Adversarial Perturbation Defense on Deep Neural Networks

Deep neural networks (DNNs) have been verified to be easily attacked by well-designed adversarial perturbations. Image objects with small perturbations that are imperceptible to human eyes can induce DNN-based image class classifiers towards making erroneous predictions with high probability. Adversarial perturbations can also fool real-world machine learning systems and transfer between different architectures and datasets. Recently, defense methods against adversarial perturbations have become a hot topic and attracted much attention. A large number of works have been put forward to defend against adversarial perturbations, enhancing DNN robustness against potential attacks, or interpreting the origin of adversarial perturbations. In this article, we provide a comprehensive survey on classical and state-of-the-art defense methods by illuminating their main concepts, in-depth algorithms, and fundamental hypotheses regarding the origin of adversarial perturbations. In addition, we further discuss potential directions of this domain for future researchers.

Isolation and Characterization of Primary Retinal Microglia From the Human Post-mortem Eyes for Future Studies of Ocular Diseases

Microglia, the primary resident immunocytes in the retina, continuously function as immune system supervisors in sustaining intraocular homeostasis. Microglia relate to many diseases, such as diabetic retinopathy, glaucoma, and optic nerve injury. To further investigate their morphology and functions in vitro, a reliable culture procedure of primary human retinal microglia is necessary. However, the culture condition of microglia from the adult retina is unclear. Researchers created several protocols, but most of them were carried out on rodents and newborns. This study describes a protocol to isolate and characterize human primary retinal microglia from human post-mortem eyes. The whole procedure started with removing the retinal vessels, mechanical separation and enzymatic dissociation, filtration, and centrifugation. Then, we cultured the cell suspensions on a T-75 flask for 18 days and then shook retinal microglia from other retinal cells. We found numerous retinal microglia grow and attach to Müller cells 10 days after seeding and increase rapidly on days 14–18. Iba1 and P2RY12 were used to qualify microglia through immunofluorescence. Moreover, CD11b and P2RY12 were positive in flow cytometry, which helps to discriminate microglia from other cells and macrophages. We also observed a robust response of retinal microglia in lipopolysaccharide (LPS) treatment with proinflammatory cytokines. In conclusion, this study provides an effective way to isolate and culture retinal microglia from adult human eyes, which may be critical for future functional investigations.

Secure machine learning against adversarial samples at test time

Deep neural networks (DNNs) are widely used to handle many difficult tasks, such as image classification and malware detection, and achieve outstanding performance. However, recent studies on adversarial examples, which have maliciously undetectable perturbations added to their original samples that are indistinguishable by human eyes but mislead the machine learning approaches, show that machine learning models are vulnerable to security attacks. Though various adversarial retraining techniques have been developed in the past few years, none of them is scalable. In this paper, we propose a new iterative adversarial retraining approach to robustify the model and to reduce the effectiveness of adversarial inputs on DNN models. The proposed method retrains the model with both Gaussian noise augmentation and adversarial generation techniques for better generalization. Furthermore, the ensemble model is utilized during the testing phase in order to increase the robust test accuracy. The results from our extensive experiments demonstrate that the proposed approach increases the robustness of the DNN model against various adversarial attacks, specifically, fast gradient sign attack, Carlini and Wagner (C&W) attack, Projected Gradient Descent (PGD) attack, and DeepFool attack. To be precise, the robust classifier obtained by our proposed approach can maintain a performance accuracy of 99% on average on the standard test set. Moreover, we empirically evaluate the runtime of two of the most effective adversarial attacks, i.e., C&W attack and BIM attack, to find that the C&W attack can utilize GPU for faster adversarial example generation than the BIM attack can. For this reason, we further develop a parallel implementation of the proposed approach. This parallel implementation makes the proposed approach scalable for large datasets and complex models.

Systematic visual analysis of groundwater hydrographs: potential benefits and challenges

Visual analysis of time series in hydrology is frequently seen as a crucial step to becoming acquainted with the nature of the data, as well as detecting unexpected errors, biases, etc. Human eyes, in particular those of a trained expert, are well suited to recognize irregularities and distinct patterns. However, there are limits as to what the eye can resolve and process; moreover, visual analysis is by definition subjective and has low reproducibility. Visual inspection is frequently mentioned in publications, but rarely described in detail, even though it may have significantly affected decisions made in the process of performing the underlying study. This paper presents a visual analysis of groundwater hydrographs that has been performed in relation to attempts to classify groundwater time series as part of developing a new concept for prediction in data-scarce groundwater systems. Within this concept, determining the similarity of groundwater hydrographs is essential. As standard approaches for similarity analysis of groundwater hydrographs do not yet exist, different approaches were developed and tested. This provided the opportunity to carry out a comparison between visual analysis and formal, automated classification approaches. The presented visual classification was carried out on two sets of time series from central Europe and Fennoscandia. It is explained why and where visual classification can be beneficial but also where the limitations and challenges associated with the approach lie. It is concluded that systematic visual analysis of time series in hydrology, despite its subjectivity and low reproducibility, should receive much more attention.

