40 research outputs found
S.U.S. You're SUS! - Identifying Influencer Hackers on Dark Web Social Networks
No full textDark web is an obscured part of the Internet, wrapped under various layers of routing making it attractive for benign usage such as anonymity and security as well as a key platform for sharing exploits, data breaches, and other means of cybercrime. Dark web forums provide opportunities to share such data and exploits similar to the social media forums within the public Internet. Users of such forums earn reputation and credibility through their participation in discussions and sharing data, exploits, and hacks. Such activities can help develop metrics to enable identification of influential mal-actors facilitating efficient and effective defence against emerging cyber threats in general and zeroday exploits in particular. In this paper, we propose a novel framework (INSPECT) to detect influential entities through intelligent analysis of user-profiles, interactions, and activities over dark web forums. INSPECT framework involves Feature Engineering, Social Network Analysis, Text Mining, Semantic Analysis, and K-means clustering and calculates an influencer score which represents the significance of the users within the dark web forum. We have used the CrimeBB dataset comprising user profiles and activities within dark web forums to evaluate effectiveness of the INSPECT framework to identify influential users on the dark web forums
Machine learning security and privacy: a review of threats and countermeasures
No full textAbstract Machine learning has become prevalent in transforming diverse aspects of our daily lives through intelligent digital solutions. Advanced disease diagnosis, autonomous vehicular systems, and automated threat detection and triage are some prominent use cases. Furthermore, the increasing use of machine learning in critical national infrastructures such as smart grids, transport, and natural resources makes it an attractive target for adversaries. The threat to machine learning systems is aggravated due to the ability of mal-actors to reverse engineer publicly available models, gaining insight into the algorithms underpinning these models. Focusing on the threat landscape for machine learning systems, we have conducted an in-depth analysis to critically examine the security and privacy threats to machine learning and the factors involved in developing these adversarial attacks. Our analysis highlighted that feature engineering, model architecture, and targeted system knowledge are crucial aspects in formulating these attacks. Furthermore, one successful attack can lead to other attacks; for instance, poisoning attacks can lead to membership inference and backdoor attacks. We have also reviewed the literature concerning methods and techniques to mitigate these threats whilst identifying their limitations including data sanitization, adversarial training, and differential privacy. Cleaning and sanitizing datasets may lead to other challenges, including underfitting and affecting model performance, whereas differential privacy does not completely preserve model’s privacy. Leveraging the analysis of attack surfaces and mitigation techniques, we identify potential research directions to improve the trustworthiness of machine learning systems
The Visitors
No full text157 pagesMFA theses in English Language and Literature are not available for direct download. Users wishing to access an MFA thesis in this collection may request access by clicking the link to the restricted file(s) and completing the request form. If we have contact information for the author, we will contact them and request permission to provide access. If we do not have contact information or the author denies or does not respond to our inquiry, we will not be able to provide access.A novel that follows four women from Karachi, Pakistan, estranged friends who come back together in the wake of a tragedy in one of their lives.10000-01-0
Threat Miner - A Text Analysis Engine for Threat Identification Using Dark Web Data
No full textCyber threats continue to grow with novel methods to attack computing systems, highlighting the need for sophisticated mechanisms and techniques to protect against such dynamic threats. Contemporary cyber defence mechanisms utilise a range of methods which rely on monitoring network or system-level events. However, with the growing use of the dark web by mal-actors to share exploits, breaches, and data leaks, the use of such information to strengthen defence mechanisms becomes an intriguing prospect. In this paper, we present our efforts to develop a text mining engine (Threat Miner) which analyses data from dark web forums and transforms it into actionable intelligence. Leveraging cutting-edge machine learning techniques and utilising a bespoke threat dictionary, Threat Miner extracts useful information from dark web forums into STIX form, enabling it to be used with threat intelligence platforms. We also present the results of a thorough evaluation
