In cognitive radio networks (CRNs), secondary users must be able to accurately and reliably track the location of small-scale mobile primary users/devices (e.g., wireless microphones) in order to efficiently utilize spatial spectrum opportunities, while protecting primary communications. However, accurate tracking of the location of mobile primary users is difficult due mainly to the CR-unique constraint, i.e., localization must rely solely on reported sensing results (i.e., measured primary signal strengths), which can easily be compromised by malicious sensors (or attackers). To cope with this challenge, we propose a new framework, called Sequential mOnte carLo combIned with shadow-faDing estimation (SOLID), for accurate, attack/fault-tolerant tracking of small-scale mobile primary users. The key idea underlying SOLID is to exploit the temporal shadow fading correlation in sensing results induced by the primary user's mobility. Specifically, SOLID augments conventional Sequential Monte Carlo (SMC)-based target tracking with shadow-fading estimation. By examining the shadow-fading gain between the primary transmitter and CRs/sensors, SOLID 1) significantly improves the accuracy of primary tracking regardless of the presence/absence of attack, and 2) successfully masks the abnormal sensing reports due to sensor faults or attacks, preserving localization accuracy and improving spatial spectrum efficiency. Our extensive evaluation in realistic wireless fading environments shows that SOLID lowers localization error by up to 88 percent in the absence of attacks, and 89 percent in the presence of the challenging slow-poisoning” attack, compared to the conventional SMC-based tracking.