Why deepfake detection is really a data problem

Filed under: lab notes · 2026-07-18

For the past few years I have spent most of my time on speech deepfake detection. The models keep getting better at fooling us — and the detectors keep getting better at catching them. But there is a pattern I keep running into, across every method I have worked on: the bottleneck is almost never the model architecture. It is the data.

The distribution gap

A detector trained on a clean, studio-recorded spoofing dataset can hit near-perfect accuracy on its own test split. Put it on a real phone call — compression, background noise, a different microphone — and performance collapses. This is the gap we studied in When AVSR Meets Video Conferencing: real-world channels introduce degradation patterns that simply do not exist in the lab.

What this means for how we work

1. Benchmarks matter. A paper’s number is only meaningful relative to the test set it is measured on.
2. Datasets are contributions. Building EchoFake taught me that a well-constructed dataset can be as valuable as a clever loss function.
3. Skepticism is a research skill. If your detector only works in one setting, you have found a feature of the test set, not a property of the world.

This is part of why the research lines on this site are organized the way they are — each direction is really asking the same question from a different angle: what can we trust, and under what conditions?

More notes to come as I keep working through it.