of our scheme which was conducted with the CrimeBB dataset to understand the feasibility of the approach as well as its effectiveness in strengthening defence capability against cyber threats
Outlier-oriented poisoning attack: a grey-box approach to disturb decision boundaries by perturbing outliers in multiclass learning
No full textPoisoning attacks are a primary threat to machine learning (ML) models, aiming to compromise their performance and reliability by manipulating training datasets. This paper introduces a novel attack—outlier-oriented poisoning (OOP) attack, which manipulates labels of most distanced samples from the decision boundaries. To ascertain the severity of the OOP attack for different degrees (5–25%) of poisoning to conduct a detailed analysis, we analyzed variance, accuracy, precision, recall, f1-score, and false positive rate for chosen ML models. Benchmarking the OOP attack, we have analyzed key characteristics of multiclass machine learning algorithms and their sensitivity to poisoning attacks. Our analysis helps understand behaviour of multiclass models against data poisoning attacks and contributes to effective mitigation against such attacks. Utilizing three publicly available datasets: IRIS, MNIST, and ISIC, our analysis shows that KNN and GNB are the most affected algorithms with a decrease in accuracy of 22.81% and 56.07% for IRIS dataset with 15% poisoning. Whereas, for same poisoning level and dataset, Decision Trees and Random Forest are the most resilient algorithms with the least accuracy disruption (12.28% and 17.52%). We have also analyzed the correlation between number of dataset classes and the performance degradation of models. Our analysis highlighted that number of classes are inversely proportional to the performance degradation, specifically the decrease in accuracy of the models, which is normalized with increasing number of classes. Further, our analysis identified that imbalanced dataset distribution can aggravate the impact of poisoning for machine learning models
Deep Behavioral Analysis of Machine Learning Algorithms Against Data Poisoning
No full textPoisoning attacks represent one of the most common and practical adversarial attempts on machine learning systems. In this paper, we have conducted a deep behavioural analysis of six machine learning (ML) algorithms, analyzing poisoning impact and correlation between poisoning levels and classification accuracy. Adopting an empirical approach, we highlight practical feasibility of data poisoning, comprehensively analyzing factors of individual algorithms affected by poisoning. We used public datasets (UNSW-NB15, BotDroid, CTU13, and CIC-IDS-2017) and varying poisoning levels (5% - 25%) to conduct rigorous analysis across different settings. In particular, we analyzed the accuracy, precision, recall, f1-score, false positive rate and ROC of the chosen algorithms. Further, we conducted a sensitivity analysis of each algorithm to understand the impact of poisoning on its performance and characteristics underpinning its susceptibility against data poisoning attacks. Our analysis shows that, for 15% poisoning of UNSW NB15 dataset, the accuracy of Decision Tree (DT) decreases by 15.04% with an increase of 14.85% in false positive rate. Further, with 25% poisoning of BotDroid dataset, accuracy of K-nearest neighbours (KNN) decreases by 15.48%. On the other hand, Random Forest (RF) is comparatively more resilient against poisoned training data with a decrease of 8.5% in accuracy with 15% poisoning of UNSW-NB15 dataset and 5.2% for BotDroid dataset. Our results highlight that 10%-15% of dataset poisoning is the most effective poisoning rate, significantly disrupting classifiers without introducing overfitting, whereas 25% is detectable because of high performance degradation and overfitting algorithms. Our analysis also helps understand how asymmetric features and noise affect the impact of data poisoning on machine learning classifiers. Our experimentation and analysis are publicly available at: https://github.com/AnumAtique/Behavioural-Analaysis-of Poisoned-ML
V kolì Andriâ Šeptic'kogo : Ukraïns'ke kul'turne vìdrodžennâ za časìv mecenatstva mitropolita
