Physical Backdoor Attack using Multi-Trigger to Same Class

Data poisoning is an attack on machine learning models that involves inserting a sample into the training data sets that does not influence the model during training but does affect it during testing. A backdoor attack aims to inject a hidden backdoor instance into deep neural networks (DNNs), causing the attacked models to perform well on benign samples while the prediction is maliciously modified if the attacker has initiated the hidden backdoor attack in neural networks. Previously, triggers were embedded digitally in data sets, therefore trigger settings were static and followed the same look and placement. Physical backdoor research is limited due to a lack of large datasets including actual photos of physical objects present with classification goals. These datasets need a significant amount of time and effort to construct. In this paper, we attempt to overcome the difficulties encountered by researchers during the physical backdoor attack. The presence of a physical trigger in an instance is very plausible, yet little study has been done on physical trigger-based attacks so far. We manually constructed a dataset with both a physical trigger and clean picture. The dataset will comprise trigger images with physical triggers, and we will be able to train the model with various sorts of triggers. The different trigger for the same class has been used to misclassify the same person using the different trigger to the target class. We have created a data set that will be used for facial recognition having a total of 1100 images of the clean images and physically triggered images with 3 volunteers including 2 physical triggers on them.