A statistical study on adversarial machine learning

In this thesis, some of the methods and basic concepts of machine learning are reviewed. Some of the algorithms are summarized with given examples. The approaches, methods and evaluation processes of machine learning and statistics discipline are briefly compared on the basis of the terminology they use. Use of artificial neural network method in supervised learning has been highlighted. Adversarial machine learning is studied by using handwritten numbers in MNIST dataset which are widely used in pattern and shape recognition problems. The recognition of handwritten numbers task is chosen as an example of adversarial machine learning application by considering white box, no-target, evasion type attack environment. In this example, machine learning was tried to be mislead by making the learned system misclassify. Attack success was evaluated by comparing machine and human perception and aesthetics, and in terms of attack execution speed. The attack algorithm was created using the PSNR (Peak Signal Noise Ratio) criterion and the Maclaurin approach of the tangent hyperbolic classification function proposed in Carlini & Wagner (2017). The new method worked 25.53% faster than the Carlini & Wagner attack method, but the prediction error rate decreased by 11.6%.