No full textTekst jest pierwszym w języku polskim studium o mecenacie Andrzeja Szeptyckiego (1865-1944), greckokatolickiego metropolity lwowskiego (1900-1944). Autor opisuje dzieła Szeptyckiego: szkolę ikon, Narodowe Muzeum Ukraińskie we Lwowie, mecenat nad ukraińską sztuką nowoczesną, ukraińskie organizacje artystyczne (SPUOM, GDUM, ANUM) i wystawy sztuki ukraińskiej w czasie II wojny światowej we Lwowie. Tekst zawiera unikatowy wykaz ok. 150 artystów ukraińskich z kręgu Andrzeja Szeptyckiego.The text is the first study in Polish on art patronage by Andrej Sheptyckyj (1865-1944), the Greek Catholic metropolite of Lwow (1900-1944). The author describes Sheptyckyj's works: the school of icons, the National Ukrainian Museum in Lwow, patronage for Ukrainian modern art, Ukrainian artistic organizations (SPUOM, GDUM, ANUM) and Ukrainian art exhibitions in Lwow in the time of the World War II. The text contains the unique register (the list) of circa 150 Ukrainian artists from the Andrej Sheptyckyj's circle
Alienation and the Dilemma of Man in Eugene O’Neill’s The Hairy Ape
No full textEugene O\u27 Neill, an American playwright was born into a troubled and an upset family on October 16, 1888. Eugene O’ Neill had a quite precarious, wobbly and uneven adolescence as his elder brother was an affirmed alcoholic whereas his mom was a drug addict. This research paper analyzes the alienation, dilemma and the futile struggle of man in the quest of his identity. O\u27Neill followed the course of a superior and advanced writer looking for a profound focus in the entirety of his significant works. His perspective on humankind in his dramatizations is basically sad and heartbreaking. The author needed to cause man to feel free from all worries and inhale outside fresh air and build up a feeling of having a place in the general public in which he lived. However, it was impractical
Automated Segmentation of Metastatic Lymph Nodes in Lymphoma Patients
No full textVed å bruke det kunstige nevrale nettverket 2D U-Net, tester denne masteroppgaven nøyaktigheten til 2D U-Net med formålet om å automatisk segmentere ondartede svultser i PET/MR-bilder av pasienter med metastatisk lymfom.
For Hodgkin- og Non-Hodgkin-lymfom er FDG PET/MR-segmenteringene viktige for prognose, stadieinndeling (staging) og responsvurdering av lymfompasienter. Manuelle segmenteringer er imidlertid tidkrevende og vanskelige i komplekse pasienttilfeller der en har høy sykdomsbyrde. Målet med dette prosjektet er å utvikle en automatisert metode for segmentering av kreftrammede lymfeknuter i PET/MR ved bruk av dyp læring (deep learning), nærmere bestemt et dypt kunstig nevralt nettverk. FDG PET/MR-baseline, interim- og behandlingsavslutningsbilder (EOT) av Hodgkin- og Non-Hodgkin-lymfompasienter ble analysert. To grupper radiologer og nukleærmedisinere har bidratt med klinisk lesing av PET/MR-bildene etter standardiserte protokoller. Imidlertid manglet de faktiske segmenteringene, m.a.o segmenterings fasiten, fra lymfomdatasettet, og disse var avgjørende for å implementere dyplæringsnettverket for en automatisert segmenteringsprosess. De manuelle segmenteringene som krevdes ble utført av forfatteren og validert av en nukleærmedisiner fra St.Olavs Hospital.
Den nevrale nettverksmodellen ble lært hvordan man utfører klassifiseringsoppgaver direkte fra bilder, dvs. nettverket ble opplært til å gjenkjenne mønstre fra et datasett bestående av 64 PET/MR-undersøkelser. Et 3-kanals multimodalt bilde, et RGB-bilde, bestående av PET, T2-HASTE og DWI med b = 800 s/mm^{2} ble brukt som input for algoritmen, og modellen ble lært til å gjenskape segmenteringene i grunnsannheten (segmenterings fasiten) ved å bruke en 2D U-Net-arkitektur.
Videre ble lymfomdatasettet delt inn i et 85/15-forhold for trening og testing som bestod av henholdsvis 53 og 11 PET/MR-undersøkelser. Både en 4-fold og 13-fold kryssvalidering ble utført for opplæringen av modellen. Valideringen resulterte i gjennomsnittlige Dice-score (overlappingsmål) på henholdsvis 0,61 og 0,63 for 4-fold og 13-fold modellene. Flere andre evaluaringer som tap, nøyaktighet, presisjon, tilbakekalling, negativ prediktiv verdi (NPV) og spesifisitet ble inkludert for voxel-nivåanalysen. Resultatene var 0,011, 0,97, 0,83, 0,11, 0,97, 0,99, henholdsvis for 4-fold valideringen og 0,065, 0,97, 0,90, 0.10, 0,97, 0,99, henholdsvis for 13-fold valideringen. Den gjennomsnittlige dice scoren til testpasientene var henholdsvis 0,29 og 0,32 for 4-fold og 13-fold kryssvalidering, noe som antyder en dårligere ytelse på nye og usette pasienter sammenlignet med PET/MR-undersøkelsene brukt i valideringen. Til tross for de generelt høye verdiene for evalueringsmetodene, ga den voxelbaserte analysen ingen god indikasjon på hvor nøyaktig modellen klarte å segmentere kreftlesjoner ettersom flertallet av vokslene i pasientene ble klassifisert som ekte negative (TN). Derfor ble det utført en lesjonsbasert analyse, og den avslørte at modellen ofte segmenterte færre lesjoner enn det som var tilstede i grunnsannheten. Dette indikerte at modellens hovedbegrensning var antallet falske negative predikerte kreftsvultser. Som en konsekvens, presterer modellen bedre på valideringsdataene enn for testdatasettet som ble ekskludert fra opplæringen.
