Intelligent Multi-Modal Spam Detection System Using Machine Learning Techniques for AI-Generated Content

We attempted to implement a spam classifier of good quality not only for classic spam but also for new generation of spam produced using AI technologies. Our classifier confirmed some of our hypotheses - SVM cannot be surpassed in terms of TF-IDF representation, preprocessing is crucial, precision-recall balance is highly dependent on the selected algorithm but also uncovered some unexpected results. What is most fascinating is the fact that classical stages of NLP pipeline, invented many years ago, before anyone even considered ChatGPT, provide an accidental yet effective protection against spam produced using AI. Stemming lowers the diversity of paraphrases. Weighting by IDF helps to overcome preference of the language model for common tokens. The margin of SVM serves as a buffer in the feature space in relation to paraphrasing perturbations. Nevertheless, these techniques have some drawbacks. They can be bypassed by an intelligent enough adversary attempting to move beyond the training vocabulary. In the long term, it will be necessary to apply not one technique but a whole strategy of defense built around a classical pipeline, including enrichment of the corpus by AI-produced spam, characteristics of AI-produced text detection by means of perplexity, and adversarial training. For future work, we plan to: 1) Train the model on diverse LLM-generated spam, enabling classifiers to witness themselves the statistical patterns created by machines. 2) Adopt the 2-step system, where the TF-IDF classifier is employed in the first step, and the BERT-based classifier [19] is applied to emails that lie on the borderline. 3) Test the system through a well-defined process based on adversarial examples created through iterative queries on the classification model. 4) Add multilingual preprocessing in the form of language-agnostic tokenizers and multilingual embeddings. 5) Add Federated Learning, ensuring that the model learns during its application to many devices while none of their communications leave the device. The code, trained model, and web app are provided for replication and further exploration.  

[1] GSMA Intelligence, "The mobile economy 2023," GSMA, London, U.K., Tech. Rep., 2023. [2] M. Delany, "Domain-based message authentication, reporting, and conformance (DMARC)," Internet Eng. Task Force, RFC 7489, Mar. 2015. [3] Statista Research Dept., "Global spam volume as percentage of total e-mail traffic from 2007 to 2023," Statista, 2023. [4] T. Brown et al., "Language models are few-shot learners," in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 33, 2020, pp. 1877–1901. [5] Z. Waseem and D. Hovy, "Hateful symbols or hateful people? Predictive features for hate speech detection on Twitter," in Proc. NAACL Student Res. Workshop, San Diego, CA, USA, 2016, pp. 88–93. [6] Apache Software Foundation, "Apache SpamAssassin," 2023. [Online]. Available: https://spamassassin.apache.org/ [7] Spamhaus Technology Ltd., "The Spamhaus project," 2023. [Online]. Available: https://www.spamhaus.org/ [8] P. O. Boykin and V. P. Roychowdhury, "Leveraging social networks to fight spam," IEEE Comput., vol. 38, no. 4, pp. 61–68, Apr. 2005. [9] F. Sebastiani, "Machine learning in automated text categorization," ACM Comput. Surv., vol. 34, no. 1, pp. 1–47, Mar. 2002. [10] M. Sahami, S. Dumais, D. Heckerman, and E. Horvitz, "A Bayesian approach to filtering junk e-mail," in Learning for Text Categorization: Papers from the 1998 Workshop, AAAI Tech. Rep. WS-98-05, Madison, WI, USA, 1998, pp. 55–62.