Dual-Domain Adaptive Input Perturbation Sensitivity for Adversarial Example Detection

Area of Science:

• Computer Vision
• Machine Learning Security
• Deep Neural Networks

Background:

• Vision-sensor-based systems are crucial for safety-critical applications like autonomous driving and surveillance.
• Deep neural networks (DNNs) are vulnerable to adversarial examples, posing significant security risks.
• Existing detection methods struggle with heterogeneous attacks and high-class density tasks due to limitations in perturbation scaling and metric analysis.

Purpose of the Study:

• To develop an advanced adversarial example detection method for vision systems.
• To address the limitations of current input-level sensitivity detection methods.
• To enhance the robustness and security of intelligent perception systems against sophisticated attacks.

Main Methods:

• Proposed a dual-domain adaptive adversarial example detection method based on Multi-scale Input Sensitivity (MSIS).
• Introduced a Manifold-Motivated Micro-scale Probing (MMP) mechanism for effective boundary probing and perturbation scaling.
• Implemented a Dual-State Sensitivity Fusion (DSF) mechanism using temperature scaling for robust detection against heterogeneous attacks.

Main Results:

• Achieved high detection performance against FGSM, PGD, and C&W attacks on CIFAR-10 and CIFAR-100 datasets without model modification.
• Demonstrated superior detection AUC (up to 99.14%) compared to non-intrusive baselines and even intrusive methods like PASA.
• Showcased significantly lower false positive rates and reduced inference latency compared to state-of-the-art methods.

Conclusions:

• The proposed MSIS method effectively enhances adversarial example detection in non-intrusive benchmark settings.
• Task-level spatial-domain scaling and temperature-state adaptation are key to improving output-probability-based detection.
• The findings provide strong algorithmic evidence for securing visual classification tasks against adversarial perturbations.