Quantum cloud platforms face the challenge of quantum circuit Trojan attack. Trojan attack in quantum circuits originates from dual-mode attack involving parameter implantation and structure implantation, threatening the secure execution of quantum programs in cloud environments. To address the difficulty of detecting dual-mode quantum circuit Trojan attack in quantum cloud platforms, we propose a dual-mode quantum circuit Trojan attack detection scheme based on unitary matrix features. By constructing dual-mode quantum Trojan attack and evaluating it with quantitative metrics (accuracy drop, JS divergence, etc.), a quantum circuit is represented by a unitary matrix, from which real and imaginary component features are extracted. A convolutional neural network model is deployed for detection on 1550 benchmark circuits. Experimental results indicate that our scheme achieves an accuracy of 91.6%, a balanced accuracy of 93.8%, and a F1-score of 0.909, exceeding the performance of traditional machine learning models. These results verify the effectiveness of deep learning for quantum security detection, establishing a defense mechanism that integrates unitary matrix features learning with four-dimensional quantitative evaluation and providing a verifiable solution for Trojan protection in quantum cloud platforms.