Hybrid reverse knowledge distillation for adversarial example detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks excel in computer vision but are susceptible to adversarial examples.
• Adversarial examples are subtle perturbations causing AI misclassification, posing risks in critical applications like medical diagnosis and autonomous driving.

Purpose of the Study:

• To propose a novel and effective method for detecting adversarial examples in deep neural networks.
• To enhance the reliability of AI systems by identifying malicious inputs.

Main Methods:

• Developed Hybrid Reverse Knowledge Distillation (Hybrid RKD) using a multi-scale decoder to reconstruct features from clean images.
• Trained the decoder on a frozen teacher encoder using only normal data to identify feature-level distortions caused by adversarial perturbations.
• Integrated Mahalanobis-style statistical distance metrics to detect distribution-level anomalies.

Main Results:

• Achieved state-of-the-art detection performance against diverse adversarial attacks (FGSM, PGD, DeepFool, C&W, AutoAttack) on CIFAR-10 and ISIC2018 datasets.
• Obtained average AUROC scores of 0.790 and 0.929 across five attack categories.
• Demonstrated superior performance compared to existing methods like LID, Mahalanobis, and ODIN.

Conclusions:

• Hybrid RKD offers a robust solution for adversarial example detection.
• The method effectively identifies adversarial perturbations by leveraging feature reconstruction and statistical anomaly detection.
• This approach does not require adversarial examples during training, making it practical for real-world deployment.