The following paper was presented this year at ICASSP 2026 in Barcelona:
Justin Dettmer, Annelot Bosman, Igor Vatolkin, Holger Hoos: Audio Classification Models are Vulnerable to Filter Perturbations
Abstract: Deep learning models are known to suffer from a lack of robustness to adversarial attacks. While the majority of research on the robustness of neural networks stems from the image classification domain, audio classifiers have also come under scrutiny, as their use in tasks such as speech processing is increasing. We introduce a novel adversarial attack method that creates band-pass filters as perturbations for mel spectrograms. These perturbations are closer to real-life conditions than conventionally created adversarial examples, as they can simulate the use of different recording devices or other acoustic properties. We demonstrate that state-of-the-art audio classifiers lack robustness against these adversarially-created filters, and we show that adversarial training with our attack method can help mitigate this effect on the widely studied NSynth, ESC-50 and SpeechCommands datasets.