http://repository.iiitd.edu.in/xmlui/handle/123456789/1017 Year-2021 2026-04-10T22:38:16Z http://repository.iiitd.edu.in/xmlui/handle/123456789/1635 Sinkhorn attacks Chaudhuri, Dibyendu Roy; Subramanyam, A V (Advisor) Adversarial attacks have been extensively investigated in the recent past. Quite interestingly, a majority of these attacks primarily work in the lp space. In this work, we propose a novel approach for generating adversarial samples using Wasserstein distance. Existing Wasserstein distance-based works generate adversarial samples using balanced optimal transport (OT). However, balanced OT requires input marginals to be of the same total probability masses these precluding its immediate application to images. Motivated by the recent unbalanced OT theory, we propose a UOT based adversarial threat model with relaxed marginal equality constraints. Our experiments on retrieval and classification tasks demonstrate significantly stronger attacks with better image quality as well as less computational overhead. 2021-07-01T00:00:00Z http://repository.iiitd.edu.in/xmlui/handle/123456789/1096 An evaluation of multi-user OFDMA performance in WiFi 6 and its optimization for deadline constrained settings Dev, Harshal; Maity, Mukulika (Advisor); Bhattacharya, Arani (Advisor) IEEE 802.11ax, popularly known as WiFi 6, introduces OFDMA (Orthogonal Frequency Division Multiple Access) that allows multiple users to transmit or receive frames concurrently via a more flexible utilization of the available bandwidth. Due to its ability to perform concurrent transmissions, prior studies suggest that OFDMA will provide reduced latency and increased throughput compared to OFDM. As our first work, we investigate this claim under various downlink traffic loads using the latest 802.11ax models supplied by the widely used open-source network simulator ns-3 (version 3.34). Our simulation results show that the actual benefits of OFDMA over OFDM can only be extracted under intermediate traffic loads. Motivated by this finding, we compare the performance of OFDMA and OFDM by simulating various application settings involving intermediate traffic rates. We find that OFDMA provides considerable improvements over OFDM under such traffic conditions for different application-specific parameters of interest, e.g., OFDMA delivers a more consistent bitrate for live video streaming applications and reduced jitter for video conferencing applications. Furthermore, for applications involving small payloads such as web-based applications and factory IoT-based motion control applications, OFDMA results in 7 − 10× lower average latency when compared to OFDM. We observe from our experiments that OFDMA experiences a substantial increase in latency as the payload size becomes small (90-30 B). In an IoT-based factory setting, applications frequently exchange small payloads with stringent deadline constraints; missing these deadlines for certain critical applications such as human-safety monitoring and equipment control have a huge impact or penalty. Our experiments make it evident that such constraints cannot be met by simple schedulers such as round-robin. Thus there is a need for intelligent scheduling techniques in such deadline-based environments. As our second work, we propose a deadline- aware scheduler for factory environments that maximizes the number of packet deadlines met for critical applications and show that our scheduler significantly reduces the overall penalty incurred due to missed deadlines compared to other deadline-based schedulers. 2021-12-01T00:00:00Z http://repository.iiitd.edu.in/xmlui/handle/123456789/1094 Scalable spatio-temporal arrival time estimation for public transit Dhingra, Karan; Biyani, Pravesh (Advisor) The problem of the ETA prediction of public transit has an essential role in improving the rider’s experience. While, it is challenging to ensure the timeliness of bus, especially during the rush hours. This thesis provides a heads-up on estimated arrival time for better planning using the open-transit data. The first step of providing a real-time scalable ETA is to design an algorithm that can preprocess the raw GTFS data of a day into a tensor. The representation aims to decouple the information about the bus, thereby enabling scalability across routes and reducing variance. The second step is to design a Spatio-temporal model (SSTG) for scalable and robust ETA prediction. In the proposed SSTG framework, we will provide answers to the following open problems. Firstly how can we exploit the spatial-temporal correlation in the ETA data? Secondly, how to scale the spatial-temporal ETA prediction framework on a large network effectively? Thirdly, How to handle sparsity in the data? Fourthly, the prediction of the ETA for the cold start stops is an unexplored problem. i.e., stops that are absent from the training dataset, how can we predict ETA for a cold start-stop? Moreover, a user would prefer waiting a bit longer than missing the bus because of underestimation. Therefore, for better customer satisfaction, we need to reduce the underestimation. The proposed framework captures the Spatiotemporal structure in the ETA data using recurrent neural networks modified with a graph convolutional. The input to the network can be sub-sampled, thereby ensuring scalable learning and further providing a solution to the cold start stops ETA prediction. The first layer of the encoder integrates GRU-D for the missing data imputation. Moreover, we use a MSLE- Weighted loss function to overestimate the ETA and fine-tune the penalty on overall performance compared to the regression loss(MSE) function. We finally conclude that the SSTG model is computationally efficient and outperforms the state-of-the-art methods on ETA and traffic datasets. 2021-09-01T00:00:00Z http://repository.iiitd.edu.in/xmlui/handle/123456789/1091 Text generation for populating semi-structured tables Mehta, Priya; Goyal, Vikram (Advisor); Chakraborty, Tanmoy (Advisor) Heterogeneous semi-structured tables are commonly used to represent data on the in- ternet. Recent years have seen a flurry of works in tasks that endeavor to comprehend such tabular information, such as table summarization, tabular question answering, and tabular fact-checking, to name a few. In this work, we proffer a new task in the realm of tabular data analysis called ‘Populating Semi-structured Tables’, wherein, given a partially filled table and related content, the aim is to generate text for the missing cells in the table. While most of the tasks that reason over semi-structured tables utilize the transformer-based sequence-to-sequence models, the table’s hier- archical structure and long-tailed nature seem to limit the performance of language models. Thus, we extend the traditional sequence-to-sequence models and propose sequence to multi-sequence models to handle multiple missing cell contents which are partially dependent on each other. Our inspiration comes from the system used for one-to-many sequence transduction problems with speech data which is yet to be experimented with for natural language generation tasks. The results show that our model, ‘Multiple Cell Filler’ (MuCeF) is better than the top baseline by a 15.44 ROUGE score and 34.54 METEOR score. Resources related to this work will be open-sourced for further research. 2021-12-01T00:00:00Z