The openness of PS systems renders them vulnerable to malicious users that can pollute the measurement collection process, in an attempt to degrade the PS system data and, overall, its usefulness. Mitigating such adversarial behavior is hard. Cryptographic protection, authentication, authorization, and access control can help but they do not fully address the problem. Reports from faulty insiders (participants with credentials) can target the process intelligently, forcing the PS system to deviate from the actual sensed phenomenon. Filtering out those faulty reports is challenging, with practically no prior knowledge on the participants' trustworthiness, dynamically changing phenomena, and possibly large numbers of compromised devices. This paper proposes SHIELD, a novel data verification framework for PS systems that can complement any security architecture. SHIELD handles available, contradicting evidence, classifies efficiently incoming reports, and effectively separates and rejects those that are faulty. As a result, the deemed correct data can accurately represent the sensed phenomena, even when 45% of the reports are faulty, intelligently selected by coordinated adversaries and targeted optimally across the system's coverage area.