Machine learning models are increasingly being deployed in practice. Machine Learning as a Service (MLaaS) providers expose such models to queries by third-party developers through application programming interfaces (APIs). Prior work has developed model extraction attacks, in which an attacker extracts an approximation of an MLaaS model by making black-box queries to it. We design ActiveThief – a model extraction framework for deep neural networks that makes use of active learning techniques and unannotated public datasets to perform model extraction. It does not expect strong domain knowledge or access to annotated data on the part of the attacker. We demonstrate that (1) it is possible to use ActiveThief to extract deep classifiers trained on a variety of datasets from image and text domains, while querying the model with as few as 10-30% of samples from public datasets, (2) the resulting model exhibits a higher transferability success rate of adversarial examples than prior work, and (3) the attack evades detection by the state-of-the-art model extraction detection method, PRADA.

Abstract. Gupta YM, Tanasarnpaiboon S, Buddhachat K, Peyachoknagul S, Inthim P, Homchan S. 2020. Development of microsatellite markers for the house cricket, Acheta domesticus (Orthoptera: Gryllidae). Biodiversitas 21: 4094-4099. The house cricket, Acheta domesticus, is one of the species of crickets commonly found in Thailand. Insect breeders in Thailand prefer to breed house cricket as food due to its better taste and popularity among local people. Moreover, largescale breeding industries also breed house cricket to produce cricket-based edible products. Insect breeding industry is growing rapidly and requires primary precaution for sustainable production. To facilitate breeding system to maintain genetic variation in the captive population, we have sequenced the whole genome of A. domesticus to search for simple sequence repeats (SSRs) in order to develop polymorphic microsatellite markers for preliminary population genetic analysis. A total of 112,157 SSRs with primer pairs were identified in our analysis.  Of these, 91 were randomly selected to check for amplification of microsatellite polymorphisms. From these, nine microsatellites were used to check genetic variation in forty-five individuals of A. domesticus from the Phitsanulok population (Thailand).  These microsatellite markers also showed cross-amplification with other three species of edible crickets, specifically Gryllus bimaculatus, Gryllus testaceus, and Brachytrupes portentosus. The microsatellite markers presented herein will facilitate future population genetic analysis of A. domesticus populations. Moreover, the transferability of these makers would also enable researchers to conduct genetic studies for other closely related species.

Rapid and accurate species diagnosis accelerates performance in numerous biological fields and associated areas. However, morphology-based species taxonomy/identification might hinder study and lead to ambiguous results. DNA barcodes (Bar) has been employed extensively for plant species identification. Recently, CRISPR-cas system can be applied for diagnostic tool to detect pathogen’s DNA based on the collateral activity of cas12a or cas13. Here, we developed barcode-coupled with cas12a assay, “Bar-cas12a” for species authentication using Phyllanthus amarus as a model. The gRNAs were designed from trnL region, namely gRNA-A and gRNA-B. As a result, gRNA-A was highly specific to P. amarus amplified by RPA in contrast to gRNA-B even in contaminated condition. Apart from the large variation of gRNA-A binding in DNA target, cas12a- specific PAM’s gRNA-A as TTTN can be found only in P. amarus. PAM site may be recognized one of the potential regions for increasing specificity to authenticate species. In addition, the sensitivity of Bar-cas12a using both gRNAs gave the same detection limit at 0.8 fg and it was 1,000 times more sensitive compared to agarose gel electrophoresis. This approach displayed the accuracy degree of 90% for species authentication. Overall, Bar-cas12a using trnL-designed gRNA offer a highly specific, sensitive, speed, and simple approach for plant species authentication. Therefore, the current method serves as a promising tool for species determination which is likely to be implemented for onsite testing.

Abstract. Homchan S, Gupta YM. 2020. Short communication: Insect detection using a machine learning model. Nusantara Bioscience 13: 69-73. The key step in characterizing any organisms and their gender highly relies on correct identification of specimens. Here we aim to classify insect and their sex by supervised machine learning (ML) model. In the present preliminary study, we used a newly developed graphical user interface (GUI) based platform to create a machine learning model for classifying two economically important cricket species. This study aims to develop ML model for Acheta domesticus and Gryllus bimaculatus species classification and sexing. An experimental investigation was conducted to use Google teachable machine GTM for preliminary cricket species detection and sexing using pre-processed 2646 still images. An alternative method for image processing is used to extract still images from high-resolution video for optimum accuracy. Out of the 2646 images, 2247 were used for training ML model and 399 were used for testing the trained model. The prediction accuracy of trained model had 100 % accuracy to identify both species and their sex. The developed trained model can be integrated into the mobile application for cricket species classification and sexing. The present study may guide professionals in the field of life science to develop ML models based on image classification, and serve as an example for researchers and taxonomists to employ machine learning for species classification and sexing in the preliminary analysis. Apart from our main goals, the paper also intends to provide the possibility of ML models in biological studies and to conduct the preliminary assessment of biodiversity.

No abstract