Akash Vartak

Deep neural networks (DNNs) are becoming commonplace in critical applications, making their susceptibility to backdoor (trojan) attacks a significant problem. In this paper, we introduce a novel backdoor attack detection pipeline, detecting attacked models using graph convolution networks (DeBUGCN). To the best of our knowledge, ours is the first use of GCNs for trojan detection. We use the static weights of a DNN to create a graph structure of its layers. A GCN is then used as a binary classifier on these graphs, yielding a trojan or clean determination for the DNN. To demonstrate the efficacy of our pipeline, we train hundreds of clean and trojaned CNN models on the MNIST handwritten digits and CIFAR-10 image datasets, and show the DNN classification results using DeBUGCN. For a true In-the-Wild use case, our pipeline is evaluated on the TrojAI dataset which consists of various CNN architectures, thus showing the robustness and model-agnostic behaviour of DeBUGCN. Furthermore, on comparing our results on several datasets with state-of-the-art trojan detection algorithms, DeBUGCN is faster and more accurate.

People identify man-made objects by their visual appearance and the text on them e.g., does a bottle say water or shampoo? We use text as an important visual cue to help distinguish between similar looking objects. This thesis explores a novel joint model of visual appearance and textual cues for image classification.We perform this in three functions - (a) Isolating an object in an input image; (b) Extracting text from the image; (c) Training a joint vision/text model. We simplify the task by extracting text separately and presenting it to the model in machine readable format. Such a joint model has utility in many real world challenges where language is interpreted through a sensory perception like vision or sound.The aim of the research is to understand whether visual percepts, when understood in the context of extracted language, will provide a better classification of image objects than using only pure vision to perform image classification. In conclusion, we show that joint classifier models can successfully make use of text present in images to classify objects, provided that the extracted text from images is of high quality and we have the number of images proportional to the number of classification classes.