Interpretable Temporal Class Activation Representation for Audio Spoofing Detection

Built-in transparency beats add-on explanations: Rather than applying explanation tools to already-trained models, this research embeds interpretability directly into the model's architecture from the start. This approach provides more reliable insights into why the system makes decisions about whether audio is genuine or fake.

Identifying which frames matter for detection: The model uses "class activation representation" to pinpoint the specific moments in an audio clip that trigger a spoofing detection. This localization capability helps auditors and users understand what acoustic features the system is actually responding to, rather than treating the model as a "black box."

Learning attack-specific patterns improves accuracy: Training the model to distinguish between different types of spoofing attacks (rather than just "real vs. fake") helps it learn the unique characteristics of each attack method. This multi-label approach achieved state-of-the-art performance with an error rate of only 0.51%, suggesting that more granular classification leads to more robust detection.

Interpretability supports trust in high-stakes applications: For audio authentication systems used in security contexts (like voice biometrics), being able to explain detection decisions is essential for user trust and regulatory compliance. This work demonstrates that transparency and performance don't have to be trade-offs.