With growing role of social platforms in shaping communicative behavior, they bring new challenges in terms of recognizing and combating harmful posts. The automatic identification of such content is challenging, especially in Dravidian language posts, where the code-switching use of languages, alternate spelling variants and different conventions employed on different platforms can occur. This work provides a framework for identifying linguistic patterns that are consistent among posts originating from different social media platforms and are written in a variety of languages. Our model uses XLM-RoBERTa contextual embeddings and is trained to do distinction of codeswitched posts with high accuracy and to identify the source platform of text. We train our model using DravidianLangTech 2025 dataset and evaluate cross-dataset generalization on HASOC FIRE 2022 and DravidianCodeMix 2020 benchmark datasets. To evaluate the versatility and robustness of our approach, we also evaluate on Telugu Code-Mixed Dataset under controlled perturbations such as alternate spellings, additional diacritics, error in punctuation, and emoji substitutions. On various platforms, our approach demonstrates better performance and resistance to different obfuscation techniques implemented in text.

Abusive Language Detection, Dravidian CodeMixed Languages, Cross-Platform Learning, Platform-Invariant Learning, Multi-Task Learning, XLM-RoBERTa, Robustness Evaluation, Social Media Text Classification.