Robust Adversarial Examples as CAPTCHA Generator
Nama Peneliti (Ketua Tim)

Nur Ulfa Maulidevi



Ringkasan Kegiatan

CAPTCHA is a type of test that mainly used to tell whether the user is a human or a computer program. Recent advancement in Deep Learning (DL) or deep neural network have become better and better to solve real-life problem. A few researches show that CAPTCHAs can be attacked using AI algorithm. Image recognition as one of the many tasks that DL can solve with such high accuracy, can be used to attack image based CAPTCHAs. Despite its ability to solve complex and complicated problems, general DL algorithm has some important shortcomings that makes it quiet vulnerable. We tested the images as image CAPTCHA, presented to human users and image recognition algorithm. The scheme that will be used is as follow: nine images will be shown and a single label is given, the user needs to pick a single image that is considered suitable to the label. After the evaluation, we received 80 responds from human users. 97.5% of the users correctly picked the right image to the label and passed the CAPTCHA test. This result shows that the images produced from our method retains most of its feature thus looks natural. The images are easily recognized by human eyes. On the other hand, the result from the test to image recognition algorithm, the machine learning model failed to recognize all of the adversarial images and failed to choose the right images (0% accuracy). We also tested the images for its robustness, after the images are transformed, the images still retain its adversarial properties, as the image recognition still failed to recognize the images.



Capaian

Penerapan Karya Tulis



Testimoni Masyarakat

A few researches show that CAPTCHAs can be attacked using AI algorithm.