DEEP LEARNING FOR SAR DATA IN PRESENCE OF ADVERSARIAL SAMPLES
The closing date for this job submission has passed.
Classification of SAR image data continues to be a big challenge. Major difficulties include the scarcity of available data, and the difficulty of semantically interpreting the SAR backscattered signal. There are no large-scale, SAR-derived image databases for Remote Sensing image analysis and knowledge discovery. Furthermore, while optical image classification has seen a breakthrough with the advent of Deep Learning methods that require Big Data, SAR-based systems have so far not experienced the same progress, likely because not enough data associated training labels is available.
The nature of the adversarial samples occurring spontaneously depends on the sensor type. In SAR, the effect of strong scattering or the model of image formation and the physical processes behind need very specific methods for dealing with adversarial samples. Given the particular nature of these samples, the solutions to avoid their insertion in the training sets or to alleviate their effects must be tailored accordingly.
The main objective of this project is to give solutions for deep learning with spontaneous adversarial samples in the case of SAR data.
The successful candidate will be employed for 3 years by University Politehnica of Bucharest and will be part of CEOSpaceTech, a very dynamic research laboratory oriented towards space applications. She or he receives a financial package, which is twice the average salary in the country and additional mobility and family allowance, granted according to the rules for Early Stage Researchers (ESRs) in an EU Marie Sklodowska-Curie Actions Innovative Training Networks (ITN). A career development plan will be prepared for her/him in accordance with the supervisor. The plan will include a choice of more than 20 streamed or registered courses, one stage in Germany, and various outreach activities. For more information please visit the Marie Sklodowska-Curie Actions Innovative Training Networks website.
• Use transformation methods for a relevant SAR data representation, in order to avoid insertion of adversarial samples.
• Design of DNNs for SAR data classification in order to achieve a given invariance to spontaneous adversarial samples.
• Define projections of features when learning the semantic axes for 3D visualization such to contextually disambiguate the meaning and to ensure a consistent training.
• A Master of Science in Computer Science is required. It could comprise the full range of mathematical, physical, engineering and technology disciplines related to sensor data acquisition and programming.
• Proficient English level.
• Image analysis, neural networks, programming languages, basic knowledge on optics could be an advantage.
• DLR, Munich, Germany, Prof. M. Datcu,10 months, theoretical aspects of DNN for SAR images and related topics on the impact of adversarial samples.
• Marmanis, Dimitrios, et al. “Artificial generation of big data for improving image classification: A generative adversarial network approach on SAR data.” arXiv preprint arXiv:1711.02010 (2017).
• Zhao, Juanping, et al. “Contrastive-Regulated CNN in the Complex Domain: A Method to Learn Physical Scattering Signatures From Flexible PolSAR Images.” IEEE Transactions on Geoscience and Remote Sensing (2019).
• Goodwin, Justin A., et al. “Learning Robust Representations for Automatic Target Recognition.” arXiv preprint arXiv:1811.10714 (2018).