Robust Adversarial Attacks Detection for Deep Learning based Relative Pose Estimation for Space Rendezvous

近年来，针对自主航天器相对导航挑战的深度学习技术的研究不断增长。采用这些技术可以提高性能。然而，这种方法也引入了对这些深度学习方法的可信度和安全性的高度担忧，因为它们容易受到对抗性攻击的影响。在这项工作中，我们提出了一种基于可解释性概念的深度神经网络相对位姿估计方案的对抗攻击检测的新方法。我们针对轨道交会场景开发了一种创新的相对位姿估计技术，采用我们提出的卷积神经网络（CNN），它可以从追逐者的机载摄像头拍摄的图像中准确地输出目标的相对位置和旋转。我们使用由快速梯度符号方法（FGSM）生成的对抗性攻击无缝扰动输入图像。然后，基于长短期记忆（LSTM）网络构建了对抗攻击检测器，该检测器采用CNN基于位姿估计器的可解释性度量（即Shapley值），并在执行时标记检测到的对抗攻击。仿真结果表明，所提出的对抗攻击检测器的检测准确度达到了99.21％。深度相对位姿估计器和对抗攻击检测器都在我们设计的实验室环境中进行了测试。实验结果表明，所提出的对抗攻击检测器在平均检测准确度方面达到了96.29％。

The paper proposes a novel approach for adversarial attack detection for deep neural network-based relative pose estimation schemes in autonomous spacecraft navigation. The problem addressed is the susceptibility of deep learning methods to adversarial attacks and the need for a reliable detection system to ensure the safety and security of autonomous spacecraft.

The key idea of the paper is to use the explainability concept to develop an adversarial attack detector based on a Long Short Term Memory (LSTM) network. The detector takes the explainability measure from a Convolutional Neural Network (CNN)-based pose estimator and flags the detection of adversarial attacks when acting. The proposed approach is tested on an orbital rendezvous scenario and achieves high detection accuracy.

The paper presents a novel approach for adversarial attack detection in deep neural network-based relative pose estimation schemes for autonomous spacecraft navigation. The proposed approach uses the explainability concept and an LSTM network to achieve high detection accuracy. The approach is tested on both simulated and real data and achieves high detection accuracy in both cases. The paper also provides a detailed description of the experimental setup and the datasets used. The proposed approach has the potential to enhance the safety and security of autonomous spacecraft navigation.

Related work in the field of adversarial attacks on deep learning in autonomous spacecraft navigation includes 'Adversarial Attacks on Autonomous Spacecraft Navigation' by M. Alzahrani et al. and 'Adversarial Machine Learning in Space: State-of-the-Art and Challenges Ahead' by M. Alzahrani et al.