Select input type
Choose an audio source
Waiting for audio.
Run the model locally in your browser, then listen to what changes when the input or the training data is attacked.
Inference
Audio is decoded and classified on this device. Nothing is uploaded. The model expects one second of 16 kHz audio and recognizes ten commands plus background and unknown.
Select input type
Waiting for audio.
Model output
Fast Gradient Sign Method Attack
Fast Gradient Sign Method changes every audio sample in the direction that increases the model's loss. These examples were generated offline with the differentiable PyTorch model, then checked here with the same weights in ONNX. The slider continues into deliberately extreme, audible settings so the perturbation is easier to explore.
Training-time attack
The poisoned model saw background clips with a low-frequency sine trigger relabeled as “yes.” Add that trigger below and compare its behavior with the clean model. At 5 Hz, the tone sits below the usual range of human hearing.
Model anatomy
Res15 learns directly from one-second waveforms. Its fixed frontend turns audio into a mel spectrogram, and a stack of dilated residual convolutions learns the patterns that separate the ten selected commands from background and unknown speech.
Unchanged Res15 architecture
Input
One mono second at 16 kHz1 × 1 × 16,000
Frontend
40 mel bins × 101 frames
log transform + normalization
Feature extractor
One input convolution + 13 residual convolutions
65 feature maps with skip
connections
Classifier
Global average pooling
65 → 12 linear layer + softmax
Experiment tracking
The selected clean and poisoned checkpoints are shown with their original metrics and confusion matrices.
Loading the run summaries…