Knowing whether a COTS-AI model is believable for use in a deployment context is crucial. The main concerns are whether actual input distribution matches what was used for training, whether there are some error-prone regions of input space where the AI is prone to fail. Significantly less attention has been given to address such lacuna in deployment scenarios. To fill this gap, we propose a first-of-its-kind novel solution known as Believable AI. This is the first paper to provide a theoretically strong algorithmic solution that will enable the COTS-AI buyer to (i) characterize inputs for which the model should be believed, (ii) certify its usage as believable given the risk tolerance in a given context and (iii) provide evidence based on above characterization for believing the model’s output. The proposed solution is based on a statistical analysis of the model performance on labeled inputs and inter-sample distances. We show with a wide variety of network architectures and publicly available datasets how buyers of COTS-AI can benefit from these measures, how designer/developers of AI can make claims about their products without revealing design or complete data related IPR, and how a regulatory compliance mechanism can be built to make COTS-AI products believable to use.
Iris recognition systems are widely deployed in various access control applications such as highly secured border control. Recent studies have demonstrated the vulnerability of iris sensors to various types of Presentation Attack Instruments (PAI) that include, including print attacks, prosthetic eyes, and textured contact lenses. Among these PAI, textured/patterned or color cosmetic contact lenses can be used to conceal the identity and thus can be a potential threat to iris recognition systems. In this study, we propose a novel approach to detect textured contact lens attacks by leveraging complementary information from a Multi-Spectral (MS) iris capture device. The method combines deep features extracted from AlexNet and ResNet50 through a fusion framework. The proposed approach utilizes a specialized MS iris capture device capable of capturing iris images across five distinct spectral wavelengths: 800 nm, 830 nm, 850 nm, 870 nm, and 980 nm. We then employed the proposed algorithm to extract the deep features using AlexNet and ResNet50 across five different spectral bands that are combined at the comparison score level to accurately detect textured contact lens attacks on iris recognition systems. To achieve this, we present a new dataset focused on presentation attacks, collected using a MS iris capture device and featuring two types of textured contact lenses. The new dataset comprises 2800 artefact samples corresponding to two different types of contact lenses and 1200 bona fide samples captured in five different spectral bands. Experimental evaluation demonstrates the outstanding performance of texture-based features on multi-spectral iris samples in detecting both known and unknown contact lens iris attacks.
Unconstrained low-resolution face recognition is a prominent area for researchers in computer vision. Due to the unavailability of abundant visual features, it is hard to improve the identification rate in low-resolution (LR) images. To boost the identification rate on LR images, we have proposed an attention network that internally consists of a local and global attention module. The local attention module focuses on the early features, while the global attention module focuses on deep features inside the network. To make up for the deficit of resolution aspects aware images, we have also generated multi-resolution images from high-resolution (HR) images where resolution aspects are covered. Experimenting on the TinyFace dataset, we have achieved a 54.1% identification rate with a margin of 6.2%. Furthermore, a train-test pipeline is also implemented to boost the identification rate on super-resolved tiny faces where state-of-the-art (SOTA) models fail. After applying the super-resolution (SR) approaches on tiny faces, the identification rate reaches 54.4% with a large margin of 6.5%. The identification rate achieved on the Surveillance Cameras Face (SCface) Database is 80.0% with a giant margin of 12.7% at a distance of 4.2 m. Heat maps are also visualized to locate the key discriminant areas in unconstrained LR images to witness the effectiveness of the attention network. A detailed quantitative experimental analysis is performed to produce qualitative results comparable to SOTA models.
In this paper, we propose NightSight: a pipeline towards enhancement and de-blurring of low-light blur images. The challenges of low-light conditions and blurriness commonly arise in nighttime photography. Improvements in nighttime photography enhance visibility in low-light conditions, benefitting applications in computer vision like security surveillance, photography enhancement, astronomical phenomena, scientific research and many more. Current techniques find challenges in providing seamless solutions for simultaneous low-light image enhancement (LLIE) and image de-blurring. To address this, we propose a pipeline that incorporates an independent low-light image enhancement model and a de-blurring model. We present the results of the proposed pipeline on real-world datasets and compare its performance with state-of-the-art methods.
Generative-Adversarial-Networks (GANs) have demonstrated notable/significant potential in the area/field of image-synthesis, especially in the areas of image-generation and image to image translation. In this work, we explore the capabilities of GANs through two distinct implementations. First, we utilize a regular GAN, to generate synthetic images belonging to MNIST and CIFAR-10 datasets, illustrating the network’s proficiency in creating diverse and realistic images from different data distributions. Second, we apply the Pix2Pix GAN model for image-to-image (I2I) translation, focusing on converting satellite images into maps, a task that highlights the model's capability/ability to learn intricate mappings between paired datasets. Our results underscore the versatility of GANs in generating high-quality images and performing complex image translation tasks. In this study, for generating images of MNIST handwritten digits, the discriminator has a loss of 0.57 and an accuracy of 82
