In this work, we consider secure communications in wireless multi-user (MU) multiple-input single-output (MISO) systems with channel coding in the presence of a multi-antenna eavesdropper (Eve), who is a legit user trying to eavesdrop other users. In this setting, we exploit machine learning (ML) tools to design soft and hard decoding schemes by using precoded pilot symbols as training data. The proposed ML frameworks allow an Eve to determine the transmitted message with high accuracy. We thereby show that MU-MISO systems are vulnerable to such eavesdropping attacks even when relatively secure transmission techniques are employed, such as symbol-level precoding (SLP). To counteract this attack, we propose two novel SLP-based schemes that increase the bit-error rate at Eve by impeding the learning process. We design these two security-enhanced schemes to meet different requirements regarding runtime, security, and power consumption. Simulation results validate both the ML-based eavesdropping attacks as well as the countermeasures, and show that the gain in security is achieved without affecting the decoding performance at the intended users.