Anomaly Detection in Unsupervised Surveillance Setting Using Ensemble of Multimodal Data with Adversarial Defense

Autonomous aerial surveillance using drone feed is an interesting and\nchallenging research domain. To ensure safety from intruders and potential\nobjects posing threats to the zone being protected, it is crucial to be able to\ndistinguish between normal and abnormal states in real-time. Additionally, we\nalso need to consider any device malfunction. However, the inherent uncertainty\nembedded within the type and level of abnormality makes supervised techniques\nless suitable since the adversary may present a unique anomaly for intrusion.\nAs a result, an unsupervised method for anomaly detection is preferable taking\nthe unpredictable nature of attacks into account. Again in our case, the\nautonomous drone provides heterogeneous data streams consisting of images and\nother analog or digital sensor data, all of which can play a role in anomaly\ndetection if they are ensembled synergistically. To that end, an ensemble\ndetection mechanism is proposed here which estimates the degree of abnormality\nof analyzing the real-time image and IMU (Inertial Measurement Unit) sensor\ndata in an unsupervised manner. First, we have implemented a Convolutional\nNeural Network (CNN) regression block, named AngleNet to estimate the angle\nbetween a reference image and current test image, which provides us with a\nmeasure of the anomaly of the device. Moreover, the IMU data are used in\nautoencoders to predict abnormality. Finally, the results from these two\npipelines are ensembled to estimate the final degree of abnormality.\nFurthermore, we have applied adversarial attack to test the robustness and\nsecurity of the proposed approach and integrated defense mechanism. The\nproposed method performs satisfactorily on the IEEE SP Cup-2020 dataset with an\naccuracy of 97.8%. Additionally, we have also tested this approach on an\nin-house dataset to validate its robustness.\n