Computer Vision News - October 2020

7 For the training of each detector, different classifiers were used: a random forests classifier for DFeat, an SVM classifier for QFeat, and finally a logistic regression classifier for both KD and LID features. Finally, for evaluating detection performance, the AUC (Area Under Curve) score is adopted. Application in Medical Imaging The experiments on adversarial attacks and detection are conducted on available medical benchmark datasets for three different applications, namely classification of diabetic retinopathy (a type of eye disease) from retinal fundoscopy, classification of thorax diseases from Chest X-rays, and classification of melanoma (a type of skin cancer) from dermoscopic photographs. The authors divided the dataset in four subsets, including train/test for the DNN model and adversarial attacks generator, and AdvTrain/AdvTest for training adversarial detectors. Understanding Adversarial Attacks on Deep Learning Based Medical Image Analysis Systems The methods used for adversarial detection consist in extracting 1) the deep features (and quantized deep features) at the second-last dense layer of the network (“DFeat”/“QFeat”), 2) the kernel density (KD) estimated from the second-last layer deep features, and 3) the local intrinsic dimensionality (LID) estimated from the output at each layer of the network features. The following equations are related to estimation of KD and LID for a given point x in a class k. ( ) = 1 | | ∑ exp( || ( , ) − ( ′ , )|| 2 2 2 ) ′ ̂ ( ) = − ( 1 ∑log ( ) ( ) =1 ) −1