The visualization of normalized layer-wise weights for Small SSL Models specifically Wav2Vec2-Small, Hubert-Small, and WavLM-Small reveals a highly concentrated reliance on the network's lower layers. Heatmap analyses show that the highest contributions consistently originate from layers 1 through 6 across all evaluated datasets and deepfake scenarios, including full speech, and song setups.

The visualization of Large SSL Models (Wav2Vec2-Large, Hubert-Large, and WavLM-Large) displays a much broader and less concentrated pattern of layer importance. The lower half of the network (specifically layers 2 through 13) still provides significant contributions to deepfake detection across most datasets. When looking at the overall averages across all setups, the highest contributions still emerge from the lower layers (1-12), with a distinct peak in importance clustered around the 4th to 7th layers.