ARCHIVES

Adversarial perturbations, including synonym substitutions, character-level edits, and token insertions, can significantly degrade the performance of modern sentiment-analysis models. Although adversarial training is widely regarded as one of the most effective defences against such attacks, it typically incurs sub- stantial computational cost and training complexity, which can limit its practicality in resource-constrained deployment settings. This work investigates whether meaningful robustness improvements can be achieved through a lightweight hybrid defence that balances robustness, accuracy, and efficiency. We propose HAT-D, a lightweight hybrid robustness framework that combines TRADES-style embedding-space adversarial training with a compact denoising module (and optional smoothing), and evaluate it under synonym, character, insertion, and mixed attack settings. The design aims to improve robustness against discrete text perturbations without the full computational cost of end-to-end adversarial fine-tuning. The framework is evaluated on the SST-2 sentiment classification benchmark under multiple adversarial attack families, including synonym substitution, character-level perturbations, token insertions, and mixed strategies. Experimental results show that full adversarial training achieves the highest overall robustness under our multi-attack evaluation, while ensemble-based defences remain competitive but incur substantial deployment overhead. HAT-D provides a single-model hybrid alternative that preserves essentially unchanged clean inference latency (at the reported precision) and avoids the multiplicative adaptation cost and footprint of ensembles. Overall, the results emphasize that robustness is attack-family dependent, motivating joint reporting of robustness, clean performance, and efficiency when selecting practical defences. These results highlight the importance of evaluating robustness methods not only by adversarial accuracy but also by their computational efficiency. By explicitly characterizing the robustness–accuracy–efficiency trade-off, this work demonstrates that lightweight hybrid defences can provide a practical operating point for robust NLP deployment in resource-constrained environments.