Search

As state-of-the-art deep neural networks are being deployed at the core level of increasingly large numbers of AI-based products and services, the incentive for “copying them” (i.e., their intellectual property, manifested through the knowledge that is encapsulated in them) either by adversaries or commercial competitors is expected to considerably increase over time. The most efficient way to extract or steal knowledge from such networks is by querying them using a large dataset of random samples and recording their output, which is followed by the training of a student network, aiming to eventually mimic these outputs, without making any assumption about the original networks. The most effective way to protect against such a mimicking attack is to answer queries with the classification result only, omitting confidence values associated with the softmax layer. In this paper, we present a novel method for generating composite images for attacking a mentor neural network using a student model. Our method assumes no information regarding the mentor’s training dataset, architecture, or weights. Furthermore, assuming no information regarding the mentor’s softmax output values, our method successfully mimics the given neural network and is capable of stealing large portions (and sometimes all) of its encapsulated knowledge. Our student model achieved 99% relative accuracy to the protected mentor model on the Cifar-10 test set. In addition, we demonstrate that our student network (which copies the mentor) is impervious to watermarking protection methods and thus would evade being detected as a stolen model by existing dedicated techniques. Our results imply that all current neural networks are vulnerable to mimicking attacks, even if they do not divulge anything but the most basic required output, and that the student model that mimics them cannot be easily detected using currently available techniques.

Coral reefs are undergoing a severe decline due to ocean acidification, seawater warming and anthropogenic eutrophication. We demonstrate the applicability of Deep Learning (DL) for following these changes. We examined the distribution and frequency appearance of the eleven most common coral species at four sites in the Gulf of Eilat. We compared deep learning with conventional census methods. The methods used in this research were natural sampling units via photographing the coral reef, line transects for estimating the cover percentage at the four test sites and deep convolutional neural networks, which proved to be an efficient sparse classification for coral species using the supervised deep learning method. The main research goal was to identify the common coral species at four test sites in the Gulf of Eilat, using DL to detect differences in coral cover and species composition among the sites, and relate these to ecological characteristics, such as depth and anthropogenic disturbance. The use of this method will produce a vital database to follow changes over time in coral reefs, identify trend lines and recommend remediation measures accordingly. We outline future monitoring needs and the corresponding system developments required to meet these.

We present thorough this review the developments in the field, point out their current limitations, and outline its timelines and unique potential. In order to do so we introduce the methods used in each of the advances in the application of deep learning (DL) to coral research that took place between the years: 2016−2018. DL has unique capability of streamlining the description, analysis, and monitoring of coral reefs, saving time, and obtaining higher reliability and accuracy compared with error-prone human performance. Coral reefs are the most diverse and complex of marine ecosystems, undergoing a severe decline worldwide resulting from the adverse synergistic influences of global climate change, ocean acidification, and seawater warming, exacerbated by anthropogenic eutrophication and pollution. DL is an extension of some of the concepts originating from machine learning that join several multilayered neural networks. Machine learning refers to algorithms that automatically detect patterns in data. In the case of corals these data are underwater photographic images. Based on “learned” patterns, such programs can recognize new images. The novelty of DL is in the use of state-of-art computerized image analyses technologies, and its fully automated methodology of dealing with large data sets of images. Automated Image recognition refers to technologies that identify and detect objects or attributes in a digital video or image automatically. Image recognition classifies data into selected categories out of many. We show that Neural Network methods are already reliable in distinguishing corals from other benthos and non-coral organisms. Automated recognition of live coral cover is a powerful indicator of reef response to slow and transient changes in the environment. Improving automated recognition of coral species, DL methods already recognize decline of coral diversity due to natural and anthropogenic stressors. Diversity indicators can document the effectiveness of reef bioremediation initiatives. We explored the current applications of deep learning for corals and benthic image classification by discussing the most recent studies conducted by researchers. We review the developments in the field, point out their current limitations, and outline their timelines and unique potential. We also discussed a few future research directions in the fields of deep learning. Future needs are the age detection of single species, in order to track trends in their population recruitment, decline, and recovery. Fine resolution, at the polyp level, is still to be developed, in order to allow separation of species with similar macroscopic features. That refinement of DL will allow such comparisons and their analyses. We conclude that the usefulness of future, more refined automatic identification will allow reef comparison, and tracking long term changes in species diversity. The hitherto unused addition of intraspecific coral color parameters, will add the inclusion of physiological coral responses to environmental conditions and change thereof. The core aim of this review was to underscore the strength and reliability of the DL approach for documenting coral reef features based on an evaluation of the currently available published uses of this method. We expect that this review will encourage researchers from computer vision and marine societies to collaborate on similar long-term joint ventures.

In recent years, large datasets of high-resolution mammalian neural images have become available, which has prompted active research on the analysis of gene expression data. Traditional image processing methods are typically applied for learning functional representations of genes, based on their expressions in these brain images. In this paper, we describe a novel end-to-end deep learning-based method for generating compact representations of in situ hybridization (ISH) images, which are invariant-to-translation. In contrast to traditional image processing methods, our method relies, instead, on deep convolutional denoising autoencoders (CDAE) for processing raw pixel inputs, and generating the desired compact image representations. We provide an in-depth description of our deep learning-based approach, and present extensive experimental results, demonstrating that representations extracted by CDAE can help learn features of functional gene ontology categories for their classification in a highly accurate manner. Our methods improve the previous state-of-the-art classification rate (Liscovitch, et al.) from an average AUC of 0.92 to 0.997, i.e., it achieves 96% reduction in error rate. Furthermore, the representation vectors generated due to our method are more compact in comparison to previous state-of-the-art methods, allowing for a more efficient high-level representation of images. These results are obtained with significantly downsampled images in comparison to the original high-resolution ones, further underscoring the robustness of our proposed method.