Mitigating Filter-Based Adversarial Attacks in BCIs through Model Compression

Brain-Computer Interface systems often employ machine learning classifiers to interpret EEG signals for real-time user interaction. However, these models can be compromised by adversarial attacks such as evasion attacks that subtly distort EEG inputs to induce misclassification. In this study, we investigate using model compression techniques to mitigate filter-based adversarial attacks. Our proposed technique utilizes structured weight pruning and quantization-aware training to systematically reduces the network's representational capacity. Our results indicate strong attack resilience, achieving a balanced classification accuracy of 72.97 percent for 8 subjects, thereby mitigating the effects of adversarial distortions while preserving baseline classification accuracy.