Temporally-Transferable Perturbations: Efficient, One-Shot Adversarial Attacks for Online Visual Object Trackers

In recent years, the trackers based on Siamese networks have emerged as\nhighly effective and efficient for visual object tracking (VOT). While these\nmethods were shown to be vulnerable to adversarial attacks, as most deep\nnetworks for visual recognition tasks, the existing attacks for VOT trackers\nall require perturbing the search region of every input frame to be effective,\nwhich comes at a non-negligible cost, considering that VOT is a real-time task.\nIn this paper, we propose a framework to generate a single temporally\ntransferable adversarial perturbation from the object template image only. This\nperturbation can then be added to every search image, which comes at virtually\nno cost, and still, successfully fool the tracker. Our experiments evidence\nthat our approach outperforms the state-of-the-art attacks on the standard VOT\nbenchmarks in the untargeted scenario. Furthermore, we show that our formalism\nnaturally extends to targeted attacks that force the tracker to follow any\ngiven trajectory by precomputing diverse directional perturbations.\n