Metaverse Target-Tracking Security: A Study on Adaptive Metalearning Adversarial Attack Methods

The rapid development of augmented reality, mixed reality, extended reality, and virtual reality technologies has provided users with unprecedented immersive, interactive experiences while also posing new security challenges. Specifically, target-tracking systems are highly susceptible to adversarial attacks. Moreover, generative model-based attacks have gained widespread attention owing to their efficiency but are often influenced by trained trackers and a lack of generality. Inspired by the concept of metalearning, this study proposes a new attack method: adaptive metalearning adversarial attack (AMLAA). AMLAA aims to cultivate the cross-model migration capability of perturbed generators to achieve an effective attack on multiple target-tracking algorithms and to generate more migratory adversarial samples while enhancing the stealthiness and practicality of the attack. We explore the target-tracking techniques based on Siamese and transformer network architectures in depth, using a metalearning framework to fuse the key features of the two network architectures, constructing two novel loss functions, and exploring the synergistic optimization and enhancement of the attack effect between different network architectures. Through experimental validation on four target-tracking datasets (LasoT, OTB100, GOT-10k, and VOT2018), we find that AMLAA demonstrates powerful attack capabilities and excellent cross-model attack capabilities.