A Light-weight Watermarking-Based Framework on Dataset Using Deep Learning Algorithms
Muhammad Tayyab, Mohsen Marjani, N. Z. Jhanjhi, Mohamed Hashem
Abstract
In most decision-based security applications Deep Learning (DL) algorithms have been widely using for improvement. For better performance, a large amount of dataset has been used for training the DL algorithms. As DL has been remained a key element in the performance of the application, hence, several privacy and security issues have reported, which have affected the performance. Such security attacks have also affected the performance by taking the advantage of the huge dataset, because it is easy for an attacker to add executable noise into the dataset to get the information of the dataset and the model used. Most common security attacks like poisoning and evasion attacks have been considered challenging attacks that have caused misclassification and wrong prediction. Hence, a secure metric is needed to mitigate the effects of such attacks from the dataset. Therefore, in this paper, a light-weight watermarking framework has been proposed that provides security to the dataset before training the DL algorithms. We have implemented our proposed framework using the most common Convolutional Neural Network (CNN) and Artificial Neural Network (ANN) against security attacks. The proposed framework has been evaluated based on accuracy, precision, and computational cost, and has maintained the accuracy up to 98.89% and a precision of 0.96, which has maintained the level as in recent literature. We have also reduced the computational cost for the proposed framework. We believed that the proposed framework can be used to mitigate the security issues in DL algorithms and enhanced toward other security applications.