17 February 2020 in IEEE Access
In the past decade, Deep Neural Networks (DNNs) have achieved breakthrough collaborations in developing smart intelligent systems within the field of computer vision, natural language processing, autonomous systems, etc. Recent research has revealed that stability of such smart systems is at greater risk when they come across to adversarial perturbations. Although, these perturbations may not be perceivable in nature when seen from naked eye, yet, they are capable enough to fool state-of-the-art DNN classifiers. Till now, much of the previous work related to fool such classifiers focuses on generating adversaries that directly change pixel values of an image in spatial-domain. In this paper, we propose a novel transform-domain imperceptible attack methodology "TDIAM" to generate adversaries based on image steganography-approach using a "single carefully selected targeted watermark". We use three different frequency-domain approaches, i.e., Discrete Wavelet Transform (DWT), Discrete Cosine Transform (DCT) and Fast Fourier Transform (FFT) to craft perturbations in selective frequency component which makes it robust and it requires less computational time as it is a non-gradient approach. We present our case study on MNIST handwritten digits dataset. Our results demonstrate that the generated perturbation vector successfully fool simple Convolutional Neural Network (CNN), LeNet-5 and AlexNet architectures by increasing probability of adversarial examples for the targeted class (to which the targeted watermark belongs) in both "black-box" and "white-box" adversarial attacks. The results have shown that among these three perturbation approaches, DWT based perturbation shown promising results by effectively fooling DNNs while ensuring the high imperceptibility as well.
23 October 2014 in IEEE Sensors Journal
The pattern recognition problem for real-life applications of gas identification is challenging due to the limited amount of data existing and the sequential variability of the mechanism mostly caused by drift and the real-time detection. These problems are commonly caused by the slow response of most of the gas sensors. In this paper, a novel gas identification approach based on the cluster-k-nearest neighbor (C-k-NN) is introduced. The effectiveness of this approach has been successfully demonstrated on the experimental data set obtained from array of gas sensors. Our classification takes advantages of both the k-NN, which is highly accurate, and the k-means cluster, which is able to reduce the classification time. In order to increase the accuracy rate, a new feature selection method is proposed. The selection of features is based on their ability to separate and distinguish between different classes. Advanced statistical metrics are introduced to quantify the classification contribution of each feature. Mostly, classifiers are suffering from misclassification detection; new statistical metrics are introduced to estimate the exactness of the classifier response, i.e., to detect the misclassification. To enhance the classification performances for gas identification, a new tree classification design is introduced, named tree C-k-NN. In order to assess the technique, experiments were conducted on six different gases. Accuracy rate of 98.7% has been obtained with the C-k-NN and 100% with the tree C-k-NN. The performance of this approach is also validated using three publicly available data sets.
17 July 2020 in IEEE Access
Urdu word spotting is among the most challenging tasks in image processing and word spotting of hand written Urdu text is even more so. When it comes to handwritten Urdu documents, variation among the same words of various writers is significant. The orientation and style of the handwriting makes it really challenging for a word spotting system to correctly recognize the instances of the keyword. In this research, we tend to overcome this hurdle. We propose a system that takes a database of hand written Urdu text and generates random, yet, similar images to improve the classifier’s ability to recognize variations caused by difference in handwriting. For image generation, we used geometric transformations and variants of Generative Adversarial Network (GAN). For the word spotting process, Histogram of Oriented Gradients (HOG) features are extracted from ligature images and then used to train a Long Short-Term Memory (LSTM) network for the classification task. This is the first study that focuses on improving word spotting by generating arbitrary samples using GANs and its variants. The system achieved a promising recognition rate of 98.96% due to the sample generation using Cycle-GANs.