• Rumaisa Syed
• Adithi S Bharadwaj
• Nayana V M
• Anushree K N
• Yuga S Gowda

Abstract : Environmental Sound Classification (ESC) is a key component in intelligent systems such as smart surveillance, public safety, and autonomous vehicles. Traditional machine learning approaches relying on handcrafted features like MFCCs and spectral centroids often underperform in noisy environments. This work employs a Convolutional Neural Network (CNN) trained on Mel-spectrograms from the UrbanSound8K dataset to enable robust, automatic feature extraction. Beyond classification, the study evaluates the model’s vulnerability to adversarial audio perturbations generated through the Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) attacks. Experimental results reveal that even imperceptible perturbations can significantly degrade model performance, exposing critical weaknesses in deep audio systems. These findings highlight both the potential of CNNs for accurate ESC and the necessity of enhancing adversarial robustness to ensure safer and more reliable audio-based AI applications.

Keywords: Adversarial Attacks, Audio Classification, White-Box Model, Adversarial Robustness, Inaudible Perturbations

"INAUDIBLE EVASION: A WHITE-BOX ADVERSARIAL SIMULATOR FOR AUDIO CLASSIFICATION", International Journal of Emerging Technologies and Innovative Research (www.jetir.org), ISSN:2349-5162, Vol.12, Issue 11, page no.f118-f125, November-2025, Available :http://www.jetir.org/papers/JETIR2511538.pdf

"INAUDIBLE EVASION: A WHITE-BOX ADVERSARIAL SIMULATOR FOR AUDIO CLASSIFICATION", International Journal of Emerging Technologies and Innovative Research (www.jetir.org | UGC and issn Approved), ISSN:2349-5162, Vol.12, Issue 11, page no. ppf118-f125, November-2025, Available at : http://www.jetir.org/papers/JETIR2511538.pdf