Reza Babaei

Deepfake technology, which employs advanced generative artificial intelligence to create hyper-realistic synthetic media, poses significant challenges across various sectors, including security, entertainment, and education. This literature review explores the evolution of deepfake generation methods, ranging from traditional techniques to state-of-the-art models such as generative adversarial networks and diffusion models. We navigate through the effectiveness and limitations of various detection approaches, including machine learning, forensic analysis, and hybrid techniques, while highlighting the critical importance of interpretability and real-time performance in detection systems. Furthermore, we discuss the ethical implications and regulatory considerations surrounding deepfake technology, emphasizing the need for comprehensive frameworks to mitigate risks associated with misinformation and manipulation. Through a systematic review of the existing literature, our aim is to identify research gaps and future directions for the development of robust, adaptable detection systems that can keep pace with rapid advancements in deepfake generation.

In the rapidly evolving landscape of cybersecurity, model extraction attacks pose a significant challenge, undermining the integrity of machine learning models by enabling adversaries to replicate proprietary algorithms without direct access. This paper presents a comprehensive study on model extraction attacks towards image classification models, focusing on the efficacy of various Deep Q-network (DQN) extensions for enhancing the performance of surrogate models. The goal is to identify the most efficient approaches for choosing images that optimize adversarial benefits. Additionally, we explore synthetic data generation techniques, including the Jacobian-based method, Linf-projected Gradient Descent (LinfPGD), and Fast Gradient Sign Method (FGSM) aiming to facilitate the training of adversary models with enhanced performance. Our investigation also extends to the realm of data-free model extraction attacks, examining their feasibility and performance under constrained query budgets. Our investigation extends to the comparison of these methods under constrained query budgets, where the Prioritized Experience Replay (PER) technique emerges as the most effective, outperforming other DQN extensions and synthetic data generation methods. Through rigorous experimentation, including multiple trials to ensure statistical significance, this work provides valuable insights into optimizing model extraction attacks.

In recent decades, there have been significant advancements in medical diagnosis and treatment techniques. However, there is still much progress to be made in effectively managing a wide range of diseases, particularly cancer. Timely diagnosis of cancer remains a critical step towards successful treatment, as it significantly impacts patients’ chances of survival. Among various types of cancer, glioma stands out as the most common primary brain tumor, exhibiting different levels of aggressiveness. One of the monitoring techniques is magnetic resonance imaging (MRI) that provides a precise visual representation of the tumor and its sub-regions (edema (ED), enhancing tumor (ET), and non-enhancing necrotic tumor core (NEC)), enabling monitoring of its location, shape, and sub- regional characteristics. In this study, we aim to investigate the underlying relationship between the maximumdiameters of tumor sub-regions and patients’ overall survival (OS) in glioblastoma cases. Using an MRI dataset of glioblastoma patients, we categorized them based on resection status: gross total resection (GTR) and unknown (NA). By employing the Euclidean distance algorithm, we estimated sub-regions’ maximum diameters. Machine learning algorithms were used to explore the correlation between sub-regions’ maximum diameters and survival outcomes. The results of the univariate prediction models showed that tumor sub-regions’ maximum diameters have a noticeable correlation with the survival rates among patients with unknown resection status with the average spearman correlation of -0.254. Also, addition of the sub-regions’ maximum diameter feature to the radiomics increased the accuracy of ML algorithms in predicting the survival rates with an average of 4.58%.

Cancer stands out as one of the fatal diseases people are facing all the time. Each year, a countless number of people die because of the late diagnosis of cancer or wrong treatments. Glioma, one of the most common primary brain tumors, has different aggressiveness and sub-regions, which can affect the risk of disease. Although prediction of overall survival based on multimodal magnetic resonance imaging (MRI) is challenging, in this study, we assess if and how location-based features of tumors can affect overall survival prediction. This approach is evaluated independently and in combination with radiomic features. The process is carried out on a data set entailing MRI images of patients with glioblastoma. To assess the impact of resection status, the data set is divided into two groups, patients were reported as gross total resection and unknown resection status. Then, different machine learning algorithms were used to evaluate how location features are linked with overall survival. Results from regression models indicate that location-based features have considerable effects on the patients’ overall survival independently. Additionally, classifier models show an improvement in prediction accuracy by the addition of location-based features to radiomic features.