Security and reliability of network protocol implementations are essential for communication services. Most of the approaches for verifying security and reliability, such as formal validation and black-box testing, are limited to checking the specification or conformance of implementation. However, in practice, a protocol implementation may contain engineering details, which are not included in the system specification but may result in security flaws. We propose a new learning-based approach to systematically and automatically test protocol implementation security properties. Protocols are specified using Symbolic Parameterized Extended Finite State Machine (SP-EFSM) model, and an important security property - message confidentiality under the general Dolev-Yao attacker model - is investigated. The new testing approach applies black-box checking theory and a supervised learning algorithm to explore the structure of an implementation under test while simulating the teacher with a conformance test generation scheme. We present the testing procedure, analyze its complexity, and report experimental results.