TY - GEN
T1 - Robust Satellite Image Classification with Bayesian Deep Learning
AU - Pang, Yutian
AU - Cheng, Sheng
AU - Hu, Jueming
AU - Liu, Yongming
N1 - Funding Information:
The research reported in this paper was supported by funds from NASA University Leadership Initiative program (Contract No. NNX17AJ86A, PI: Yongming Liu, Technical Officer: Anupa Bajwa). The support is gratefully acknowledged.
Publisher Copyright:
© 2022 IEEE.
PY - 2022
Y1 - 2022
N2 - Image-based object detection and classification are essential for satellite-based monitoring, which spans multiple safety-critical engineering applications. Meanwhile, state-of-the-art deep learning significant improves the accuracy for image classification tasks thus has been deployed in various scenarios. However, it's well-known deep learning-based image classifiers are vulnerable to small perturbations along specific directions, known as adversarial attacks. These attacks are exceptionally effective to fool image classifiers. In extreme cases, merely one pixel's change can lead to a attacker-desired wrong prediction label. In this work, we show that deep learning with Bayesian formulation can extend the deep learning adversarial robustness by a large margin, without the need of adversarial training. Moreover, we show that the stochastic classifier after the deterministic CNN extractor has sufficient robustness enhancement rather than a stochastic feature extractor before the stochastic classifier. This advises on utilizing stochastic layers in building decision-making pipelines within a safety-critical domain. Additionally, we show that the Bayesian posterior can act as the safety precursor of ongoing malicious activities towards a deployed image classification system. This leads to the detection of adversarial samples in cybersecurity. With lots of potentials, we leave them as the future studies.
AB - Image-based object detection and classification are essential for satellite-based monitoring, which spans multiple safety-critical engineering applications. Meanwhile, state-of-the-art deep learning significant improves the accuracy for image classification tasks thus has been deployed in various scenarios. However, it's well-known deep learning-based image classifiers are vulnerable to small perturbations along specific directions, known as adversarial attacks. These attacks are exceptionally effective to fool image classifiers. In extreme cases, merely one pixel's change can lead to a attacker-desired wrong prediction label. In this work, we show that deep learning with Bayesian formulation can extend the deep learning adversarial robustness by a large margin, without the need of adversarial training. Moreover, we show that the stochastic classifier after the deterministic CNN extractor has sufficient robustness enhancement rather than a stochastic feature extractor before the stochastic classifier. This advises on utilizing stochastic layers in building decision-making pipelines within a safety-critical domain. Additionally, we show that the Bayesian posterior can act as the safety precursor of ongoing malicious activities towards a deployed image classification system. This leads to the detection of adversarial samples in cybersecurity. With lots of potentials, we leave them as the future studies.
UR - http://www.scopus.com/inward/record.url?scp=85130713371&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85130713371&partnerID=8YFLogxK
U2 - 10.1109/ICNS54818.2022.9771496
DO - 10.1109/ICNS54818.2022.9771496
M3 - Conference contribution
AN - SCOPUS:85130713371
T3 - Integrated Communications, Navigation and Surveillance Conference, ICNS
BT - 2022 Integrated Communication, Navigation and Surveillance Conference, ICNS 2022
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 2022 Integrated Communication, Navigation and Surveillance Conference, ICNS 2022
Y2 - 5 April 2022 through 7 April 2022
ER -