Quang-Tien Tran

Machine learning (ML) has become the core of many applications thanks to the continuous advancements in the field. Various tasks in Remote Sensing (RS) are being completed with the help of ML models, including crop monitoring, flood prediction, etc.  However, deploying a ML-based model can be extremely challenging because of concept drift, which can degrade the model’s accuracy over time. One of the most popular types of concept drift is data drift [1] (aka distribution shift). This drift frequently happens when the data collected for model training is over or under-represented (biased data collection), or when the testing environment is different from the training environment (non-stationary environment) that frequently happens with continual learning ML systems integrated in RS and Earth Observation applications. Recent efforts to deal with the problem use external sources of knowledge such as domain or environment proxy data, protected attributes, or sub population groups. However, these approaches do not solve the problem completely and the models remain
sensitive to data drifts. The major issue behind this observed behaviour is the lack of clarity and applicability for the underlying assumptions regarding the problem statement
and the data-generating process [2]. This becomes more severe when ML models are vulnerable to adversarial attacks such as data poisoning attacks [3][4], which are increasingly frequent with RS data [5]. Adversaries can deliberately alter the decision boundary of the model, leading to "adversarial drift". Therefore, a thorough understanding of the data and relevant threats of drifting data are necessary to keep the model robust and maintain good accuracy. Casual inference is a powerful tool to model the data generation process and comprehend the causal relations in the training data. Causal models are expected to generalize under certain distribution shifts since they explicitly model interventions [6]. Recently, neural networks have shown the ability to expand causal inference to new settings where confounding variables/data is high-dimensional, non-linear, time-varying, or even embedded in texts and images. This project will research and develop new approaches for inspecting data drift in RS. It will perform a thorough background survey about the RS data, threat models of adversarial attacks on this data, together with mitigation/detection action plans. This project aims to propose a new deep learning model, using casual inference for the adversarial drift detection problem in RS.