A deepfake detector with averaged weights reaches 97.04% accuracy

Researchers built PairCLIP-SWA, a detector for deepfake images that uses one OpenCLIP model. It averages the model weights during training. It does not keep the one checkpoint with the best score on check data. A checkpoint is a saved copy of the model. On the IEEE SP Cup 2025 benchmark it reached 97.04% accuracy. The best published team of three models reached 94.46%.

How it works

The authors train the model on fake images and on images that are not fake, in equal numbers. They then average the weights from the last epochs, which are cycles through the training data.

The numbers

The best single checkpoint changed by 5.34 points of accuracy in training. On a different test of 10,905 images, the checkpoint that validation chose had 88.85% and rejected 21.1% of photos that are not fake. The average had 94.76%.

What is not known

Secondary face cropping cost about 34 points. Frequency inputs transferred poorly. The authors tested set-ups on one fixed set of data. They only recommend selection rules for each class.

Insight: The best score on check data can pick the worst detector

In October 2026, a test found that three of six deepfake detectors flag a robot face as fake. A different study found that continual learning frequently increases unfairness. A detector is chosen on one number that can be incorrect, and the buyer who tests on new data holds the proof. Do not trust one validation score.