An empirical study of how adapting pre-trained language models to new tasks through transfer learning affects their vulnerability to adversarial attacks. Across several architectures and bias-detection datasets, the work finds that transfer learning can raise standard accuracy while simultaneously making models easier to fool with adversarial text, and that larger models tend to resist this degradation. Students learn why gains in task performance do not guarantee security, and how model size and adaptation method interact with robustness.
We investigate the adversarial robustness of LLMs in transfer learning scenarios. Through comprehensive experiments on multiple datasets (MBIB Hate Speech, MBIB Political Bias, MBIB Gender Bias) and …