Self-Supervised Adversarial Training of Monocular Depth Estimation Against Physical-World Attacks

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Monocular Depth Estimation (MDE) is crucial for autonomous driving but vulnerable to physical attacks.
• Existing adversarial training methods often require ground-truth depth, which is unavailable for MDE.
• Current self-supervised techniques may neglect MDE-specific domain knowledge, limiting effectiveness.

Purpose of the Study:

• To develop a novel self-supervised adversarial training approach for MDE models.
• To improve the robustness of MDE models against real-world physical attacks.
• To address the limitations of existing methods that require ground-truth depth or overlook MDE domain knowledge.

Main Methods:

• Introduced a self-supervised adversarial training strategy for MDE.
• Utilized view synthesis to eliminate the need for ground-truth depth.
• Incorporated L0-norm-bounded perturbation during training to enhance adversarial robustness.

Main Results:

• The proposed method demonstrated improved robustness against diverse adversarial attacks compared to supervised and contrastive learning approaches.
• Achieved enhanced adversarial resilience with minimal degradation in benign performance.
• Validated effectiveness on two representative MDE network architectures.

Conclusions:

• The novel self-supervised adversarial training approach effectively enhances MDE model robustness against physical attacks.
• View synthesis and L0-norm perturbation offer a viable solution for hardening MDE without ground-truth depth.
• This work provides a significant advancement in securing MDE systems for safety-critical applications.