Dipartimento di Ingegneria informatica, automatica e gestionale

Microphone identification based on the intrinsic physical features has received significant attention by the research community in recent years. Such properties can be exploited in security and forensics applications in order to assess the authenticity of a certain audio track or for audio attribution. The detection is possible since the specific characteristics of the microphone components slightly change from one microphone to another due to the manufacturing process. Various techniques have been proposed to implement physical microphone identification from the use of hand-tailored features (e.g., entropy measure) to spectral representation (e.g., cepstral coefficients) in combination with machine learning algorithms. In recent times, the application of deep learning to microphone identification was successfully demonstrated especially in comparison to shallow machine learning algorithms. On the other hand, deep learning requires significant computing resources especially with large data sets, as in the case of audio recordings for microphone identification. Then, dimensionality reduction could benefit the computing time efficiency for this task. The proposed study envisaged the combined use of Convolutional Neural Networks with spectral entropy features extraction to improve time efficiency while preserving a high identification accuracy. Spectral features, based on Shannon entropy and Renyi entropy, are proposed in combination with the ReliefF algorithm to implement a dimensionality reduction of the spectral representation of the audio signals recorded from 34 different microphones. Then, the reduced spectral representation is fed to a custom Convolutional Neural Network to perform the classification. The results show that this approach is able to reduce significantly the processing time in comparison with the state of the art while preserving a comparable identification accuracy and an increased robustness to the presence of noise.