Urgent.News

What's breaking now, across thousands of outlets.

Tech

Fine-tuning a certified network without losing its robustness guarantee

A certified classifier comes with a proof for each input: inside a radius r around the image, no perturbation can change the predicted class. In medical imaging that radius has to cover what happens in a clinic every day, such as a second photo of the same patch of skin taken at a slightly different angle. The proof rests on the Lipschitz constant L of the network. The certificate divides the…

Certified classifiers provide a guarantee that no perturbation within a certain radius around an image will alter the predicted class. This radius depends on the Lipschitz constant L of the network. When fine-tuning, L increases, causing the radius to shrink proportionally. Fine-tuning a certified network requires adjusting L based on the amount of data available - LoRA raises L by 1.9 to 2.5 times, while full fine-tuning raises it by 5 to 8.5 times.

Butterfly adapters maintain the singular values of frozen weights during training. They are created by multiplying each frozen weight matrix by an orthogonal matrix R(θ), ensuring the spectral norm remains unchanged. Givens rotations, which are orthogonal for any angle, are used to construct the adapters, allowing Adam to train the angles directly without projection or penalty terms. Each adapter can be represented as a quantum circuit of depth log₂ n.

In a study involving trichoscopic scalp images from a clinic, certified BRONet was adapted using butterfly adapters and LoRA. The butterfly adapter achieved 63.8% accuracy at ε = 36/255, compared to 31.6% for LoRA and 20.4% for full fine-tuning. Full fine-tuning reached 35.6% accuracy at ε = 108/255. Trained LoRA adapters achieved 2.764 L/L₀, while the butterfly adapter achieved 1.000 L/L₀. The butterfly adapter improved certified accuracy by 8.6 points and demonstrated robustness against adversarial attacks.

Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at dev.to →

More in Tech

More from Tuesday 17 November →