Conv-Reluplex: A Verification Framework for Convolutional Neural Networks
Conv-Reluplex: A Verification Framework For Convolution Neural Networks
Translation in progress — untranslated text is omitted.
The researchers propose the Conv-Reluplex framework, which extends the Reluplex verification tool from deep neural networks to convolutional neural networks, can verify convolutional and pooling layers, and generates adversarial examples when classification errors occur. Experiments on MNIST show that adversarial examples generated using this tool can improve the original CNN, with classification accuracy significantly improved.
- Reluplex can only verify fully connected DNNs and cannot handle CNNs containing convolutional and pooling layers, and therefore cannot perform adversarial robustness verification on image classification networks or generate counterexamples in image form.
- Adding Unpooling and Deconvolution inversion algorithms on top of Reluplex: after obtaining intermediate adversarial examples via verification at the fully connected layer, invert them layer by layer back to the image; the generated adversarial examples are then used for adversarial training.
- On the self-built CNN on MNIST, robustness is satisfied at δ=0.2, and misclassification occurs at δ≥0.3; after adversarial training, adversarial example accuracy rises from 0% to 99.65%, and original sample accuracy drops from 94.98% to 94.89%.
- The authors state that SMT solving complexity is exponential, and currently only smaller networks can be handled (the MNIST network has more than 4000 ReLU nodes); ImageNet-scale networks have too many nodes to handle; future goals include handling color images and RNNs.