state-of-the-art algorithms on CIFAR-10, CIFAR-100, and SVHN datasets using ResNet and Inception networks.

Main Results:

• The proposed detector consistently outperformed state-of-the-art detection algorithms in identifying the latest adversarial attacks.
• Demonstrated the effectiveness of using sentiment analysis on feature map perturbations for detection.
• Achieved high detection accuracy with a compact model (approx. 2 million parameters) and fast inference time (less than 4.6 ms).

Conclusions:

• The novel sentiment analysis-based detector offers a significant advancement in identifying adversarial examples against DNNs.
• The method provides a robust and efficient solution for detecting sophisticated adversarial attacks on image datasets.
• This approach shows promise for enhancing the security and reliability of deep learning models in real-world applications.