Reconstruction Error Fingerprints for Few-Shot Neural Audio Codec Attribution

Modern music generation systems frequently rely on neural audio codecs or reconstruction backends to transform compact latent codes into waveforms. Previous work has shown that these reconstruction stages are highly informative for binary AI-generated music detection. In this paper, we study a complementary attribution question: given a reconstructed music excerpt, which codec family or codec configuration produced it? This problem is important because source attribution can support provenance analysis, auditing of generative pipelines, and diagnosis of reconstruction backends, while detection alone only answers whether an excerpt appears synthetic. To the best of our knowledge, existing work on AI-generated music has primarily addressed binary detection or text-to-music generator attribution, whereas neural audio codec attribution has not been systematically studied. Our approach builds on the frozen-decoder/source-specific encoder principle introduced for few-shot generative model attribution. A base autoencoder is pretrained on real music spectrogram patches, its decoder is frozen, and a compact encoder is trained for each candidate codec source. At inference time, selected patches are reconstructed by all source-specific encoder-decoder pairs, producing a vector of mean reconstruction errors that is classified by a linear support vector machine. We evaluate family-level attribution among Encodec, GriffinMel, and the LAC/DAC family, as well as model-level attribution among eight codec configurations, on a Free Music Archive (FMA)-based reconstruction benchmark. We further compare the autoencoder fingerprint with a full-spectrogram ResNet18 baseline motivated by machine-generated music detection. The results show a clear data-regime trade-off: autoencoder fingerprints are more data-efficient in the lowest-shot settings, while ResNet18 becomes stronger when hundreds of labelled examples per source are available.