Skip to content

Quantum-Inspired Orthogonal Adapters for Certifiably Robust Medical Image Classification

Quantum-inspired orthogonal adapters for certifiably robust medical image classification

arXiv: 其他 AI 安全论文· arXiv:2611.18775· Olga Tsurkan, Laia Domingo, Victoria Orlova, Asel Sagingalieva, Iris Zalaudek, Raffaele Dante Caposiena CaroEigenOrbit Sàrl; Dermatology Clinic, Hospital Maggiore of Trieste, University of TriesteIndexed by this site · 2 days ago· Originally publishedDefenses· Its place in the scan record →

Translation in progress — untranslated text is omitted.

Quick read

Defenses

Based on the paper PDF (AI-generated); refer to the original

AI summaryThe researchers propose a quantum-inspired orthogonal butterfly layer fine-tuning method that preserves the Lipschitz bound when adapting pretrained certified networks. In trichoscopy image classification for alopecia diagnosis and alopecia areata activity assessment, its certified accuracy is approximately twice that of conventional fine-tuning and LoRA, with comparable classification accuracy. On MNIST and MedMNIST, certified robustness is better than conventional methods, and in DermaMNIST transfer it is more resistant to the tested attacks than LoRA.

Questions
Medical image classifiers are sensitive to small input perturbations; certified robustness can provide a mathematical guarantee of prediction stability, but conventional fine-tuning changes the Lipschitz bound of a pretrained certified network, invalidating the guarantee.
Methods
The authors implement an orthogonal adapter using a butterfly circuit composed of quantum-inspired RBS gates, left-multiplying the pretrained weights by an orthogonal matrix R(θ); singular values are unchanged, thereby preserving the Lipschitz bound; each layer requires only n/2·log2 n trainable angles, and training uses a margin loss.
Experiments and results
On the controlled transfer tasks of MNIST/MedMNIST and the BRONet dermoscopy task, butterfly maintains L/L0=1, and on the diagnostic task at ε=36/255 the certified accuracy is 63.8%, about twice that of LoRA (31.6%) and full fine-tuning (35.6%); the empirical robustness of each method under the tested attacks is similar.
Limitations and future work
The authors note: following the original study's image-level split may overestimate accuracy; the clinical datasets are small (143 and 57 patients); the attack evaluation uses only a single seed and a test subset; only the BRONet fully connected layers are adapted; and the certified backbone's accuracy is lower than that of conventional networks.

Related papers

Show 3 more papers
Read originalarxiv.org