A Transfer Learning Approach for Securing Resource-Constrained IoT Devices

Abstract : In recent years, Internet of Things (IoT) security has attracted significant interest by researchers due to new characteristics of IoT such as heterogeneity of devices, resource constraints, and new types of attacks targeting IoT. Intrusion detection, which is an indispensable part of a security system, is also included in these studies. In order to explore the complex characteristics of IoT, machine learning methods, which rely on long training time to generate intrusion detection models, are proposed in the literature. Furthermore, these systems need to learn a new/fresh model from scratch when the environment changes. This study explores the use of transfer learning in order to generate intrusion detection algorithms for such dynamically changing IoT. Transfer learning is an approach that stores knowledge learned from a problem domain/task and applies that knowledge to another problem domain/task. Here, it is employed in the following two settings: transferring knowledge for generating suitable intrusion algorithms for new devices, transferring knowledge for detecting new types of attacks. In this study, Routing Protocol for Low-Power and Lossy Network (RPL), a routing protocol for resource-constrained wireless networks, is used as an exemplar protocol and specific attacks against RPL are targeted. The experimental results show that the transfer learning approach gives better performance than the traditional approach. Moreover, the proposed approach significantly reduces learning time, which is an important factor for putting devices/networks in operation in a timely manner. Even though transfer learning has been considered a potential candidate for improving IoT security, to the best of our knowledge, this is the first application of transfer learning under these two settings in RPL-based IoT networks.

 EXISTING SYSTEM :

 ? It is possible that most of the proficient clients go out of the network, and existing clients do not fulfill system requirements. ? Existing methods in the literature that are proposed to reduce communication overhead can be categorized into decentralized training, compression, and local updating. ? Most existing studies and implementations consider synchronous FL, where the progress of the iteration round’s training period depends on the slowest device within the network. ? Due to mobility of clients, new clients may join a network that is more competent than the existing clients, and any client may leave the network during communication that can hamper the model training.

 DISADVANTAGE :

 ? It is a population-based search algorithm in order to evolve better individuals that correspond to candidate solutions for a targeted problem at each generation. ? The individuals, which represent candidate solutions for the problem at hand, are usually generated randomly in the first population. ? Transfer learning is a way to handle this problem by transferring the information from a source domain to a target domain that has limited data. ? Therefore, the problem at hand becomes a multi-objective problem with the goals of higher accuracy and lower energy consumption. ? This problem is not desired in many GP applications, hence the tree depth parameter is often decreased.

 PROPOSED SYSTEM :

 • In this work, they eliminate the straggler clients using an optimization technique named ‘soft-training’ that dynamically masks different neurons based on model updates, and their proposed aggregation scheme speeds up the collaborative convergence. • To ensure convergence in a non-IID scenario, particularly for asynchronous learning, loss functions of the non-convex problem need to be considered, and supportive algorithms should be proposed. • We need to select clients for training purposes based on system requirements. • However, due to strict cost and energy requirements, only a few clients might end up meeting the required criterion.

 ADVANTAGE :

 ? It is unsurprising to observe that the comparative convergence capability of the proposed approach is proportional to its initial performance by showing early convergence behavior. ? Moreover, it is expected to produce higher initial and final performances for the learned model in the new task/domain compared to learning without transfer. ? Attacker node contrarily prefers the worst parent to send or forward packets, which degrades the performance of the network (e.g., end-to-end delay, delivery ratio). ? It is statistically proven that the proposed approach shows much better performance than the traditional approach. ? DR is observed to generally degrade the performance of the proposed approach.