Abstract:
With the increasing capabilities of artificial intelligence, the risk of impersonation and deepfakes has grown, particularly in audio, where ”spoofing attacks” involve altering or faking voices to deceive systems or individuals. This study introduces a novel deep learning architecture designed to detect such spoofing in audio. Focusing on Logical Access (LA) attacks—where digital spoof ing is directly introduced into audio without tampering with physical systems—our approach addresses challenges like voice recordings, machine-generated voices, and voice conversion. Building on traditional feature extraction and advanced deep neural networks such as Con volutional Neural Networks (CNNs) and attention-based architectures, our model leverages Large Language Models (LLMs) to enhance detection. The integration of LLMs introduces context-aware predictions, adding robustness to spoof detection. During training, a label-based ”oracle” mechanism is used, injecting controlled impurity to prevent overreliance. Inference involves prompt-tuned LLMs that classify samples as bonafide, spoofed, or ”unknown,” with post-inference adjustments amplifying score distinctions. Evaluated on the ASVSpoof dataset, our enhanced architecture demonstrates significant improvements in performance, surpassing state-of-the-art benchmarks in detection accuracy and reliability. This advancement strengthens voice authentication systems, ensuring enhanced protection against sophisticated spoofing attacks.
