Abstract:

Existing malicious SMS detection methods often focus on improving the detection accuracy or speed, ignoring the security problems of the model itself, thus likely to suffer from adversarial examples attack in real-world scenarios. To alleviate this pain point, this paper proposed a malicious SMS detection model that blended adversarial enhancement and multi-task optimization. During the input stage, a random matching pool was used to generate “original text-adversarial example” pairs as input, and the semantic type encoding technique was adopted to help the model distinguish the data boundaries. Then, a single-tower neural network based on ChineseBERT was used as the backbone model to excavate the semantic, pinyin, and glyph features of the SMS. In the output stage, the supervised classification cross-entropy loss and the unsupervised input consistency loss were used as multi-task optimization objectives to help the model learn the correlated features of text pairs and complete the classification. Experimental results based on the public datasets show that the proposed method outperforms a variety of machine learning and deep learning detection methods in terms of accuracy and robustness.

Key words: malicious SMS, robustness, adversarial examples, multi-task learning