For å konkludere, så segmenterer den trente 2D U-Net-modellen automatisk ondartede lymfeknutesvultser i 3-kanals multimodale bilder. Fremtidig arbeid bør fokusere på å forbedre overlappingsmålet dice score, co-registreringen, redusere antall uoppdagede tumorlesjoner, samt øke datasettet for å sikre en større variasjon i kohorten. Dette kan forbedre både treningen og resultatene.Using the deep learning artificial neural network 2D U-Net, this project tests the accuracy of the 2D U-Net for the purpose of automatically segmenting malignant lesions in PET/MR images of patients with metastatic lymphoma.
For Hodgkin and Non-Hodgkin lymphoma, the FDG PET/MRI segmentations are important for prognosis, staging, and response assessment of lymphoma patients. However, manually segmentations are time-consuming and difficult in complex patient cases and for high disease burden. The aim of this project is to develop an automated method for segmentation of cancer-affected lymph-nodes in PET/MRI using a deep neural network.
FDG PET/MRI baseline, interim, and End-Of-Treatment (EOT) images of Hodgkin and Non-Hodgkin lymphoma patients were analyzed. Two groups of radiologist and nuclear medicine physicians have contributed with clinical reading of the PET/MR images following standardized protocols. However, the segmentation ground truth was missing from the lymphoma dataset, and it was crucial for implementing the deep learning network for an automated segmentation process. The manual segmentation required has therefore been performed by the author and validated by a nuclear medicine physician from St. Olavs Hospital.
The neural network model was taught how to perform classification tasks directly from images, i.e., the network was trained to recognize patterns from a dataset consisting of 64 PET/MRI examinations. A 3-channel multi-modal image, i.e., an RGB image, consisting of a PET, a T2-HASTE, and a DWI with b = 800 s/mm^{2} was used as input for the algorithm. The model was trained to replicate the segmentations of the ground truth by using a 2D U-Net architecture.
Furthermore, the lymphoma dataset was divided in a 85/15 ratio for training and testing consisting of 53 and 11 PET/MRI examinations, respectively. Both a 4-fold and 13-fold cross-validation were performed for the training of the model. The validation resulted in average dice scores of 0.61 and 0.63 respectively for the 4-fold and 13-fold trained models. Several other metrics such as loss, accuracy, precision, recall, Negative Predictive Value (NPV), and specificity were included for the voxel level analysis. The scores were 0.011, 0.97, 0.83, 0.11, 0.97, 0.99, respectively for the 4-fold validation and 0.065, 0.97, 0.90, 0.10, 0.97, 0.99, respectively for the 13-fold validation. The average dice score of the testing patient were 0.29 and 0.32 respectively for the 4-fold and 13-fold cross-validation which suggested an inferior performance on unseen patients compared to the PET/MRI examinations used in the validation. Despite the overall high scores for the evaluation metrics, the voxel based analysis did not give a great indication of how well the model managed to segment cancer lesions due to the majority of the voxels in a patient being classified as true negative. Therefore, a lesion-based analysis were conducted and it revealed that the model often segmented fewer lesions than in the ground truth. This indicated that the model's main limitation was the number of false negative predicted lesions. As a consequence, the model performs better on the validation data than for the testing dataset which was excluded from the training.
In conclusion, the trained 2D U-Net model automatically segments malignant lymph node lesions in the 3-channel multi-modal images. However, future research should focus on improving the dice score, co-registration, decrease the number of undetected tumor lesions, and increase the dataset to ensure a larger variation in the cohort. This will benefit the training and yield better results
