In reconstruction attacks, the goal is to reconstruct the data on which the model has performed the model learning process. In a meta-learning process, each task has its own data and does not provide it to the meta-learner. But due to the fact that in order to participate in the learning process, they have to provide information such as gradients or loss to meta-learner, the privacy of the model is compromised. In this project, the goal is to present an attack to reconstruct data in a meta-learning process.

At various stages of a learning process, the adversary may intend to violate privacy. Privacy preserving methods help to prevent privacy attacks. One of the machine learning frameworks that has been widely used in recent years is meta-learning algorithms. With the increasing applications of these algorithms, specially in applications such as personalization, the importance of model privacy has also increased. In this project, the aim is to provide a method of maintaining privacy for meta-learning algorithms that prevents membership inference attacks. One of the important points in presenting this method is to preserve a good balance between privacy and utility.

Images are highly related to graphs. As there have been lots of works in CNN domain, maybe one option is to convert works in dynamic CNN domain to graphs. Knowing exactly what has been done in the area of CNNs can help us design better models in GNN domain.

In dynamic graph neural networks, information propagation from graphs of different times is aggregated and combined with data from current time graph and the output is created. Knowing which time step has the most effect on the prediction of current time, helps to know how is our model working and predicting, which addresses the problem of interpretability

Pathological images are captured from the samples derived from tumors in different locations of the body, biopsy, to determine whether they have become malignant. It is the gold standard in determining the state of the tumor in prostate cancer, breast cancer, determining the state of thyroid, and many other fields. The images are captured from the cells on a slide and the decision is made based on the states of the cells. In most of the fields, the number of cancerous cells versus the total number of cells plays an important role in determining the state, and in some other cases, there must be a minimum number of cells in the slide to consider it as informative. Therefore it is essential to segment and count the cells automatically. In this project, we intend to do a literature review on the existing methods and datasets and do a comparative study between the existing methods. The limits of the project can also be extended to propose a novel method excelling the existing ones.

Capsule endoscopy is a new technology developed for diagnosing small bowel. The patient swallows a small capsule and it captures a video of the whole way till expulsion. Due to long videos, it takes a great effort of experts to analyze them carefully. AI which has proved its usefulness in various fields of medical sciences can be adopted to do automatic detection of anomalies to help in accurately fastening the diagnosis. In this project, we aim at detecting ulcers, bleedings, and polyps from the tagged videos of 150 patients provided by Shahid Beheshti University of Medical Sciences using interpretable machine learning models. Each video is nearly O(60,000-80,000 frames). The system should also do automatic measurements of the sizes of anomalies and prepare a summarized report and help in speeding up the operations of experts. In addition to these challenges, only the frames containing a lesion are marked but the exact locations of the lesions are not available. This puts the problem in the category of weakly supervised training in which finding the exact boundaries are important for the automatic measurements. There are also many additional challenges regarding this project including highly imbalanced data as the lesions are only present in a few frames, highly correlated images, the problem with the large number of frames per sample requiring a fast and efficient solution, and the three-dimensional input. The previous studies have not considered detecting polyps, lesions, and bleeding altogether in one model. They are also trained on videos from devices other than the ones we are using in Iran and so may not be applicable. Therefore it is essential for us to prepare a CADx system to detect these anomalies and notify them in capsule endoscopy videos.

Graph data structures are fundamental for studying connected entities in many research areas. With an increase in the number of applications where data is represented as graphs, the problem of graph generation has recently become a hot topic. However, despite its significance, conditional graph generation is relatively less explored. The MSc project can addresses the problem of class-conditional molecular graph generation that uses class labels as generation constraints. This way, we will be able to generate molecular graphs of desired classes with specific chemical properties like smells or medicinal effects.

Using a proper node ordering strategy plays a crucial role in the training of graph generation methods. The purpose of this bachelor thesis is to study current strategies and propose a novel one to outperform them.

Million packages are delivered every day due to the fast development of e-commerce. The goal of this project is proposing a spatial-temporal network for clustering the parcels based on reducing the delivery time, distance and number of vehicles for multiple destinations. Ref: Wu, Fan, and Lixia Wu. "DeepETA: a spatial-temporal sequential neural network model for estimating time of arrival in package delivery system." Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). Vol. 33. No. 01. 2019. A. C. de Araujo and A. Etemad, "End-to-End Prediction of Parcel Delivery Time With Deep Learning for Smart-City Applications," in IEEE Internet of Things Journal, vol. 8, no. 23, pp. 17043-17056, 2021.

With the ongoing popularization of online services, the digital document images have been used in various applications. Detecting forgery and similarities in these documents is the goal of this project. Ref: Zhao, Lin, Changsheng Chen, and Jiwu Huang. "Deep Learning-Based Forgery Attack on Document Images." IEEE Transactions on Image Processing 30 (2021): 7964-7979. Rahiche, Abderrahmane, and Mohamed Cheriet. "Forgery detection in hyperspectral document images using graph orthogonal nonnegative matrix factorization." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR) workshops. 2020.

Removing shadows and background in document images enhances both the visual quality and readability of digital copies of documents. Most existing shadow removal algorithms for document images use hand-crafted heuristics and are often not robust to documents with different characteristics. The goal of this project is removing the shadow and background of various document images (handwriting and digital documents (. Ref: Lin, Yun-Hsuan, Wen-Chin Chen, and Yung-Yu Chuang. "Bedsr-net: A deep shadow removal network from a single document image." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2020.