By @dnl0x00 State-of-the-art computer vision models are vulnerable to adversarial examples, i.e. for every image that is classified correctly you can find very small perturbations which causes the model to misclassify the image. The perturbations are so small that the original image and the modified image are indistinguishable for a... [Read More]

By @dnl0x00 On August a paper was published which describes ways to inject backdoors into machine learning models. This might be a problem especially for deep learning networks which are usually used for image recognition, speech recognition and machine translation because these networks are usually computationally expensive (building a model... [Read More]

By @dnl0x00 Here’s a short summary of a paper three machine learning researchers from the Kyushu University in Japan have published in October 2017 in which they describe a novel technique based on differential evolution to create adversarial images to force deep neural networks to misclassify objects shown in natural... [Read More]

By @dnl0x00 Today, I run into docker DNS problems. Whenever I tried to build an image with docker run I got an error message which said: docker build --no-cache -t ml/3 -f /tmp/tmp.FxOR4qC1OY/Dockerfile /tmp/tmp.FxOR4qC1OY Sending build context to Docker daemon 6.144kB Step 1/55 : FROM ubuntu:17.10 Get https://registry-1.docker.io/v2/: dial tcp:... [Read More]

By @dnl0x00 I’ve recently created a small Python library mnistdb which can be used to easily load the MNIST database of handwritten digits in Python into numpy arrays without any manual effort. It works for Python 2 and Python3. Just install the library via pip: pip install mnistdb Here’s an... [Read More]