Physical Characterization and In Vitro Toxicity Test of PDMS Synthesized from Low-Grade D4 Monomer as a Vitreous Substitute in the Human Eyes

Polydimethylsiloxane (PDMS) is one of the most superior materials and has been used as a substitute for vitreous humor in the human eye. In previous research, we have succeeded in producing PDMS with low and medium viscosity using octamethylcyclotetrasiloxane (D4) monomer with a low grade of 96%. Both have good physical properties and are comparable to commercial product PDMS and PDMS synthesized using D4 monomer with a high grade of 98%. An improvement of the synthesis process is needed to ensure that PDMS synthesized from a low-grade D4 monomer under specific synthesis conditions can repeatedly produce high-quality PDMS. Apart from good physical properties, the PDMS as a substitute for vitreous humor must also be safe and not cause other disturbances to the eyes. Here, we reported the process of synthesizing and characterizing the physical properties of low- and medium-viscosity PDMS using a low-grade D4 monomer. We also reported for the first time the in vitro toxicity test using the Hen’s Egg Test Chorioallantoic Membrane (HET-CAM) test method. We have succeeded in obtaining PDMS with viscosities of 1.15 Pa.s, 1.17 Pa.s, and 1.81 Pa.s. All samples have good physical properties such as refractive index, surface tension, and functional groups that are similar to commercial PDMS. The HET-CAM test results showed that all samples did not show signs of irritation indicating that samples were non-toxic. From the results of this study, it can be concluded that PDMS synthesized from a low-grade D4 monomer under specific synthesis conditions by the ROP method is very safe and has the potential to be developed as a substitute for vitreous humor in human eyes.

Research on Video Quality Diagnosis Technology Based on Artificial Intelligence and Internet of Things

The remote video diagnosis system based on the Internet of Things is based on the Internet of Things and integrates advanced intelligent technology. To better promote a harmonious society, constructing a video surveillance system is accelerating in our country. Many enterprises and government agencies have invested much money to build video surveillance systems. The quality of video images is an important index to evaluate the video surveillance system. However, as the number of cameras continues to increase, the monitoring time continues to extend. In the face of many cameras, it is not realistic to rely on human eyes to diagnose video-solely quality. Besides, due to human eyes’ subjectivity, there will be some deviation in diagnosis through human eyes, and these factors bring new challenges to system maintenance. Therefore, relying on artificial intelligence technology and digital image processing technology, the intelligent diagnosis system of monitoring video quality is born using the computer’s efficient mathematical operation ability. Based on artificial intelligence, this paper focuses on studying video quality diagnosis technology and establishes a video quality diagnosis system for video definition detection and noise detection. This article takes the artificial intelligence algorithm in the diagnosis of video quality effect. Compared with the improved algorithm, the improved video quality diagnosis algorithm has excellent improvement and can well finish video quality inspection work. The accuracy of the improved definition evaluation function for the definition detection of surveillance video and noise detection is as high as 95.56%.

Detection of SARS-CoV-2-RNA in post-mortem samples of human eyes

Purpose To detect SARS-CoV-2 RNA in post-mortem human eyes. Ocular symptoms are common in patients with COVID-19. In some cases, they can occur before the onset of respiratory and other symptoms. Accordingly, SARS-CoV-2 RNA has been detected in conjunctival samples and tear film of patients suffering from COVID-19. However, the detection and clinical relevance of intravitreal SARS-CoV-2 RNA still remain unclear due to so far contradictory reports in the literature. Methods In our study 20 patients with confirmed diagnosis of COVID-19 were evaluated post-mortem to assess the conjunctival and intraocular presence of SARS-CoV-2 RNA using sterile pulmonary and conjunctival swabs as well as intravitreal biopsies (IVB) via needle puncture. SARS-CoV-2 PCR and whole genome sequencing from the samples of the deceased patients were performed. Medical history and comorbidities of all subjects were recorded and analyzed for correlations with viral data. Results SARS-CoV-2 RNA was detected in 10 conjunctival (50%) and 6 vitreal (30%) samples. SARS-CoV-2 whole genome sequencing showed the distribution of cases largely reflecting the frequency of circulating lineages in the Munich area at the time of examination with no preponderance of specific variants. Especially there was no association between the presence of SARS-CoV-2 RNA in IVBs and infection with the variant of concern (VOC) alpha. Viral load in bronchial samples correlated positively with load in conjunctiva but not the vitreous. Conclusion SARS-CoV-2 RNA can be detected post mortem in conjunctival tissues and IVBs. This is relevant to the planning of ophthalmologic surgical procedures in COVID-19 patients, such as pars plana vitrectomy or corneal transplantation. Furthermore, not only during surgery but also in an outpatient setting it is important to emphasize the need for personal protection in order to avoid infection and spreading of SARS-CoV-2. Prospective studies are needed, especially to determine the clinical relevance of conjunctival and intravitreal SARS-CoV-2 detection concerning intraocular affection in active COVID-19 state and in post-COVID syndrome.

Photoacoustic Imaging of Enucleated Eyes from Patients with Uveal Melanoma can Reveal Extrascleral Growth

Background: Uveal melanoma is treated by either enucleation (removal of the eye) or local eye-sparing therapies, depending on tumor size and whether there are signs of extrascleral growth. Photoacoustic (PA) imaging is a novel imaging modality that provides high-resolution images of the molecular composition of tissues. Objective: In this study, the feasibility of PA imaging for uveal melanomas and detection of extrascleral growth was explored. Methods: Seven enucleated human eyes with uveal melanomas were examined using PA imaging. The spectral signatures of the melanomas and the layers of the normal eyewall were characterized using 59 excitation wavelengths from 680 to 970 nm. Results: Significant differences were seen between the spectra obtained from melanoma and the healthy eyewall. Using spectral unmixing, melanin, hemoglobin and collagen could be mapped out, showing the architecture of the tumor in relation to the eyewall. This allowed visualization of regions where the tumor extended into the extrascleral space. Conclusion: PA imaging appears to have the potential to aid in assessing uveal melanomas and as a diagnostic tool for the detection of extrascleral growth.